{ "cells": [ { "cell_type": "markdown", "id": "08afcefc", "metadata": {}, "source": [ "# Классификация астероидов прототипной нейросетью\n", "\n", "В работе автоматически определяется ключевая зависимая переменная, выполняется предобработка данных и строится **прототипная нейросеть типа RBF**:\n", "- для каждого класса выделяются прототипы методом `MiniBatchKMeans`;\n", "- расстояния до прототипов преобразуются в радиальные отклики;\n", "- выходной слой обучается логистической регрессией по откликам прототипов.\n", "\n", "Далее вычисляются:\n", "- метрики качества на train / validation / test;\n", "- глобальная важность исходных признаков методом перестановок;\n", "- основные прототипы, их опорные объекты и вклады признаков внутри прототипов.\n", "\n", "Финальный текстовый вывод в ноутбуке **не приводится**." ] }, { "cell_type": "code", "execution_count": 1, "id": "9b6fb9eb", "metadata": {}, "outputs": [], "source": [ "import os\n", "import re\n", "import glob\n", "import warnings\n", "from pathlib import Path\n", "\n", "import numpy as np\n", "import pandas as pd\n", "import matplotlib.pyplot as plt\n", "\n", "from sklearn.model_selection import train_test_split\n", "from sklearn.compose import ColumnTransformer\n", "from sklearn.pipeline import Pipeline\n", "from sklearn.impute import SimpleImputer\n", "from sklearn.preprocessing import OneHotEncoder, StandardScaler\n", "from sklearn.cluster import MiniBatchKMeans\n", "from sklearn.linear_model import LogisticRegression\n", "from sklearn.metrics import (\n", " accuracy_score,\n", " balanced_accuracy_score,\n", " precision_score,\n", " recall_score,\n", " f1_score,\n", " roc_auc_score,\n", " confusion_matrix,\n", " classification_report,\n", ")\n", "from sklearn.metrics import pairwise_distances, pairwise_distances_argmin_min\n", "\n", "warnings.filterwarnings(\"ignore\")\n", "\n", "SEED = 42\n", "np.random.seed(SEED)" ] }, { "cell_type": "code", "execution_count": 2, "id": "dd08087c", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Используемый файл: Asteroid_Updated.csv\n", "Форма датасета: (839714, 31)\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
nameaeiomwqadper_ydata_arc...UBIRspec_Bspec_TGmoidclassnperma
0Ceres2.7691650.07600910.59406780.30553273.5976942.5586842.9796474.6082028822.0...0.426NaNCG0.121.59478MBA0.2138851683.14570877.372096
1Pallas2.7724660.23033734.836234173.080063310.0488572.1338653.4110674.61644472318.0...0.284NaNBB0.111.23324MBA0.2135031686.15599959.699133
2Juno2.6691500.25694212.988919169.852760248.1386261.9833323.3549674.36081472684.0...0.433NaNSkS0.321.03454MBA0.2260191592.78728534.925016
3Vesta2.3614180.0887217.141771103.810804150.7285412.1519092.5709263.62883724288.0...0.492NaNVV0.321.13948MBA0.2716091325.43276595.861936
4Astraea2.5742490.1910955.366988141.576605358.6876072.0823243.0661744.13032363507.0...0.411NaNSSNaN1.09589MBA0.2386321508.600458282.366289
\n", "

5 rows × 31 columns

\n", "
" ], "text/plain": [ " name a e i om w q \\\n", "0 Ceres 2.769165 0.076009 10.594067 80.305532 73.597694 2.558684 \n", "1 Pallas 2.772466 0.230337 34.836234 173.080063 310.048857 2.133865 \n", "2 Juno 2.669150 0.256942 12.988919 169.852760 248.138626 1.983332 \n", "3 Vesta 2.361418 0.088721 7.141771 103.810804 150.728541 2.151909 \n", "4 Astraea 2.574249 0.191095 5.366988 141.576605 358.687607 2.082324 \n", "\n", " ad per_y data_arc ... UB IR spec_B spec_T G moid \\\n", "0 2.979647 4.608202 8822.0 ... 0.426 NaN C G 0.12 1.59478 \n", "1 3.411067 4.616444 72318.0 ... 0.284 NaN B B 0.11 1.23324 \n", "2 3.354967 4.360814 72684.0 ... 0.433 NaN Sk S 0.32 1.03454 \n", "3 2.570926 3.628837 24288.0 ... 0.492 NaN V V 0.32 1.13948 \n", "4 3.066174 4.130323 63507.0 ... 0.411 NaN S S NaN 1.09589 \n", "\n", " class n per ma \n", "0 MBA 0.213885 1683.145708 77.372096 \n", "1 MBA 0.213503 1686.155999 59.699133 \n", "2 MBA 0.226019 1592.787285 34.925016 \n", "3 MBA 0.271609 1325.432765 95.861936 \n", "4 MBA 0.238632 1508.600458 282.366289 \n", "\n", "[5 rows x 31 columns]" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Поиск CSV-файла\n", "candidate_files = [\n", " \"Asteroid_Updated.csv\",\n", " \"asteroid_updated.csv\",\n", " \"Asteroid Updated.csv\",\n", " \"asteroid.csv\",\n", " \"asteroid_dataset.csv\",\n", " \"Asteroid_Dataset.csv\"\n", "]\n", "\n", "csv_path = None\n", "for cand in candidate_files:\n", " if Path(cand).exists():\n", " csv_path = cand\n", " break\n", "\n", "if csv_path is None:\n", " matches = glob.glob(\"*.csv\")\n", " if len(matches) == 1:\n", " csv_path = matches[0]\n", " elif len(matches) > 1:\n", " # стараемся выбрать наиболее подходящий по имени\n", " scored = []\n", " for m in matches:\n", " score = 0\n", " low = m.lower()\n", " if \"asteroid\" in low:\n", " score += 5\n", " if \"updated\" in low:\n", " score += 2\n", " scored.append((score, m))\n", " scored = sorted(scored, reverse=True)\n", " csv_path = scored[0][1]\n", "\n", "if csv_path is None:\n", " raise FileNotFoundError(\"CSV-файл не найден. Поместите файл Asteroid_Updated.csv в ту же папку, что и ноутбук.\")\n", "\n", "print(\"Используемый файл:\", csv_path)\n", "df_raw = pd.read_csv(csv_path)\n", "print(\"Форма датасета:\", df_raw.shape)\n", "display(df_raw.head())" ] }, { "cell_type": "code", "execution_count": 3, "id": "8a8386bf", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
original_namenormalized_name
0namename
1aa
2ee
3ii
4omom
5ww
6qq
7adad
8per_yper_y
9data_arcdata_arc
10condition_codecondition_code
11n_obs_usedn_obs_used
12Hh
13neoneo
14phapha
15diameterdiameter
16extentextent
17albedoalbedo
18rot_perrot_per
19GMgm
20BVbv
21UBub
22IRir
23spec_Bspec_b
24spec_Tspec_t
25Gg
26moidmoid
27classclass
28nn
29perper
30mama
\n", "
" ], "text/plain": [ " original_name normalized_name\n", "0 name name\n", "1 a a\n", "2 e e\n", "3 i i\n", "4 om om\n", "5 w w\n", "6 q q\n", "7 ad ad\n", "8 per_y per_y\n", "9 data_arc data_arc\n", "10 condition_code condition_code\n", "11 n_obs_used n_obs_used\n", "12 H h\n", "13 neo neo\n", "14 pha pha\n", "15 diameter diameter\n", "16 extent extent\n", "17 albedo albedo\n", "18 rot_per rot_per\n", "19 GM gm\n", "20 BV bv\n", "21 UB ub\n", "22 IR ir\n", "23 spec_B spec_b\n", "24 spec_T spec_t\n", "25 G g\n", "26 moid moid\n", "27 class class\n", "28 n n\n", "29 per per\n", "30 ma ma" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Типы данных после нормализации:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
nameaeiomwqadper_ydata_arc...ubirspec_bspec_tgmoidclassnperma
dtypeobjectfloat64float64float64float64float64float64float64float64float64...float64float64objectobjectfloat64float64objectfloat64float64float64
\n", "

1 rows × 31 columns

\n", "
" ], "text/plain": [ " name a e i om w q ad \\\n", "dtype object float64 float64 float64 float64 float64 float64 float64 \n", "\n", " per_y data_arc ... ub ir spec_b spec_t g \\\n", "dtype float64 float64 ... float64 float64 object object float64 \n", "\n", " moid class n per ma \n", "dtype float64 object float64 float64 float64 \n", "\n", "[1 rows x 31 columns]" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Нормализация названий столбцов и попытка приведения object-столбцов к числам\n", "def normalize_col(col: str) -> str:\n", " col = str(col).strip().lower()\n", " col = col.replace(\"%\", \"pct\")\n", " col = re.sub(r\"[^a-z0-9]+\", \"_\", col)\n", " col = re.sub(r\"_+\", \"_\", col).strip(\"_\")\n", " return col\n", "\n", "df = df_raw.copy()\n", "orig_cols = df.columns.tolist()\n", "df.columns = [normalize_col(c) for c in df.columns]\n", "\n", "col_map = pd.DataFrame({\n", " \"original_name\": orig_cols,\n", " \"normalized_name\": df.columns\n", "})\n", "display(col_map)\n", "\n", "for col in df.columns:\n", " if df[col].dtype == \"object\":\n", " cleaned = (\n", " df[col]\n", " .astype(str)\n", " .str.replace(\",\", \"\", regex=False)\n", " .str.strip()\n", " .replace({\"\": np.nan, \"nan\": np.nan, \"none\": np.nan, \"None\": np.nan, \"null\": np.nan})\n", " )\n", " parsed = pd.to_numeric(cleaned, errors=\"coerce\")\n", " share_numeric = parsed.notna().mean()\n", " if share_numeric >= 0.95:\n", " df[col] = parsed\n", "\n", "print(\"Типы данных после нормализации:\")\n", "display(df.dtypes.rename(\"dtype\").to_frame().T)" ] }, { "cell_type": "code", "execution_count": 4, "id": "1e02249a", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Выбранная зависимая переменная: pha\n", "Распределение классов:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
share
pha
00.997552
10.002448
\n", "
" ], "text/plain": [ " share\n", "pha \n", "0 0.997552\n", "1 0.002448" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Автоматический выбор зависимой переменной\n", "target_candidates = [\n", " \"pha\", \"is_hazardous\", \"hazardous\", \"hazardous_asteroid\",\n", " \"potentially_hazardous\", \"risk\", \"target\"\n", "]\n", "\n", "target_col = None\n", "for cand in target_candidates:\n", " if cand in df.columns:\n", " target_col = cand\n", " break\n", "\n", "if target_col is None:\n", " for col in df.columns:\n", " nun = df[col].nunique(dropna=True)\n", " if nun == 2 and any(key in col for key in [\"haz\", \"risk\", \"target\", \"pha\"]):\n", " target_col = col\n", " break\n", "\n", "if target_col is None:\n", " raise ValueError(\"Не удалось автоматически определить целевую переменную для бинарной классификации.\")\n", "\n", "print(\"Выбранная зависимая переменная:\", target_col)\n", "\n", "target_series = df[target_col].copy()\n", "if target_series.dtype == \"object\":\n", " target_clean = (\n", " target_series.astype(str)\n", " .str.strip()\n", " .str.lower()\n", " .replace({\"\": np.nan, \"nan\": np.nan, \"none\": np.nan, \"null\": np.nan})\n", " )\n", " mapping = {\n", " \"y\": 1, \"yes\": 1, \"true\": 1, \"1\": 1, \"t\": 1,\n", " \"n\": 0, \"no\": 0, \"false\": 0, \"0\": 0, \"f\": 0\n", " }\n", " uniq = pd.Series(target_clean.dropna().unique())\n", " if set(uniq.tolist()).issubset(set(mapping.keys())):\n", " y = target_clean.map(mapping)\n", " else:\n", " vals = list(uniq.sort_values())\n", " if len(vals) != 2:\n", " raise ValueError(f\"Целевая переменная {target_col} не является бинарной. Найденные значения: {vals[:10]}\")\n", " y = (target_clean == vals[-1]).astype(float)\n", "else:\n", " uniq = sorted(pd.Series(target_series.dropna().unique()).tolist())\n", " if len(uniq) != 2:\n", " raise ValueError(f\"Целевая переменная {target_col} не является бинарной.\")\n", " y = (target_series == uniq[-1]).astype(float)\n", "\n", "mask = y.notna()\n", "df = df.loc[mask].reset_index(drop=True)\n", "y = y.loc[mask].astype(int).reset_index(drop=True)\n", "\n", "print(\"Распределение классов:\")\n", "display(y.value_counts(normalize=True).sort_index().rename(\"share\").to_frame())" ] }, { "cell_type": "code", "execution_count": 5, "id": "faa5136c", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Датасет уменьшен до 250000 объектов для устойчивого обучения с сохранением баланса классов.\n" ] } ], "source": [ "# При необходимости уменьшаем объём данных для устойчивого обучения\n", "MAX_ROWS_FOR_MODELING = 250_000\n", "\n", "if len(df) > MAX_ROWS_FOR_MODELING:\n", " tmp = pd.DataFrame({\"target\": y})\n", " sampled_idx = (\n", " tmp.groupby(\"target\", group_keys=False)\n", " .apply(lambda g: g.sample(\n", " n=max(1, int(round(len(g) / len(tmp) * MAX_ROWS_FOR_MODELING))),\n", " random_state=SEED\n", " ))\n", " .index\n", " )\n", " sampled_idx = np.sort(sampled_idx)\n", " df = df.loc[sampled_idx].reset_index(drop=True)\n", " y = y.loc[sampled_idx].reset_index(drop=True)\n", " print(f\"Датасет уменьшен до {len(df)} объектов для устойчивого обучения с сохранением баланса классов.\")\n", "else:\n", " print(\"Используется полный датасет.\")" ] }, { "cell_type": "code", "execution_count": 6, "id": "4b2d11fb", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Удалённые признаки: ['albedo', 'bv', 'diameter', 'extent', 'g', 'gm', 'ir', 'name', 'rot_per', 'spec_b', 'spec_t', 'ub']\n", "\n", "Число исходных содержательных признаков: 18\n", "Числовые признаки: ['a', 'e', 'i', 'om', 'w', 'q', 'ad', 'per_y', 'data_arc', 'condition_code', 'n_obs_used', 'h', 'moid', 'n', 'per', 'ma']\n", "Категориальные признаки: ['neo', 'class']\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
aeiomwqadper_ydata_arccondition_coden_obs_usedhmoidnpermaneoclass
02.7691650.07600910.59406780.30553273.5976942.5586842.9796474.6082028822.00.010023.341.5947800.2138851683.14570877.372096NMBA
12.3853340.2312065.523651259.563231145.2651061.8338312.9368373.68410562452.00.052065.510.8461000.2675351345.619196140.419656NMBA
22.2017640.1564995.886955110.889330285.2874621.8571902.5463393.26711562655.00.027446.490.8741760.3016811193.313717194.882895NMBA
33.1415390.1124613.831560283.202167312.3152062.7882403.4948395.56829162175.00.034095.431.7783900.1770072033.818284152.184851NMBA
42.5759810.08512116.53612543.22191380.5448232.3567102.7952524.13449261680.00.023856.741.4363300.2383911510.123380187.488522NMBA
\n", "
" ], "text/plain": [ " a e i om w q ad \\\n", "0 2.769165 0.076009 10.594067 80.305532 73.597694 2.558684 2.979647 \n", "1 2.385334 0.231206 5.523651 259.563231 145.265106 1.833831 2.936837 \n", "2 2.201764 0.156499 5.886955 110.889330 285.287462 1.857190 2.546339 \n", "3 3.141539 0.112461 3.831560 283.202167 312.315206 2.788240 3.494839 \n", "4 2.575981 0.085121 16.536125 43.221913 80.544823 2.356710 2.795252 \n", "\n", " per_y data_arc condition_code n_obs_used h moid n \\\n", "0 4.608202 8822.0 0.0 1002 3.34 1.594780 0.213885 \n", "1 3.684105 62452.0 0.0 5206 5.51 0.846100 0.267535 \n", "2 3.267115 62655.0 0.0 2744 6.49 0.874176 0.301681 \n", "3 5.568291 62175.0 0.0 3409 5.43 1.778390 0.177007 \n", "4 4.134492 61680.0 0.0 2385 6.74 1.436330 0.238391 \n", "\n", " per ma neo class \n", "0 1683.145708 77.372096 N MBA \n", "1 1345.619196 140.419656 N MBA \n", "2 1193.313717 194.882895 N MBA \n", "3 2033.818284 152.184851 N MBA \n", "4 1510.123380 187.488522 N MBA " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Отбор содержательных признаков\n", "name_series = df[\"name\"].copy() if \"name\" in df.columns else pd.Series(np.arange(len(df)), name=\"object_name\")\n", "\n", "X = df.drop(columns=[target_col]).copy()\n", "\n", "explicit_drop = {\n", " \"name\", \"id\", \"identifier\", \"full_name\", \"spkid\", \"pdes\",\n", " \"orbit_id\", \"extent\"\n", "}\n", "to_drop = [c for c in X.columns if c in explicit_drop]\n", "\n", "# Удаляем почти пустые столбцы\n", "missing_share = X.isna().mean()\n", "to_drop += missing_share[missing_share > 0.80].index.tolist()\n", "\n", "# Удаляем высококардинальные строковые идентификаторы\n", "candidate_keep = {\"neo\", \"class\", \"spec_b\", \"spec_t\"}\n", "obj_cols = X.select_dtypes(include=\"object\").columns.tolist()\n", "for col in obj_cols:\n", " nun = X[col].nunique(dropna=True)\n", " ratio = nun / max(len(X), 1)\n", " avg_len = X[col].dropna().astype(str).str.len().mean() if X[col].notna().any() else 0\n", " if col in candidate_keep:\n", " continue\n", " if nun > 50 or ratio > 0.10 or avg_len > 25:\n", " to_drop.append(col)\n", "\n", "to_drop = sorted(set([c for c in to_drop if c in X.columns]))\n", "X = X.drop(columns=to_drop, errors=\"ignore\")\n", "\n", "numeric_features = X.select_dtypes(include=[np.number]).columns.tolist()\n", "categorical_features = [c for c in X.columns if c not in numeric_features]\n", "\n", "cat_keep = []\n", "for col in categorical_features:\n", " nun = X[col].nunique(dropna=True)\n", " if 2 <= nun <= 30:\n", " cat_keep.append(col)\n", "\n", "categorical_features = cat_keep\n", "X = X[numeric_features + categorical_features].copy()\n", "\n", "print(\"Удалённые признаки:\", to_drop)\n", "print(\"\\nЧисло исходных содержательных признаков:\", X.shape[1])\n", "print(\"Числовые признаки:\", numeric_features)\n", "print(\"Категориальные признаки:\", categorical_features)\n", "display(X.head())" ] }, { "cell_type": "code", "execution_count": 7, "id": "c6931339", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Размеры выборок:\n", "train: (150000, 18) (150000,)\n", "val: (50000, 18) (50000,)\n", "test: (50000, 18) (50000,)\n" ] } ], "source": [ "# Разбиение на train / validation / test\n", "X_trainval, X_test, y_trainval, y_test, names_trainval, names_test = train_test_split(\n", " X, y, name_series,\n", " test_size=0.20,\n", " random_state=SEED,\n", " stratify=y\n", ")\n", "\n", "X_train, X_val, y_train, y_val, names_train, names_val = train_test_split(\n", " X_trainval, y_trainval, names_trainval,\n", " test_size=0.25, # 0.25 * 0.80 = 0.20\n", " random_state=SEED,\n", " stratify=y_trainval\n", ")\n", "\n", "print(\"Размеры выборок:\")\n", "print(\"train:\", X_train.shape, y_train.shape)\n", "print(\"val:\", X_val.shape, y_val.shape)\n", "print(\"test:\", X_test.shape, y_test.shape)" ] }, { "cell_type": "code", "execution_count": 8, "id": "76262ee1", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Число признаков после кодирования: 30\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
num__anum__enum__inum__omnum__wnum__qnum__adnum__per_ynum__data_arcnum__condition_codenum__n_obs_usednum__h
00.0311880.806925-1.1288520.3146690.0598680.0230070.029438-0.0038204.396863-0.4791872.890778-1.740824
1-0.016760-0.608113-0.471353-0.340761-1.1257700.030330-0.020955-0.006358-0.012200-0.479187-0.4159830.067917
20.032497-0.4832321.726338-1.054857-0.0842240.2038220.009041-0.003748-0.408712-0.479187-0.558353-0.315755
3-0.060804-0.890887-0.3763641.356538-0.673594-0.112042-0.049310-0.0085120.492364-0.4791870.148128-0.151324
4-0.0472001.4794580.571230-0.8026310.340934-0.305853-0.011950-0.007866-1.3603393.006464-0.6738620.719612
\n", "
" ], "text/plain": [ " num__a num__e num__i num__om num__w num__q num__ad \\\n", "0 0.031188 0.806925 -1.128852 0.314669 0.059868 0.023007 0.029438 \n", "1 -0.016760 -0.608113 -0.471353 -0.340761 -1.125770 0.030330 -0.020955 \n", "2 0.032497 -0.483232 1.726338 -1.054857 -0.084224 0.203822 0.009041 \n", "3 -0.060804 -0.890887 -0.376364 1.356538 -0.673594 -0.112042 -0.049310 \n", "4 -0.047200 1.479458 0.571230 -0.802631 0.340934 -0.305853 -0.011950 \n", "\n", " num__per_y num__data_arc num__condition_code num__n_obs_used num__h \n", "0 -0.003820 4.396863 -0.479187 2.890778 -1.740824 \n", "1 -0.006358 -0.012200 -0.479187 -0.415983 0.067917 \n", "2 -0.003748 -0.408712 -0.479187 -0.558353 -0.315755 \n", "3 -0.008512 0.492364 -0.479187 0.148128 -0.151324 \n", "4 -0.007866 -1.360339 3.006464 -0.673862 0.719612 " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Предобработка\n", "try:\n", " ohe = OneHotEncoder(handle_unknown=\"ignore\", sparse_output=False)\n", "except TypeError:\n", " ohe = OneHotEncoder(handle_unknown=\"ignore\", sparse=False)\n", "\n", "num_pipe = Pipeline([\n", " (\"imputer\", SimpleImputer(strategy=\"median\")),\n", " (\"scaler\", StandardScaler())\n", "])\n", "\n", "cat_pipe = Pipeline([\n", " (\"imputer\", SimpleImputer(strategy=\"most_frequent\")),\n", " (\"ohe\", ohe)\n", "])\n", "\n", "preprocessor = ColumnTransformer(\n", " transformers=[\n", " (\"num\", num_pipe, numeric_features),\n", " (\"cat\", cat_pipe, categorical_features)\n", " ],\n", " remainder=\"drop\"\n", ")\n", "\n", "X_train_p = preprocessor.fit_transform(X_train).astype(np.float32)\n", "X_val_p = preprocessor.transform(X_val).astype(np.float32)\n", "X_test_p = preprocessor.transform(X_test).astype(np.float32)\n", "X_trainval_p = preprocessor.transform(X_trainval).astype(np.float32)\n", "\n", "feature_names = preprocessor.get_feature_names_out().tolist()\n", "\n", "print(\"Число признаков после кодирования:\", X_train_p.shape[1])\n", "display(pd.DataFrame(X_train_p[:5, :min(12, X_train_p.shape[1])], columns=feature_names[:min(12, len(feature_names))]))" ] }, { "cell_type": "code", "execution_count": 9, "id": "193784e9", "metadata": {}, "outputs": [], "source": [ "# Вспомогательные функции\n", "def evaluate_binary(y_true, y_pred, y_proba):\n", " out = {\n", " \"accuracy\": accuracy_score(y_true, y_pred),\n", " \"balanced_accuracy\": balanced_accuracy_score(y_true, y_pred),\n", " \"precision_macro\": precision_score(y_true, y_pred, average=\"macro\", zero_division=0),\n", " \"recall_macro\": recall_score(y_true, y_pred, average=\"macro\", zero_division=0),\n", " \"f1_macro\": f1_score(y_true, y_pred, average=\"macro\", zero_division=0),\n", " }\n", " try:\n", " out[\"roc_auc\"] = roc_auc_score(y_true, y_proba)\n", " except Exception:\n", " out[\"roc_auc\"] = np.nan\n", " return out\n", "\n", "def choose_sigma_from_centers(centers):\n", " if len(centers) <= 1:\n", " return 1.0\n", " d = pairwise_distances(centers)\n", " d[d == 0] = np.nan\n", " nearest = np.nanmin(d, axis=1)\n", " nearest = nearest[np.isfinite(nearest)]\n", " if len(nearest) == 0:\n", " return 1.0\n", " return float(np.nanmedian(nearest))\n", "\n", "class PrototypeRBFClassifier:\n", " def __init__(self, n_prototypes_per_class=4, sigma_multiplier=1.0, C=1.0, random_state=42):\n", " self.n_prototypes_per_class = n_prototypes_per_class\n", " self.sigma_multiplier = sigma_multiplier\n", " self.C = C\n", " self.random_state = random_state\n", "\n", " def _build_prototypes(self, X, y):\n", " centers = []\n", " labels = []\n", " unique_classes = np.sort(np.unique(y))\n", " for cls in unique_classes:\n", " Xc = X[y == cls]\n", " n_clusters = min(self.n_prototypes_per_class, len(Xc))\n", " if n_clusters <= 0:\n", " continue\n", " km = MiniBatchKMeans(\n", " n_clusters=n_clusters,\n", " random_state=self.random_state,\n", " batch_size=min(4096, max(128, len(Xc))),\n", " n_init=5\n", " )\n", " km.fit(Xc)\n", " centers.append(km.cluster_centers_)\n", " labels.extend([cls] * n_clusters)\n", " centers = np.vstack(centers).astype(np.float32)\n", " labels = np.array(labels)\n", " return centers, labels\n", "\n", " def _rbf(self, X):\n", " d2 = pairwise_distances(X, self.prototypes_, metric=\"euclidean\") ** 2\n", " denom = max(self.sigma_ ** 2, 1e-8) * 2.0\n", " return np.exp(-d2 / denom).astype(np.float32)\n", "\n", " def fit(self, X, y):\n", " X = np.asarray(X, dtype=np.float32)\n", " y = np.asarray(y)\n", " self.prototypes_, self.prototype_labels_ = self._build_prototypes(X, y)\n", " self.sigma_base_ = choose_sigma_from_centers(self.prototypes_)\n", " self.sigma_ = float(max(self.sigma_base_ * self.sigma_multiplier, 1e-6))\n", "\n", " Z = self._rbf(X)\n", " self.output_layer_ = LogisticRegression(\n", " max_iter=1000,\n", " C=self.C,\n", " class_weight=\"balanced\",\n", " random_state=self.random_state\n", " )\n", " self.output_layer_.fit(Z, y)\n", "\n", " self.assign_idx_, self.assign_dist_ = pairwise_distances_argmin_min(X, self.prototypes_)\n", " self.train_support_ = pd.Series(self.assign_idx_).value_counts().sort_index()\n", " return self\n", "\n", " def predict_proba(self, X):\n", " Z = self._rbf(np.asarray(X, dtype=np.float32))\n", " return self.output_layer_.predict_proba(Z)\n", "\n", " def predict(self, X):\n", " return self.predict_proba(X)[:, 1] >= 0.5" ] }, { "cell_type": "code", "execution_count": 10, "id": "ca93866f", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Топ-10 конфигураций:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
n_prototypes_per_classtotal_prototypessigma_multipliersigmaCval_accuracyval_balanced_accuracyval_precision_macroval_recall_macroval_f1_macroval_roc_auc
05102.05.9481372.00.979340.9896450.5528140.9896450.5903050.996080
15102.05.9481371.00.978900.9894240.5518270.9894240.5885740.996067
2482.05.8079962.00.978600.9892740.5511740.9892740.5874250.995346
35102.05.9481370.50.978240.9890930.5504130.9890930.5860780.995906
4481.02.9039982.00.978220.9849950.5500310.9849950.5853910.995502
5482.05.8079961.00.978020.9889830.5499590.9889830.5852720.995145
6361.03.7456382.00.977860.9889030.5496340.9889030.5846930.994570
7481.02.9039981.00.977700.9888230.5493130.9888230.5841200.995309
8482.05.8079960.50.977680.9888130.5492730.9888130.5840490.994914
9362.07.4912762.00.977580.9887630.5490750.9887630.5836950.993315
\n", "
" ], "text/plain": [ " n_prototypes_per_class total_prototypes sigma_multiplier sigma C \\\n", "0 5 10 2.0 5.948137 2.0 \n", "1 5 10 2.0 5.948137 1.0 \n", "2 4 8 2.0 5.807996 2.0 \n", "3 5 10 2.0 5.948137 0.5 \n", "4 4 8 1.0 2.903998 2.0 \n", "5 4 8 2.0 5.807996 1.0 \n", "6 3 6 1.0 3.745638 2.0 \n", "7 4 8 1.0 2.903998 1.0 \n", "8 4 8 2.0 5.807996 0.5 \n", "9 3 6 2.0 7.491276 2.0 \n", "\n", " val_accuracy val_balanced_accuracy val_precision_macro val_recall_macro \\\n", "0 0.97934 0.989645 0.552814 0.989645 \n", "1 0.97890 0.989424 0.551827 0.989424 \n", "2 0.97860 0.989274 0.551174 0.989274 \n", "3 0.97824 0.989093 0.550413 0.989093 \n", "4 0.97822 0.984995 0.550031 0.984995 \n", "5 0.97802 0.988983 0.549959 0.988983 \n", "6 0.97786 0.988903 0.549634 0.988903 \n", "7 0.97770 0.988823 0.549313 0.988823 \n", "8 0.97768 0.988813 0.549273 0.988813 \n", "9 0.97758 0.988763 0.549075 0.988763 \n", "\n", " val_f1_macro val_roc_auc \n", "0 0.590305 0.996080 \n", "1 0.588574 0.996067 \n", "2 0.587425 0.995346 \n", "3 0.586078 0.995906 \n", "4 0.585391 0.995502 \n", "5 0.585272 0.995145 \n", "6 0.584693 0.994570 \n", "7 0.584120 0.995309 \n", "8 0.584049 0.994914 \n", "9 0.583695 0.993315 " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Подбор гиперпараметров прототипной сети по validation\n", "n_proto_grid = [2, 3, 4, 5]\n", "sigma_mult_grid = [0.5, 1.0, 2.0]\n", "C_grid = [0.5, 1.0, 2.0]\n", "\n", "search_rows = []\n", "best_model = None\n", "best_tuple = None\n", "\n", "for n_proto in n_proto_grid:\n", " for sigma_mult in sigma_mult_grid:\n", " for C in C_grid:\n", " model = PrototypeRBFClassifier(\n", " n_prototypes_per_class=n_proto,\n", " sigma_multiplier=sigma_mult,\n", " C=C,\n", " random_state=SEED\n", " )\n", " model.fit(X_train_p, y_train.values)\n", "\n", " val_proba = model.predict_proba(X_val_p)[:, 1]\n", " val_pred = (val_proba >= 0.5).astype(int)\n", " val_metrics = evaluate_binary(y_val.values, val_pred, val_proba)\n", "\n", " row = {\n", " \"n_prototypes_per_class\": n_proto,\n", " \"total_prototypes\": len(model.prototypes_),\n", " \"sigma_multiplier\": sigma_mult,\n", " \"sigma\": model.sigma_,\n", " \"C\": C,\n", " **{f\"val_{k}\": v for k, v in val_metrics.items()}\n", " }\n", " search_rows.append(row)\n", "\n", " candidate_tuple = (\n", " val_metrics[\"f1_macro\"],\n", " val_metrics[\"balanced_accuracy\"],\n", " val_metrics[\"roc_auc\"] if pd.notna(val_metrics[\"roc_auc\"]) else -np.inf\n", " )\n", " if best_tuple is None or candidate_tuple > best_tuple:\n", " best_tuple = candidate_tuple\n", " best_model = model\n", "\n", "search_df = pd.DataFrame(search_rows).sort_values(\n", " [\"val_f1_macro\", \"val_balanced_accuracy\", \"val_roc_auc\"],\n", " ascending=False\n", ").reset_index(drop=True)\n", "\n", "print(\"Топ-10 конфигураций:\")\n", "display(search_df.head(10))" ] }, { "cell_type": "code", "execution_count": 11, "id": "accbc804", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Метрики модели:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
splitaccuracybalanced_accuracyprecision_macrorecall_macrof1_macroroc_auc
0trainval0.9793750.9886440.5530000.9886440.5906030.995785
1test0.9793200.9855460.5524160.9855460.5895910.996248
\n", "
" ], "text/plain": [ " split accuracy balanced_accuracy precision_macro recall_macro \\\n", "0 trainval 0.979375 0.988644 0.553000 0.988644 \n", "1 test 0.979320 0.985546 0.552416 0.985546 \n", "\n", " f1_macro roc_auc \n", "0 0.590603 0.995785 \n", "1 0.589591 0.996248 " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Матрица ошибок на test:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
01
0488451033
11121
\n", "
" ], "text/plain": [ " 0 1\n", "0 48845 1033\n", "1 1 121" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Классификационный отчёт на test:\n", " precision recall f1-score support\n", "\n", " 0 1.0000 0.9793 0.9895 49878\n", " 1 0.1049 0.9918 0.1897 122\n", "\n", " accuracy 0.9793 50000\n", " macro avg 0.5524 0.9855 0.5896 50000\n", "weighted avg 0.9978 0.9793 0.9876 50000\n", "\n" ] } ], "source": [ "# Финальная модель: переобучение на train+validation с лучшими параметрами\n", "best_params = search_df.iloc[0].to_dict()\n", "\n", "final_model = PrototypeRBFClassifier(\n", " n_prototypes_per_class=int(best_params[\"n_prototypes_per_class\"]),\n", " sigma_multiplier=float(best_params[\"sigma_multiplier\"]),\n", " C=float(best_params[\"C\"]),\n", " random_state=SEED\n", ")\n", "final_model.fit(X_trainval_p, y_trainval.values)\n", "\n", "# Метрики\n", "trainval_proba = final_model.predict_proba(X_trainval_p)[:, 1]\n", "trainval_pred = (trainval_proba >= 0.5).astype(int)\n", "\n", "test_proba = final_model.predict_proba(X_test_p)[:, 1]\n", "test_pred = (test_proba >= 0.5).astype(int)\n", "\n", "metrics_trainval = evaluate_binary(y_trainval.values, trainval_pred, trainval_proba)\n", "metrics_test = evaluate_binary(y_test.values, test_pred, test_proba)\n", "\n", "results_df = pd.DataFrame([\n", " {\"split\": \"trainval\", **metrics_trainval},\n", " {\"split\": \"test\", **metrics_test}\n", "])\n", "\n", "print(\"Метрики модели:\")\n", "display(results_df)\n", "print(\"Матрица ошибок на test:\")\n", "cm = confusion_matrix(y_test.values, test_pred)\n", "display(pd.DataFrame(cm))\n", "print(\"Классификационный отчёт на test:\")\n", "print(classification_report(y_test.values, test_pred, digits=4, zero_division=0))" ] }, { "cell_type": "code", "execution_count": 12, "id": "1b7d4810", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Топ-10 признаков по permutation importance:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
featurepermutation_importance_meanpermutation_importance_std
0e0.0476640.002647
1condition_code0.0078150.000350
2data_arc0.0036830.000315
3n0.0022030.003038
4n_obs_used0.0015050.000192
5ma0.0004220.000162
6w0.0000260.000099
7per0.0000000.000000
8per_y0.0000000.000000
9ad-0.0000260.000037
\n", "
" ], "text/plain": [ " feature permutation_importance_mean permutation_importance_std\n", "0 e 0.047664 0.002647\n", "1 condition_code 0.007815 0.000350\n", "2 data_arc 0.003683 0.000315\n", "3 n 0.002203 0.003038\n", "4 n_obs_used 0.001505 0.000192\n", "5 ma 0.000422 0.000162\n", "6 w 0.000026 0.000099\n", "7 per 0.000000 0.000000\n", "8 per_y 0.000000 0.000000\n", "9 ad -0.000026 0.000037" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Глобальная важность исходных признаков методом перестановок\n", "def score_from_raw(X_raw_df, y_true):\n", " Xp = preprocessor.transform(X_raw_df).astype(np.float32)\n", " proba = final_model.predict_proba(Xp)[:, 1]\n", " pred = (proba >= 0.5).astype(int)\n", " return f1_score(y_true, pred, average=\"macro\", zero_division=0)\n", "\n", "baseline_f1 = score_from_raw(X_test.copy(), y_test.values)\n", "perm_rows = []\n", "n_repeats = 3\n", "rng = np.random.default_rng(SEED)\n", "\n", "for col in X_test.columns:\n", " scores = []\n", " for _ in range(n_repeats):\n", " X_perm = X_test.copy()\n", " X_perm[col] = rng.permutation(X_perm[col].values)\n", " s = score_from_raw(X_perm, y_test.values)\n", " scores.append(baseline_f1 - s)\n", " perm_rows.append({\n", " \"feature\": col,\n", " \"permutation_importance_mean\": float(np.mean(scores)),\n", " \"permutation_importance_std\": float(np.std(scores))\n", " })\n", "\n", "perm_df = pd.DataFrame(perm_rows).sort_values(\n", " \"permutation_importance_mean\", ascending=False\n", ").reset_index(drop=True)\n", "\n", "print(\"Топ-10 признаков по permutation importance:\")\n", "display(perm_df.head(10))" ] }, { "cell_type": "code", "execution_count": 13, "id": "fddae509", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Основные прототипы:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
prototype_idprototype_classsupport_nsupport_share_trainvalprototype_output_weightimpact_scoreassigned_positive_rate_trainvalnearest_objectsecondition_codedata_arcnn_obs_usedma
030276750.138375-22.3402983.0913390.000000nan, nan, nan0.1737806.712051253.6742300.23730218.095537181.147050
100717650.358825-3.4363011.2330310.000000nan, nan, nan0.1438260.1872505927.5718530.226685199.988755189.886354
240487250.243625-3.3567720.8177940.000000nan, nan, nan0.1552310.2375375965.4589020.233649198.257712162.908708
320191610.0958056.9531260.6661440.000052nan, nan, nan0.1376370.20652412629.8022550.2179451149.485048176.631783
410272210.136105-3.9708710.5404550.000000nan, nan, nan0.1241960.2366196007.0144740.235234199.583740192.906197
55120120.01006010.9443900.1101010.086481nan, nan, nan0.5943845.1888671152.3109660.26427792.923956172.768315
67112370.00618511.8300920.0731690.099434nan, nan, nan0.4573774.5610351747.5696720.455755109.617623176.214085
7918170.0040858.2973590.0338950.083231nan, nan, nan0.3702354.0697672102.9447170.508083116.062424174.708108
8811900.0009507.4017950.0070320.178947nan, nan, nan0.4100744.2789471910.5978841.374819152.363158172.634906
96111970.0059850.6844610.0040960.075188nan, nan, nan0.2473855.0359231103.2247670.83558986.230576176.160290
\n", "
" ], "text/plain": [ " prototype_id prototype_class support_n support_share_trainval \\\n", "0 3 0 27675 0.138375 \n", "1 0 0 71765 0.358825 \n", "2 4 0 48725 0.243625 \n", "3 2 0 19161 0.095805 \n", "4 1 0 27221 0.136105 \n", "5 5 1 2012 0.010060 \n", "6 7 1 1237 0.006185 \n", "7 9 1 817 0.004085 \n", "8 8 1 190 0.000950 \n", "9 6 1 1197 0.005985 \n", "\n", " prototype_output_weight impact_score assigned_positive_rate_trainval \\\n", "0 -22.340298 3.091339 0.000000 \n", "1 -3.436301 1.233031 0.000000 \n", "2 -3.356772 0.817794 0.000000 \n", "3 6.953126 0.666144 0.000052 \n", "4 -3.970871 0.540455 0.000000 \n", "5 10.944390 0.110101 0.086481 \n", "6 11.830092 0.073169 0.099434 \n", "7 8.297359 0.033895 0.083231 \n", "8 7.401795 0.007032 0.178947 \n", "9 0.684461 0.004096 0.075188 \n", "\n", " nearest_objects e condition_code data_arc n \\\n", "0 nan, nan, nan 0.173780 6.712051 253.674230 0.237302 \n", "1 nan, nan, nan 0.143826 0.187250 5927.571853 0.226685 \n", "2 nan, nan, nan 0.155231 0.237537 5965.458902 0.233649 \n", "3 nan, nan, nan 0.137637 0.206524 12629.802255 0.217945 \n", "4 nan, nan, nan 0.124196 0.236619 6007.014474 0.235234 \n", "5 nan, nan, nan 0.594384 5.188867 1152.310966 0.264277 \n", "6 nan, nan, nan 0.457377 4.561035 1747.569672 0.455755 \n", "7 nan, nan, nan 0.370235 4.069767 2102.944717 0.508083 \n", "8 nan, nan, nan 0.410074 4.278947 1910.597884 1.374819 \n", "9 nan, nan, nan 0.247385 5.035923 1103.224767 0.835589 \n", "\n", " n_obs_used ma \n", "0 18.095537 181.147050 \n", "1 199.988755 189.886354 \n", "2 198.257712 162.908708 \n", "3 1149.485048 176.631783 \n", "4 199.583740 192.906197 \n", "5 92.923956 172.768315 \n", "6 109.617623 176.214085 \n", "7 116.062424 174.708108 \n", "8 152.363158 172.634906 \n", "9 86.230576 176.160290 " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Основные прототипы: поддержка, влияние, ближайшие объекты и интерпретация значений признаков\n", "assign_idx_trainval, assign_dist_trainval = pairwise_distances_argmin_min(X_trainval_p, final_model.prototypes_)\n", "support_counts = pd.Series(assign_idx_trainval).value_counts().sort_index()\n", "support_share = support_counts / len(X_trainval)\n", "\n", "coef = final_model.output_layer_.coef_.ravel()\n", "impact = np.abs(coef) * support_share.reindex(range(len(coef)), fill_value=0).values\n", "\n", "# Для интерпретации берём 6 наиболее важных исходных признаков\n", "top_proto_features = perm_df.head(6)[\"feature\"].tolist()\n", "\n", "prototype_rows = []\n", "for pid in range(len(final_model.prototypes_)):\n", " mask = assign_idx_trainval == pid\n", " subset_X = X_trainval.loc[mask]\n", " subset_y = y_trainval.loc[mask]\n", " subset_names = names_trainval.loc[mask] if isinstance(names_trainval, pd.Series) else pd.Series(names_trainval)[mask]\n", "\n", " nearest_idx = np.argsort(np.where(mask, assign_dist_trainval, np.inf))[:3]\n", " nearest_objects = \", \".join(pd.Series(names_trainval).iloc[nearest_idx].astype(str).tolist())\n", "\n", " row = {\n", " \"prototype_id\": pid,\n", " \"prototype_class\": int(final_model.prototype_labels_[pid]),\n", " \"support_n\": int(mask.sum()),\n", " \"support_share_trainval\": float(mask.mean()),\n", " \"prototype_output_weight\": float(coef[pid]),\n", " \"impact_score\": float(impact[pid]),\n", " \"assigned_positive_rate_trainval\": float(subset_y.mean()) if len(subset_y) else np.nan,\n", " \"nearest_objects\": nearest_objects\n", " }\n", " for feat in top_proto_features:\n", " if feat in subset_X.columns:\n", " if pd.api.types.is_numeric_dtype(subset_X[feat]):\n", " row[feat] = float(subset_X[feat].mean()) if len(subset_X) else np.nan\n", " else:\n", " row[feat] = subset_X[feat].mode(dropna=True).iloc[0] if len(subset_X) and subset_X[feat].notna().any() else np.nan\n", " prototype_rows.append(row)\n", "\n", "prototype_df = pd.DataFrame(prototype_rows).sort_values(\n", " [\"impact_score\", \"support_n\"], ascending=False\n", ").reset_index(drop=True)\n", "\n", "print(\"Основные прототипы:\")\n", "display(prototype_df.head(10))" ] }, { "cell_type": "code", "execution_count": 14, "id": "4d7d44d5", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Топ-5 вкладов признаков для основных прототипов:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
prototype_idfeaturecontribution_scoreinterpreted_valuerank
03condition_code47.3544886.7120511
13data_arc29.003140253.674232
23h15.00028218.0055363
33n_obs_used14.65050918.0955374
43i5.87660910.67045
50om2.82364284.0402741
60condition_code1.3966860.187252
70w0.829923205.946873
80n_obs_used0.577087199.9887554
90n0.4607640.2266855
104w2.97605088.6003461
114om2.472264244.5934222
124condition_code1.2990070.2375373
134n_obs_used0.579315198.2577124
144ma0.565179162.9087085
152n_obs_used16.5392341149.4850481
162data_arc11.54341512629.8022552
172h10.73302013.9513563
182q4.3113343.7801624
192moid4.2987702.7915475
201w3.681780277.272541
211om3.497154259.4515022
221condition_code1.5380640.2366193
231e1.3355890.1241964
241i0.9466167.3497165
255e51.1633960.5943841
265h29.82275821.7639822
275condition_code16.7446315.1888673
285data_arc11.8574911152.3109664
295neo9.156986Y5
307h39.83129622.9390381
317e38.0296360.4573772
327n30.9394060.4557553
337condition_code15.2241774.5610354
347neo11.453909Y5
359i32.00346834.4710531
369n26.9114280.5080832
379e18.9669810.3702353
389h17.86519820.7174014
399condition_code9.0997454.0697675
\n", "
" ], "text/plain": [ " prototype_id feature contribution_score interpreted_value rank\n", "0 3 condition_code 47.354488 6.712051 1\n", "1 3 data_arc 29.003140 253.67423 2\n", "2 3 h 15.000282 18.005536 3\n", "3 3 n_obs_used 14.650509 18.095537 4\n", "4 3 i 5.876609 10.6704 5\n", "5 0 om 2.823642 84.040274 1\n", "6 0 condition_code 1.396686 0.18725 2\n", "7 0 w 0.829923 205.94687 3\n", "8 0 n_obs_used 0.577087 199.988755 4\n", "9 0 n 0.460764 0.226685 5\n", "10 4 w 2.976050 88.600346 1\n", "11 4 om 2.472264 244.593422 2\n", "12 4 condition_code 1.299007 0.237537 3\n", "13 4 n_obs_used 0.579315 198.257712 4\n", "14 4 ma 0.565179 162.908708 5\n", "15 2 n_obs_used 16.539234 1149.485048 1\n", "16 2 data_arc 11.543415 12629.802255 2\n", "17 2 h 10.733020 13.951356 3\n", "18 2 q 4.311334 3.780162 4\n", "19 2 moid 4.298770 2.791547 5\n", "20 1 w 3.681780 277.27254 1\n", "21 1 om 3.497154 259.451502 2\n", "22 1 condition_code 1.538064 0.236619 3\n", "23 1 e 1.335589 0.124196 4\n", "24 1 i 0.946616 7.349716 5\n", "25 5 e 51.163396 0.594384 1\n", "26 5 h 29.822758 21.763982 2\n", "27 5 condition_code 16.744631 5.188867 3\n", "28 5 data_arc 11.857491 1152.310966 4\n", "29 5 neo 9.156986 Y 5\n", "30 7 h 39.831296 22.939038 1\n", "31 7 e 38.029636 0.457377 2\n", "32 7 n 30.939406 0.455755 3\n", "33 7 condition_code 15.224177 4.561035 4\n", "34 7 neo 11.453909 Y 5\n", "35 9 i 32.003468 34.471053 1\n", "36 9 n 26.911428 0.508083 2\n", "37 9 e 18.966981 0.370235 3\n", "38 9 h 17.865198 20.717401 4\n", "39 9 condition_code 9.099745 4.069767 5" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Вклады признаков внутри прототипов\n", "global_stats = {}\n", "for feat in X_trainval.columns:\n", " if pd.api.types.is_numeric_dtype(X_trainval[feat]):\n", " global_stats[feat] = {\n", " \"type\": \"num\",\n", " \"mean\": float(X_trainval[feat].mean()),\n", " \"std\": float(X_trainval[feat].std(ddof=0) if X_trainval[feat].std(ddof=0) > 1e-8 else 1.0)\n", " }\n", " else:\n", " global_mode = X_trainval[feat].mode(dropna=True)\n", " global_stats[feat] = {\n", " \"type\": \"cat\",\n", " \"mode\": global_mode.iloc[0] if len(global_mode) else np.nan,\n", " \"freq\": float((X_trainval[feat] == (global_mode.iloc[0] if len(global_mode) else np.nan)).mean()) if len(global_mode) else np.nan\n", " }\n", "\n", "contrib_rows = []\n", "important_proto_ids = prototype_df.head(min(8, len(prototype_df)))[\"prototype_id\"].tolist()\n", "\n", "for pid in important_proto_ids:\n", " mask = assign_idx_trainval == pid\n", " subset_X = X_trainval.loc[mask]\n", " if len(subset_X) == 0:\n", " continue\n", "\n", " proto_weight_abs = abs(coef[pid])\n", "\n", " feat_rows = []\n", " for feat in X_trainval.columns:\n", " if pd.api.types.is_numeric_dtype(subset_X[feat]):\n", " proto_mean = float(subset_X[feat].mean())\n", " score = abs(proto_mean - global_stats[feat][\"mean\"]) / global_stats[feat][\"std\"]\n", " interpreted_value = proto_mean\n", " else:\n", " mode_series = subset_X[feat].mode(dropna=True)\n", " proto_mode = mode_series.iloc[0] if len(mode_series) else np.nan\n", " proto_freq = float((subset_X[feat] == proto_mode).mean()) if pd.notna(proto_mode) else 0.0\n", " global_freq = float((X_trainval[feat] == proto_mode).mean()) if pd.notna(proto_mode) else 0.0\n", " score = abs(proto_freq - global_freq)\n", " interpreted_value = proto_mode\n", "\n", " feat_rows.append({\n", " \"prototype_id\": pid,\n", " \"feature\": feat,\n", " \"contribution_score\": float(score * proto_weight_abs),\n", " \"interpreted_value\": interpreted_value\n", " })\n", "\n", " feat_df = pd.DataFrame(feat_rows).sort_values(\"contribution_score\", ascending=False).head(5).reset_index(drop=True)\n", " feat_df[\"rank\"] = np.arange(1, len(feat_df) + 1)\n", " contrib_rows.append(feat_df)\n", "\n", "prototype_contrib_df = pd.concat(contrib_rows, ignore_index=True) if contrib_rows else pd.DataFrame()\n", "\n", "print(\"Топ-5 вкладов признаков для основных прототипов:\")\n", "display(prototype_contrib_df)" ] }, { "cell_type": "code", "execution_count": 15, "id": "90a44731", "metadata": {}, "outputs": [ { "data": { "image/png": "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", "text/plain": [ "
" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Иллюстрация качества\n", "fig, axes = plt.subplots(1, 2, figsize=(13, 5))\n", "\n", "axes[0].bar(results_df[\"split\"], results_df[\"f1_macro\"])\n", "axes[0].set_title(\"F1-macro\")\n", "axes[0].set_ylim(0, 1)\n", "\n", "axes[1].bar(results_df[\"split\"], results_df[\"roc_auc\"])\n", "axes[1].set_title(\"ROC-AUC\")\n", "axes[1].set_ylim(0, 1)\n", "\n", "plt.tight_layout()\n", "plt.show()" ] }, { "cell_type": "code", "execution_count": 16, "id": "63615e54", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Первые 10 фактических и предсказанных значений на test:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
object_namey_truey_predy_pred_proba
0nan000.000041
1nan000.000116
2nan000.000245
3nan000.000022
4nan000.000150
5nan000.000084
6nan000.000121
7Dinotina000.001224
8nan000.000040
9nan000.000272
\n", "
" ], "text/plain": [ " object_name y_true y_pred y_pred_proba\n", "0 nan 0 0 0.000041\n", "1 nan 0 0 0.000116\n", "2 nan 0 0 0.000245\n", "3 nan 0 0 0.000022\n", "4 nan 0 0 0.000150\n", "5 nan 0 0 0.000084\n", "6 nan 0 0 0.000121\n", "7 Dinotina 0 0 0.001224\n", "8 nan 0 0 0.000040\n", "9 nan 0 0 0.000272" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Примеры предсказаний на test\n", "pred_df = pd.DataFrame({\n", " \"object_name\": pd.Series(names_test).astype(str).values,\n", " \"y_true\": y_test.values,\n", " \"y_pred\": test_pred,\n", " \"y_pred_proba\": test_proba\n", "})\n", "print(\"Первые 10 фактических и предсказанных значений на test:\")\n", "display(pred_df.head(10))" ] }, { "cell_type": "code", "execution_count": 17, "id": "5c82aace", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "=== КЛЮЧЕВАЯ ИНФОРМАЦИЯ ДЛЯ ВЫВОДА ===\n", "Зависимая переменная: pha\n", "\n", "Число исходных признаков: 18\n", "Число признаков после кодирования: 30\n", "Лучшая конфигурация прототипной сети:\n", "n_prototypes_per_class = 5\n", "total_prototypes = 10\n", "sigma_multiplier = 2.0\n", "sigma = 6.271090984344482\n", "C = 2.0\n", "\n", "Распределение классов:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
share
pha
00.997552
10.002448
\n", "
" ], "text/plain": [ " share\n", "pha \n", "0 0.997552\n", "1 0.002448" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Использованные исходные признаки:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
feature
0a
1e
2i
3om
4w
5q
6ad
7per_y
8data_arc
9condition_code
10n_obs_used
11h
12moid
13n
14per
15ma
16neo
17class
\n", "
" ], "text/plain": [ " feature\n", "0 a\n", "1 e\n", "2 i\n", "3 om\n", "4 w\n", "5 q\n", "6 ad\n", "7 per_y\n", "8 data_arc\n", "9 condition_code\n", "10 n_obs_used\n", "11 h\n", "12 moid\n", "13 n\n", "14 per\n", "15 ma\n", "16 neo\n", "17 class" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Метрики модели:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
splitaccuracybalanced_accuracyprecision_macrorecall_macrof1_macroroc_auc
0trainval0.9793750.9886440.5530000.9886440.5906030.995785
1test0.9793200.9855460.5524160.9855460.5895910.996248
\n", "
" ], "text/plain": [ " split accuracy balanced_accuracy precision_macro recall_macro \\\n", "0 trainval 0.979375 0.988644 0.553000 0.988644 \n", "1 test 0.979320 0.985546 0.552416 0.985546 \n", "\n", " f1_macro roc_auc \n", "0 0.590603 0.995785 \n", "1 0.589591 0.996248 " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Топ-10 конфигураций по validation:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
n_prototypes_per_classtotal_prototypessigma_multipliersigmaCval_accuracyval_balanced_accuracyval_precision_macroval_recall_macroval_f1_macroval_roc_auc
05102.05.9481372.00.979340.9896450.5528140.9896450.5903050.996080
15102.05.9481371.00.978900.9894240.5518270.9894240.5885740.996067
2482.05.8079962.00.978600.9892740.5511740.9892740.5874250.995346
35102.05.9481370.50.978240.9890930.5504130.9890930.5860780.995906
4481.02.9039982.00.978220.9849950.5500310.9849950.5853910.995502
5482.05.8079961.00.978020.9889830.5499590.9889830.5852720.995145
6361.03.7456382.00.977860.9889030.5496340.9889030.5846930.994570
7481.02.9039981.00.977700.9888230.5493130.9888230.5841200.995309
8482.05.8079960.50.977680.9888130.5492730.9888130.5840490.994914
9362.07.4912762.00.977580.9887630.5490750.9887630.5836950.993315
\n", "
" ], "text/plain": [ " n_prototypes_per_class total_prototypes sigma_multiplier sigma C \\\n", "0 5 10 2.0 5.948137 2.0 \n", "1 5 10 2.0 5.948137 1.0 \n", "2 4 8 2.0 5.807996 2.0 \n", "3 5 10 2.0 5.948137 0.5 \n", "4 4 8 1.0 2.903998 2.0 \n", "5 4 8 2.0 5.807996 1.0 \n", "6 3 6 1.0 3.745638 2.0 \n", "7 4 8 1.0 2.903998 1.0 \n", "8 4 8 2.0 5.807996 0.5 \n", "9 3 6 2.0 7.491276 2.0 \n", "\n", " val_accuracy val_balanced_accuracy val_precision_macro val_recall_macro \\\n", "0 0.97934 0.989645 0.552814 0.989645 \n", "1 0.97890 0.989424 0.551827 0.989424 \n", "2 0.97860 0.989274 0.551174 0.989274 \n", "3 0.97824 0.989093 0.550413 0.989093 \n", "4 0.97822 0.984995 0.550031 0.984995 \n", "5 0.97802 0.988983 0.549959 0.988983 \n", "6 0.97786 0.988903 0.549634 0.988903 \n", "7 0.97770 0.988823 0.549313 0.988823 \n", "8 0.97768 0.988813 0.549273 0.988813 \n", "9 0.97758 0.988763 0.549075 0.988763 \n", "\n", " val_f1_macro val_roc_auc \n", "0 0.590305 0.996080 \n", "1 0.588574 0.996067 \n", "2 0.587425 0.995346 \n", "3 0.586078 0.995906 \n", "4 0.585391 0.995502 \n", "5 0.585272 0.995145 \n", "6 0.584693 0.994570 \n", "7 0.584120 0.995309 \n", "8 0.584049 0.994914 \n", "9 0.583695 0.993315 " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Топ-10 признаков по важности:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
featurepermutation_importance_meanpermutation_importance_std
0e0.0476640.002647
1condition_code0.0078150.000350
2data_arc0.0036830.000315
3n0.0022030.003038
4n_obs_used0.0015050.000192
5ma0.0004220.000162
6w0.0000260.000099
7per0.0000000.000000
8per_y0.0000000.000000
9ad-0.0000260.000037
\n", "
" ], "text/plain": [ " feature permutation_importance_mean permutation_importance_std\n", "0 e 0.047664 0.002647\n", "1 condition_code 0.007815 0.000350\n", "2 data_arc 0.003683 0.000315\n", "3 n 0.002203 0.003038\n", "4 n_obs_used 0.001505 0.000192\n", "5 ma 0.000422 0.000162\n", "6 w 0.000026 0.000099\n", "7 per 0.000000 0.000000\n", "8 per_y 0.000000 0.000000\n", "9 ad -0.000026 0.000037" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Основные прототипы:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
prototype_idprototype_classsupport_nsupport_share_trainvalprototype_output_weightimpact_scoreassigned_positive_rate_trainvalnearest_objectsecondition_codedata_arcnn_obs_usedma
030276750.138375-22.3402983.0913390.000000nan, nan, nan0.1737806.712051253.6742300.23730218.095537181.147050
100717650.358825-3.4363011.2330310.000000nan, nan, nan0.1438260.1872505927.5718530.226685199.988755189.886354
240487250.243625-3.3567720.8177940.000000nan, nan, nan0.1552310.2375375965.4589020.233649198.257712162.908708
320191610.0958056.9531260.6661440.000052nan, nan, nan0.1376370.20652412629.8022550.2179451149.485048176.631783
410272210.136105-3.9708710.5404550.000000nan, nan, nan0.1241960.2366196007.0144740.235234199.583740192.906197
55120120.01006010.9443900.1101010.086481nan, nan, nan0.5943845.1888671152.3109660.26427792.923956172.768315
67112370.00618511.8300920.0731690.099434nan, nan, nan0.4573774.5610351747.5696720.455755109.617623176.214085
7918170.0040858.2973590.0338950.083231nan, nan, nan0.3702354.0697672102.9447170.508083116.062424174.708108
8811900.0009507.4017950.0070320.178947nan, nan, nan0.4100744.2789471910.5978841.374819152.363158172.634906
96111970.0059850.6844610.0040960.075188nan, nan, nan0.2473855.0359231103.2247670.83558986.230576176.160290
\n", "
" ], "text/plain": [ " prototype_id prototype_class support_n support_share_trainval \\\n", "0 3 0 27675 0.138375 \n", "1 0 0 71765 0.358825 \n", "2 4 0 48725 0.243625 \n", "3 2 0 19161 0.095805 \n", "4 1 0 27221 0.136105 \n", "5 5 1 2012 0.010060 \n", "6 7 1 1237 0.006185 \n", "7 9 1 817 0.004085 \n", "8 8 1 190 0.000950 \n", "9 6 1 1197 0.005985 \n", "\n", " prototype_output_weight impact_score assigned_positive_rate_trainval \\\n", "0 -22.340298 3.091339 0.000000 \n", "1 -3.436301 1.233031 0.000000 \n", "2 -3.356772 0.817794 0.000000 \n", "3 6.953126 0.666144 0.000052 \n", "4 -3.970871 0.540455 0.000000 \n", "5 10.944390 0.110101 0.086481 \n", "6 11.830092 0.073169 0.099434 \n", "7 8.297359 0.033895 0.083231 \n", "8 7.401795 0.007032 0.178947 \n", "9 0.684461 0.004096 0.075188 \n", "\n", " nearest_objects e condition_code data_arc n \\\n", "0 nan, nan, nan 0.173780 6.712051 253.674230 0.237302 \n", "1 nan, nan, nan 0.143826 0.187250 5927.571853 0.226685 \n", "2 nan, nan, nan 0.155231 0.237537 5965.458902 0.233649 \n", "3 nan, nan, nan 0.137637 0.206524 12629.802255 0.217945 \n", "4 nan, nan, nan 0.124196 0.236619 6007.014474 0.235234 \n", "5 nan, nan, nan 0.594384 5.188867 1152.310966 0.264277 \n", "6 nan, nan, nan 0.457377 4.561035 1747.569672 0.455755 \n", "7 nan, nan, nan 0.370235 4.069767 2102.944717 0.508083 \n", "8 nan, nan, nan 0.410074 4.278947 1910.597884 1.374819 \n", "9 nan, nan, nan 0.247385 5.035923 1103.224767 0.835589 \n", "\n", " n_obs_used ma \n", "0 18.095537 181.147050 \n", "1 199.988755 189.886354 \n", "2 198.257712 162.908708 \n", "3 1149.485048 176.631783 \n", "4 199.583740 192.906197 \n", "5 92.923956 172.768315 \n", "6 109.617623 176.214085 \n", "7 116.062424 174.708108 \n", "8 152.363158 172.634906 \n", "9 86.230576 176.160290 " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Топ-5 вкладов признаков для основных прототипов:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
prototype_idfeaturecontribution_scoreinterpreted_valuerank
03condition_code47.3544886.7120511
13data_arc29.003140253.674232
23h15.00028218.0055363
33n_obs_used14.65050918.0955374
43i5.87660910.67045
50om2.82364284.0402741
60condition_code1.3966860.187252
70w0.829923205.946873
80n_obs_used0.577087199.9887554
90n0.4607640.2266855
104w2.97605088.6003461
114om2.472264244.5934222
124condition_code1.2990070.2375373
134n_obs_used0.579315198.2577124
144ma0.565179162.9087085
152n_obs_used16.5392341149.4850481
162data_arc11.54341512629.8022552
172h10.73302013.9513563
182q4.3113343.7801624
192moid4.2987702.7915475
201w3.681780277.272541
211om3.497154259.4515022
221condition_code1.5380640.2366193
231e1.3355890.1241964
241i0.9466167.3497165
255e51.1633960.5943841
265h29.82275821.7639822
275condition_code16.7446315.1888673
285data_arc11.8574911152.3109664
295neo9.156986Y5
307h39.83129622.9390381
317e38.0296360.4573772
327n30.9394060.4557553
337condition_code15.2241774.5610354
347neo11.453909Y5
359i32.00346834.4710531
369n26.9114280.5080832
379e18.9669810.3702353
389h17.86519820.7174014
399condition_code9.0997454.0697675
\n", "
" ], "text/plain": [ " prototype_id feature contribution_score interpreted_value rank\n", "0 3 condition_code 47.354488 6.712051 1\n", "1 3 data_arc 29.003140 253.67423 2\n", "2 3 h 15.000282 18.005536 3\n", "3 3 n_obs_used 14.650509 18.095537 4\n", "4 3 i 5.876609 10.6704 5\n", "5 0 om 2.823642 84.040274 1\n", "6 0 condition_code 1.396686 0.18725 2\n", "7 0 w 0.829923 205.94687 3\n", "8 0 n_obs_used 0.577087 199.988755 4\n", "9 0 n 0.460764 0.226685 5\n", "10 4 w 2.976050 88.600346 1\n", "11 4 om 2.472264 244.593422 2\n", "12 4 condition_code 1.299007 0.237537 3\n", "13 4 n_obs_used 0.579315 198.257712 4\n", "14 4 ma 0.565179 162.908708 5\n", "15 2 n_obs_used 16.539234 1149.485048 1\n", "16 2 data_arc 11.543415 12629.802255 2\n", "17 2 h 10.733020 13.951356 3\n", "18 2 q 4.311334 3.780162 4\n", "19 2 moid 4.298770 2.791547 5\n", "20 1 w 3.681780 277.27254 1\n", "21 1 om 3.497154 259.451502 2\n", "22 1 condition_code 1.538064 0.236619 3\n", "23 1 e 1.335589 0.124196 4\n", "24 1 i 0.946616 7.349716 5\n", "25 5 e 51.163396 0.594384 1\n", "26 5 h 29.822758 21.763982 2\n", "27 5 condition_code 16.744631 5.188867 3\n", "28 5 data_arc 11.857491 1152.310966 4\n", "29 5 neo 9.156986 Y 5\n", "30 7 h 39.831296 22.939038 1\n", "31 7 e 38.029636 0.457377 2\n", "32 7 n 30.939406 0.455755 3\n", "33 7 condition_code 15.224177 4.561035 4\n", "34 7 neo 11.453909 Y 5\n", "35 9 i 32.003468 34.471053 1\n", "36 9 n 26.911428 0.508083 2\n", "37 9 e 18.966981 0.370235 3\n", "38 9 h 17.865198 20.717401 4\n", "39 9 condition_code 9.099745 4.069767 5" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Матрица ошибок на test:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
01
0488451033
11121
\n", "
" ], "text/plain": [ " 0 1\n", "0 48845 1033\n", "1 1 121" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Первые 10 фактических и предсказанных значений на test:\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
object_namey_truey_predy_pred_proba
0nan000.000041
1nan000.000116
2nan000.000245
3nan000.000022
4nan000.000150
5nan000.000084
6nan000.000121
7Dinotina000.001224
8nan000.000040
9nan000.000272
\n", "
" ], "text/plain": [ " object_name y_true y_pred y_pred_proba\n", "0 nan 0 0 0.000041\n", "1 nan 0 0 0.000116\n", "2 nan 0 0 0.000245\n", "3 nan 0 0 0.000022\n", "4 nan 0 0 0.000150\n", "5 nan 0 0 0.000084\n", "6 nan 0 0 0.000121\n", "7 Dinotina 0 0 0.001224\n", "8 nan 0 0 0.000040\n", "9 nan 0 0 0.000272" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# === КЛЮЧЕВАЯ ИНФОРМАЦИЯ ДЛЯ ВЫВОДА ===\n", "print(\"=== КЛЮЧЕВАЯ ИНФОРМАЦИЯ ДЛЯ ВЫВОДА ===\")\n", "print(f\"Зависимая переменная: {target_col}\")\n", "print()\n", "print(f\"Число исходных признаков: {X.shape[1]}\")\n", "print(f\"Число признаков после кодирования: {X_train_p.shape[1]}\")\n", "print(f\"Лучшая конфигурация прототипной сети:\")\n", "print(f\"n_prototypes_per_class = {int(best_params['n_prototypes_per_class'])}\")\n", "print(f\"total_prototypes = {int(prototype_df.shape[0])}\")\n", "print(f\"sigma_multiplier = {float(best_params['sigma_multiplier'])}\")\n", "print(f\"sigma = {float(final_model.sigma_)}\")\n", "print(f\"C = {float(best_params['C'])}\")\n", "print()\n", "\n", "print(\"Распределение классов:\")\n", "display(y.value_counts(normalize=True).sort_index().rename(\"share\").to_frame())\n", "\n", "print(\"Использованные исходные признаки:\")\n", "display(pd.DataFrame({\"feature\": X.columns.tolist()}))\n", "\n", "print(\"Метрики модели:\")\n", "display(results_df)\n", "\n", "print(\"Топ-10 конфигураций по validation:\")\n", "display(search_df.head(10))\n", "\n", "print(\"Топ-10 признаков по важности:\")\n", "display(perm_df.head(10))\n", "\n", "print(\"Основные прототипы:\")\n", "display(prototype_df.head(10))\n", "\n", "print(\"Топ-5 вкладов признаков для основных прототипов:\")\n", "display(prototype_contrib_df)\n", "\n", "print(\"Матрица ошибок на test:\")\n", "display(pd.DataFrame(cm))\n", "\n", "print(\"Первые 10 фактических и предсказанных значений на test:\")\n", "display(pred_df.head(10))" ] }, { "cell_type": "markdown", "id": "aaf705be", "metadata": {}, "source": [ "Итог\n", "\n", "Зависимой переменной выбрана pha. Прототипная нейросеть с 10 прототипами (по 5 на класс) показала на test accuracy = 0.979, balanced accuracy = 0.986, F1-macro = 0.590, ROC-AUC = 0.996. С учётом сильного дисбаланса классов (0.245% опасных объектов) модель даёт высокое качество именно по сбалансированным метрикам, а не только по accuracy.\n", "\n", "Наиболее важные признаки: e, condition_code, data_arc, n, n_obs_used. Для опасных астероидов прототипы характеризуются, как правило, более высокой эксцентриситетностью, повышенными значениями condition_code, меньшей наблюдательной историей (data_arc, n_obs_used) и признаками класса neo = Y. Для неопасных объектов характерны более стабильные орбитальные параметры и значительно большая представленность в обучающей выборке.\n", "\n", "Следовательно, прототипная сеть выделяет интерпретируемые группы астероидов, а ключевой вклад в распознавание опасных объектов вносят орбитальная эксцентриситетность и характеристики наблюдаемости орбиты." ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "name": "python" } }, "nbformat": 4, "nbformat_minor": 5 }