{ "cells": [ { "cell_type": "markdown", "id": "5c1bd341", "metadata": {}, "source": [ "# Классификация объектов KOI глубокой нейросетью с PSO и дообучением\n", "\n", "В работе автоматически выбирается ключевая зависимая переменная для задачи классификации, далее строится глубокая нейросеть с количеством скрытых слоёв больше 3. \n", "Сначала веса подбираются методом **PSO (Particle Swarm Optimization)**, затем модель **дообучается градиентным методом**. \n", "В конце выводятся все ключевые метрики на тестовой выборке для написания итогового вывода.\n" ] }, { "cell_type": "code", "execution_count": 1, "id": "5528408b", "metadata": {}, "outputs": [], "source": [ "\n", "import os\n", "import re\n", "import copy\n", "import math\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.compose import ColumnTransformer\n", "from sklearn.impute import SimpleImputer\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", " log_loss,\n", ")\n", "from sklearn.model_selection import train_test_split\n", "from sklearn.neural_network import MLPClassifier\n", "from sklearn.pipeline import Pipeline\n", "from sklearn.preprocessing import OneHotEncoder, StandardScaler, LabelEncoder\n", "\n", "warnings.filterwarnings(\"ignore\")\n", "RANDOM_STATE = 42\n", "np.random.seed(RANDOM_STATE)\n" ] }, { "cell_type": "code", "execution_count": 2, "id": "b4170ddb", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Используемый файл: NASA Exoplanet.csv\n", "Форма датасета: (9564, 49)\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", "
kepidkepoi_namekepler_namekoi_dispositionkoi_pdispositionkoi_scorekoi_fpflag_ntkoi_fpflag_sskoi_fpflag_cokoi_fpflag_ec...koi_steff_err2koi_sloggkoi_slogg_err1koi_slogg_err2koi_sradkoi_srad_err1koi_srad_err2radeckoi_kepmag
010797460K00752.01Kepler-227 bCONFIRMEDCANDIDATE1.0000000...-81.04.4670.064-0.0960.9270.105-0.061291.9342348.14165115.347
110797460K00752.02Kepler-227 cCONFIRMEDCANDIDATE0.9690000...-81.04.4670.064-0.0960.9270.105-0.061291.9342348.14165115.347
210811496K00753.01NaNCANDIDATECANDIDATE0.0000000...-176.04.5440.044-0.1760.8680.233-0.078297.0048248.13412915.436
310848459K00754.01NaNFALSE POSITIVEFALSE POSITIVE0.0000100...-174.04.5640.053-0.1680.7910.201-0.067285.5346148.28521015.597
410854555K00755.01Kepler-664 bCONFIRMEDCANDIDATE1.0000000...-211.04.4380.070-0.2101.0460.334-0.133288.7548848.22620015.509
\n", "

5 rows × 49 columns

\n", "
" ], "text/plain": [ " kepid kepoi_name kepler_name koi_disposition koi_pdisposition \\\n", "0 10797460 K00752.01 Kepler-227 b CONFIRMED CANDIDATE \n", "1 10797460 K00752.02 Kepler-227 c CONFIRMED CANDIDATE \n", "2 10811496 K00753.01 NaN CANDIDATE CANDIDATE \n", "3 10848459 K00754.01 NaN FALSE POSITIVE FALSE POSITIVE \n", "4 10854555 K00755.01 Kepler-664 b CONFIRMED CANDIDATE \n", "\n", " koi_score koi_fpflag_nt koi_fpflag_ss koi_fpflag_co koi_fpflag_ec ... \\\n", "0 1.000 0 0 0 0 ... \n", "1 0.969 0 0 0 0 ... \n", "2 0.000 0 0 0 0 ... \n", "3 0.000 0 1 0 0 ... \n", "4 1.000 0 0 0 0 ... \n", "\n", " koi_steff_err2 koi_slogg koi_slogg_err1 koi_slogg_err2 koi_srad \\\n", "0 -81.0 4.467 0.064 -0.096 0.927 \n", "1 -81.0 4.467 0.064 -0.096 0.927 \n", "2 -176.0 4.544 0.044 -0.176 0.868 \n", "3 -174.0 4.564 0.053 -0.168 0.791 \n", "4 -211.0 4.438 0.070 -0.210 1.046 \n", "\n", " koi_srad_err1 koi_srad_err2 ra dec koi_kepmag \n", "0 0.105 -0.061 291.93423 48.141651 15.347 \n", "1 0.105 -0.061 291.93423 48.141651 15.347 \n", "2 0.233 -0.078 297.00482 48.134129 15.436 \n", "3 0.201 -0.067 285.53461 48.285210 15.597 \n", "4 0.334 -0.133 288.75488 48.226200 15.509 \n", "\n", "[5 rows x 49 columns]" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "\n", "Типы данных:\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", "
dtype
kepidint64
kepoi_nameobject
kepler_nameobject
koi_dispositionobject
koi_pdispositionobject
koi_scorefloat64
koi_fpflag_ntint64
koi_fpflag_ssint64
koi_fpflag_coint64
koi_fpflag_ecint64
koi_periodfloat64
koi_period_err1float64
koi_period_err2float64
koi_time0bkfloat64
koi_time0bk_err1float64
koi_time0bk_err2float64
koi_impactfloat64
koi_impact_err1float64
koi_impact_err2float64
koi_durationfloat64
koi_duration_err1float64
koi_duration_err2float64
koi_depthfloat64
koi_depth_err1float64
koi_depth_err2float64
koi_pradfloat64
koi_prad_err1float64
koi_prad_err2float64
koi_teqfloat64
koi_teq_err1float64
koi_teq_err2float64
koi_insolfloat64
koi_insol_err1float64
koi_insol_err2float64
koi_model_snrfloat64
koi_tce_plnt_numfloat64
koi_tce_delivnameobject
koi_stefffloat64
koi_steff_err1float64
koi_steff_err2float64
koi_sloggfloat64
koi_slogg_err1float64
koi_slogg_err2float64
koi_sradfloat64
koi_srad_err1float64
koi_srad_err2float64
rafloat64
decfloat64
koi_kepmagfloat64
\n", "
" ], "text/plain": [ " dtype\n", "kepid int64\n", "kepoi_name object\n", "kepler_name object\n", "koi_disposition object\n", "koi_pdisposition object\n", "koi_score float64\n", "koi_fpflag_nt int64\n", "koi_fpflag_ss int64\n", "koi_fpflag_co int64\n", "koi_fpflag_ec int64\n", "koi_period float64\n", "koi_period_err1 float64\n", "koi_period_err2 float64\n", "koi_time0bk float64\n", "koi_time0bk_err1 float64\n", "koi_time0bk_err2 float64\n", "koi_impact float64\n", "koi_impact_err1 float64\n", "koi_impact_err2 float64\n", "koi_duration float64\n", "koi_duration_err1 float64\n", "koi_duration_err2 float64\n", "koi_depth float64\n", "koi_depth_err1 float64\n", "koi_depth_err2 float64\n", "koi_prad float64\n", "koi_prad_err1 float64\n", "koi_prad_err2 float64\n", "koi_teq float64\n", "koi_teq_err1 float64\n", "koi_teq_err2 float64\n", "koi_insol float64\n", "koi_insol_err1 float64\n", "koi_insol_err2 float64\n", "koi_model_snr float64\n", "koi_tce_plnt_num float64\n", "koi_tce_delivname object\n", "koi_steff float64\n", "koi_steff_err1 float64\n", "koi_steff_err2 float64\n", "koi_slogg float64\n", "koi_slogg_err1 float64\n", "koi_slogg_err2 float64\n", "koi_srad float64\n", "koi_srad_err1 float64\n", "koi_srad_err2 float64\n", "ra float64\n", "dec float64\n", "koi_kepmag float64" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "def find_csv_file():\n", " candidates = [\n", " \"NASA Exoplanet.csv\",\n", " \"nasa_exoplanet.csv\",\n", " \"NASA_Exoplanet.csv\",\n", " \"koi.csv\",\n", " \"kepler.csv\",\n", " ]\n", " for name in candidates:\n", " if Path(name).exists():\n", " return name\n", "\n", " csv_files = sorted(Path(\".\").glob(\"*.csv\"))\n", " preferred = []\n", " for p in csv_files:\n", " low = p.name.lower()\n", " if any(key in low for key in [\"nasa\", \"exo\", \"kepler\", \"koi\"]):\n", " preferred.append(p)\n", " if preferred:\n", " return preferred[0].as_posix()\n", " if csv_files:\n", " return csv_files[0].as_posix()\n", " raise FileNotFoundError(\"CSV-файл не найден в текущей папке.\")\n", "\n", "csv_path = find_csv_file()\n", "print(\"Используемый файл:\", csv_path)\n", "\n", "df = pd.read_csv(csv_path)\n", "print(\"Форма датасета:\", df.shape)\n", "display(df.head())\n", "print(\"\\nТипы данных:\")\n", "display(df.dtypes.to_frame(\"dtype\"))\n" ] }, { "cell_type": "code", "execution_count": 3, "id": "5cc71ee1", "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", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \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
0kepidkepid
1kepoi_namekepoi_name
2kepler_namekepler_name
3koi_dispositionkoi_disposition
4koi_pdispositionkoi_pdisposition
5koi_scorekoi_score
6koi_fpflag_ntkoi_fpflag_nt
7koi_fpflag_sskoi_fpflag_ss
8koi_fpflag_cokoi_fpflag_co
9koi_fpflag_eckoi_fpflag_ec
10koi_periodkoi_period
11koi_period_err1koi_period_err1
12koi_period_err2koi_period_err2
13koi_time0bkkoi_time0bk
14koi_time0bk_err1koi_time0bk_err1
15koi_time0bk_err2koi_time0bk_err2
16koi_impactkoi_impact
17koi_impact_err1koi_impact_err1
18koi_impact_err2koi_impact_err2
19koi_durationkoi_duration
20koi_duration_err1koi_duration_err1
21koi_duration_err2koi_duration_err2
22koi_depthkoi_depth
23koi_depth_err1koi_depth_err1
24koi_depth_err2koi_depth_err2
25koi_pradkoi_prad
26koi_prad_err1koi_prad_err1
27koi_prad_err2koi_prad_err2
28koi_teqkoi_teq
29koi_teq_err1koi_teq_err1
30koi_teq_err2koi_teq_err2
31koi_insolkoi_insol
32koi_insol_err1koi_insol_err1
33koi_insol_err2koi_insol_err2
34koi_model_snrkoi_model_snr
35koi_tce_plnt_numkoi_tce_plnt_num
36koi_tce_delivnamekoi_tce_delivname
37koi_steffkoi_steff
38koi_steff_err1koi_steff_err1
39koi_steff_err2koi_steff_err2
40koi_sloggkoi_slogg
41koi_slogg_err1koi_slogg_err1
42koi_slogg_err2koi_slogg_err2
43koi_sradkoi_srad
44koi_srad_err1koi_srad_err1
45koi_srad_err2koi_srad_err2
46rara
47decdec
48koi_kepmagkoi_kepmag
\n", "
" ], "text/plain": [ " original_name normalized_name\n", "0 kepid kepid\n", "1 kepoi_name kepoi_name\n", "2 kepler_name kepler_name\n", "3 koi_disposition koi_disposition\n", "4 koi_pdisposition koi_pdisposition\n", "5 koi_score koi_score\n", "6 koi_fpflag_nt koi_fpflag_nt\n", "7 koi_fpflag_ss koi_fpflag_ss\n", "8 koi_fpflag_co koi_fpflag_co\n", "9 koi_fpflag_ec koi_fpflag_ec\n", "10 koi_period koi_period\n", "11 koi_period_err1 koi_period_err1\n", "12 koi_period_err2 koi_period_err2\n", "13 koi_time0bk koi_time0bk\n", "14 koi_time0bk_err1 koi_time0bk_err1\n", "15 koi_time0bk_err2 koi_time0bk_err2\n", "16 koi_impact koi_impact\n", "17 koi_impact_err1 koi_impact_err1\n", "18 koi_impact_err2 koi_impact_err2\n", "19 koi_duration koi_duration\n", "20 koi_duration_err1 koi_duration_err1\n", "21 koi_duration_err2 koi_duration_err2\n", "22 koi_depth koi_depth\n", "23 koi_depth_err1 koi_depth_err1\n", "24 koi_depth_err2 koi_depth_err2\n", "25 koi_prad koi_prad\n", "26 koi_prad_err1 koi_prad_err1\n", "27 koi_prad_err2 koi_prad_err2\n", "28 koi_teq koi_teq\n", "29 koi_teq_err1 koi_teq_err1\n", "30 koi_teq_err2 koi_teq_err2\n", "31 koi_insol koi_insol\n", "32 koi_insol_err1 koi_insol_err1\n", "33 koi_insol_err2 koi_insol_err2\n", "34 koi_model_snr koi_model_snr\n", "35 koi_tce_plnt_num koi_tce_plnt_num\n", "36 koi_tce_delivname koi_tce_delivname\n", "37 koi_steff koi_steff\n", "38 koi_steff_err1 koi_steff_err1\n", "39 koi_steff_err2 koi_steff_err2\n", "40 koi_slogg koi_slogg\n", "41 koi_slogg_err1 koi_slogg_err1\n", "42 koi_slogg_err2 koi_slogg_err2\n", "43 koi_srad koi_srad\n", "44 koi_srad_err1 koi_srad_err1\n", "45 koi_srad_err2 koi_srad_err2\n", "46 ra ra\n", "47 dec dec\n", "48 koi_kepmag koi_kepmag" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "def normalize_col_name(col):\n", " col = str(col).strip().lower()\n", " col = col.replace(\"%\", \"percent\")\n", " col = re.sub(r\"[^0-9a-zA-Zа-яА-Я]+\", \"_\", col)\n", " col = re.sub(r\"_+\", \"_\", col).strip(\"_\")\n", " return col\n", "\n", "original_cols = df.columns.tolist()\n", "normalized_cols = [normalize_col_name(c) for c in original_cols]\n", "mapping_df = pd.DataFrame({\"original_name\": original_cols, \"normalized_name\": normalized_cols})\n", "display(mapping_df)\n", "\n", "df.columns = normalized_cols\n" ] }, { "cell_type": "code", "execution_count": 4, "id": "680f5053", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Выбранная зависимая переменная: koi_pdisposition\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", "
scoren_uniquecolumn
01002koi_pdisposition
11003koi_disposition
2202koi_fpflag_co
3202koi_fpflag_ec
4202koi_fpflag_ss
5203koi_fpflag_nt
6203koi_tce_delivname
7208koi_tce_plnt_num
802745kepler_name
9-1009564kepoi_name
\n", "
" ], "text/plain": [ " score n_unique column\n", "0 100 2 koi_pdisposition\n", "1 100 3 koi_disposition\n", "2 20 2 koi_fpflag_co\n", "3 20 2 koi_fpflag_ec\n", "4 20 2 koi_fpflag_ss\n", "5 20 3 koi_fpflag_nt\n", "6 20 3 koi_tce_delivname\n", "7 20 8 koi_tce_plnt_num\n", "8 0 2745 kepler_name\n", "9 -100 9564 kepoi_name" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "def choose_target_column(dataframe):\n", " cols = dataframe.columns.tolist()\n", " priority_patterns = [\n", " r\"disposition\",\n", " r\"target\",\n", " r\"label\",\n", " r\"class$\",\n", " r\"status\",\n", " r\"outcome\",\n", " r\"risk\",\n", " r\"quality\",\n", " ]\n", "\n", " candidates = []\n", " n = len(dataframe)\n", "\n", " for col in cols:\n", " s = dataframe[col]\n", " nunique = s.dropna().nunique()\n", " if nunique < 2:\n", " continue\n", "\n", " is_class_like = (\n", " s.dtype == \"object\"\n", " or str(s.dtype).startswith(\"category\")\n", " or nunique <= 20\n", " )\n", " if not is_class_like:\n", " continue\n", "\n", " score = 0\n", " for i, pat in enumerate(priority_patterns[::-1], start=1):\n", " if re.search(pat, col):\n", " score += 10 * i\n", "\n", " # более предпочтительны столбцы с небольшим числом классов\n", " if 2 <= nunique <= 10:\n", " score += 20\n", " elif nunique <= 20:\n", " score += 10\n", "\n", " # целевая переменная не должна быть почти уникальной\n", " if nunique / max(n, 1) > 0.3:\n", " score -= 100\n", "\n", " candidates.append((score, nunique, col))\n", "\n", " if not candidates:\n", " raise ValueError(\"Не удалось автоматически определить зависимую переменную для классификации.\")\n", "\n", " candidates = sorted(candidates, key=lambda x: (-x[0], x[1], x[2]))\n", " return candidates[0][2], pd.DataFrame(candidates, columns=[\"score\", \"n_unique\", \"column\"])\n", "\n", "target_col, target_candidates = choose_target_column(df)\n", "print(\"Выбранная зависимая переменная:\", target_col)\n", "display(target_candidates.head(15))\n" ] }, { "cell_type": "code", "execution_count": 5, "id": "675462f3", "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", "
class_indexclass_label
00CANDIDATE
11FALSE POSITIVE
\n", "
" ], "text/plain": [ " class_index class_label\n", "0 0 CANDIDATE\n", "1 1 FALSE POSITIVE" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Распределение классов:\n" ] }, { "data": { "text/plain": [ "koi_pdisposition\n", "CANDIDATE 0.493204\n", "FALSE POSITIVE 0.506796\n", "Name: share, dtype: float64" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "# Очистка и подготовка целевой переменной\n", "target_raw = df[target_col].copy()\n", "\n", "if target_raw.dtype == \"object\":\n", " target_clean = (\n", " target_raw.astype(str)\n", " .str.strip()\n", " .replace({\"\": np.nan, \"nan\": np.nan, \"None\": np.nan, \"none\": np.nan, \"NULL\": np.nan, \"null\": np.nan})\n", " )\n", "else:\n", " target_clean = target_raw.copy()\n", "\n", "mask_target = target_clean.notna()\n", "df = df.loc[mask_target].reset_index(drop=True)\n", "target_clean = target_clean.loc[mask_target].reset_index(drop=True)\n", "\n", "label_encoder = LabelEncoder()\n", "y = label_encoder.fit_transform(target_clean.astype(str))\n", "\n", "class_mapping = pd.DataFrame({\n", " \"class_index\": np.arange(len(label_encoder.classes_)),\n", " \"class_label\": label_encoder.classes_\n", "})\n", "display(class_mapping)\n", "\n", "class_share = pd.Series(target_clean).value_counts(normalize=True).sort_index().rename(\"share\")\n", "print(\"Распределение классов:\")\n", "display(class_share)\n" ] }, { "cell_type": "code", "execution_count": 6, "id": "0ff25260", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Удаляемые идентификаторы / несодержательные поля: ['kepoi_name', 'koi_teq_err1', 'koi_teq_err2']\n", "Число исходных признаков: 45\n", "Числовые признаки: 42\n", "Категориальные признаки: 3\n", "Первые числовые признаки: ['kepid', 'koi_score', 'koi_fpflag_nt', 'koi_fpflag_ss', 'koi_fpflag_co', 'koi_fpflag_ec', 'koi_period', 'koi_period_err1', 'koi_period_err2', 'koi_time0bk', 'koi_time0bk_err1', 'koi_time0bk_err2', 'koi_impact', 'koi_impact_err1', 'koi_impact_err2']\n", "Первые категориальные признаки: ['kepler_name', 'koi_disposition', 'koi_tce_delivname']\n" ] } ], "source": [ "\n", "# Отделяем признаки и убираем очевидные идентификаторы / почти уникальные текстовые поля\n", "X = df.drop(columns=[target_col]).copy()\n", "\n", "def find_drop_columns(dataframe):\n", " to_drop = []\n", " n = len(dataframe)\n", " for col in dataframe.columns:\n", " s = dataframe[col]\n", " nunique = s.nunique(dropna=True)\n", " unique_ratio = nunique / max(n, 1)\n", "\n", " if re.search(r\"(^id$|_id$|^id_|name)\", col) and unique_ratio > 0.3:\n", " to_drop.append(col)\n", " continue\n", "\n", " if s.dtype == \"object\" and unique_ratio > 0.95:\n", " to_drop.append(col)\n", " continue\n", "\n", " if nunique <= 1:\n", " to_drop.append(col)\n", " continue\n", "\n", " return sorted(set(to_drop))\n", "\n", "drop_cols = find_drop_columns(X)\n", "print(\"Удаляемые идентификаторы / несодержательные поля:\", drop_cols)\n", "\n", "X = X.drop(columns=drop_cols, errors=\"ignore\").copy()\n", "\n", "# Попытка перевести object-столбцы в числовой формат, если это возможно\n", "for col in X.columns:\n", " if X[col].dtype == \"object\":\n", " temp = X[col].astype(str).str.strip()\n", " temp = temp.replace({\"\": np.nan, \"nan\": np.nan, \"None\": np.nan, \"none\": np.nan, \"NULL\": np.nan, \"null\": np.nan})\n", " temp_num = pd.to_numeric(temp, errors=\"coerce\")\n", " share_parsed = temp_num.notna().mean()\n", " if share_parsed >= 0.85:\n", " X[col] = temp_num\n", " else:\n", " X[col] = temp\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", "print(\"Число исходных признаков:\", X.shape[1])\n", "print(\"Числовые признаки:\", len(numeric_features))\n", "print(\"Категориальные признаки:\", len(categorical_features))\n", "print(\"Первые числовые признаки:\", numeric_features[:15])\n", "print(\"Первые категориальные признаки:\", categorical_features[:15])\n" ] }, { "cell_type": "code", "execution_count": 7, "id": "9886ad12", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Размер train: (6120, 45)\n", "Размер val: (1531, 45)\n", "Размер test: (1913, 45)\n" ] } ], "source": [ "\n", "# Разбиение данных\n", "X_trainval, X_test, y_trainval, y_test = train_test_split(\n", " X, y,\n", " test_size=0.2,\n", " random_state=RANDOM_STATE,\n", " stratify=y\n", ")\n", "\n", "X_train, X_val, y_train, y_val = train_test_split(\n", " X_trainval, y_trainval,\n", " test_size=0.2,\n", " random_state=RANDOM_STATE,\n", " stratify=y_trainval\n", ")\n", "\n", "print(\"Размер train:\", X_train.shape)\n", "print(\"Размер val:\", X_val.shape)\n", "print(\"Размер test:\", X_test.shape)\n" ] }, { "cell_type": "code", "execution_count": 8, "id": "c7333379", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Число признаков после кодирования: 2240\n", "Первые 25 признаков после кодирования:\n", "['num__kepid' 'num__koi_score' 'num__koi_fpflag_nt' 'num__koi_fpflag_ss'\n", " 'num__koi_fpflag_co' 'num__koi_fpflag_ec' 'num__koi_period'\n", " 'num__koi_period_err1' 'num__koi_period_err2' 'num__koi_time0bk'\n", " 'num__koi_time0bk_err1' 'num__koi_time0bk_err2' 'num__koi_impact'\n", " 'num__koi_impact_err1' 'num__koi_impact_err2' 'num__koi_duration'\n", " 'num__koi_duration_err1' 'num__koi_duration_err2' 'num__koi_depth'\n", " 'num__koi_depth_err1' 'num__koi_depth_err2' 'num__koi_prad'\n", " 'num__koi_prad_err1' 'num__koi_prad_err2' 'num__koi_teq']\n" ] } ], "source": [ "\n", "# 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", "numeric_transformer = Pipeline(steps=[\n", " (\"imputer\", SimpleImputer(strategy=\"median\")),\n", " (\"scaler\", StandardScaler())\n", "])\n", "\n", "categorical_transformer = Pipeline(steps=[\n", " (\"imputer\", SimpleImputer(strategy=\"most_frequent\")),\n", " (\"onehot\", ohe)\n", "])\n", "\n", "preprocessor = ColumnTransformer(\n", " transformers=[\n", " (\"num\", numeric_transformer, numeric_features),\n", " (\"cat\", categorical_transformer, categorical_features)\n", " ],\n", " remainder=\"drop\"\n", ")\n", "\n", "X_train_proc = preprocessor.fit_transform(X_train)\n", "X_val_proc = preprocessor.transform(X_val)\n", "X_test_proc = preprocessor.transform(X_test)\n", "X_trainval_proc = preprocessor.fit_transform(X_trainval)\n", "X_test_proc_final = preprocessor.transform(X_test)\n", "\n", "feature_names = preprocessor.get_feature_names_out()\n", "\n", "print(\"Число признаков после кодирования:\", len(feature_names))\n", "print(\"Первые 25 признаков после кодирования:\")\n", "print(feature_names[:25])\n" ] }, { "cell_type": "code", "execution_count": 9, "id": "fbcd8812", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Размер train после oversampling: (6202, 1795)\n", "Распределение классов после oversampling:\n" ] }, { "data": { "text/plain": [ "0 3101\n", "1 3101\n", "Name: count, dtype: int64" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "def oversample_multiclass(X_array, y_array, random_state=42):\n", " rng = np.random.default_rng(random_state)\n", " classes, counts = np.unique(y_array, return_counts=True)\n", " max_count = counts.max()\n", "\n", " X_parts = []\n", " y_parts = []\n", "\n", " for cls, cnt in zip(classes, counts):\n", " idx = np.where(y_array == cls)[0]\n", " if cnt < max_count:\n", " extra_idx = rng.choice(idx, size=max_count - cnt, replace=True)\n", " idx = np.concatenate([idx, extra_idx])\n", " rng.shuffle(idx)\n", " X_parts.append(X_array[idx])\n", " y_parts.append(y_array[idx])\n", "\n", " X_bal = np.vstack(X_parts)\n", " y_bal = np.concatenate(y_parts)\n", "\n", " perm = rng.permutation(len(y_bal))\n", " return X_bal[perm], y_bal[perm]\n", "\n", "X_train_bal, y_train_bal = oversample_multiclass(np.asarray(X_train_proc, dtype=np.float32), y_train, random_state=RANDOM_STATE)\n", "X_val_np = np.asarray(X_val_proc, dtype=np.float32)\n", "X_test_np = np.asarray(X_test_proc, dtype=np.float32)\n", "X_trainval_np = np.asarray(X_trainval_proc, dtype=np.float32)\n", "X_test_final_np = np.asarray(X_test_proc_final, dtype=np.float32)\n", "\n", "print(\"Размер train после oversampling:\", X_train_bal.shape)\n", "print(\"Распределение классов после oversampling:\")\n", "display(pd.Series(y_train_bal).value_counts().sort_index())\n" ] }, { "cell_type": "markdown", "id": "8d4bf58b", "metadata": {}, "source": [ "## Глубокая нейросеть и PSO\n", "\n", "Используется архитектура с числом скрытых слоёв больше 3: **(128, 64, 32, 16)**. \n", "PSO выполняется на сбалансированной подвыборке train, после чего лучшая найденная конфигурация весов используется для инициализации модели и дальнейшего дообучения.\n" ] }, { "cell_type": "code", "execution_count": 10, "id": "e0b60499", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Архитектура скрытых слоёв: (128, 64, 32, 16)\n", "Полная структура сети: [1795, 128, 64, 32, 16, 1]\n", "Бинарная классификация: True\n" ] } ], "source": [ "\n", "hidden_layers = (128, 64, 32, 16)\n", "input_dim = X_train_bal.shape[1]\n", "classes_unique = np.unique(y)\n", "n_classes = len(classes_unique)\n", "is_binary = (n_classes == 2)\n", "output_dim = 1 if is_binary else n_classes\n", "layer_sizes = [input_dim] + list(hidden_layers) + [output_dim]\n", "\n", "print(\"Архитектура скрытых слоёв:\", hidden_layers)\n", "print(\"Полная структура сети:\", layer_sizes)\n", "print(\"Бинарная классификация:\", is_binary)\n" ] }, { "cell_type": "code", "execution_count": 11, "id": "3f8d880e", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Число параметров сети: 240769\n" ] } ], "source": [ "\n", "# Вспомогательные функции для ручного прямого прохода\n", "def relu(x):\n", " return np.maximum(0.0, x)\n", "\n", "def sigmoid(z):\n", " return 1.0 / (1.0 + np.exp(-np.clip(z, -40, 40)))\n", "\n", "def softmax(z):\n", " z = z - np.max(z, axis=1, keepdims=True)\n", " exp_z = np.exp(z)\n", " return exp_z / np.clip(exp_z.sum(axis=1, keepdims=True), 1e-12, None)\n", "\n", "def build_shapes(layer_sizes):\n", " shapes = []\n", " for in_dim, out_dim in zip(layer_sizes[:-1], layer_sizes[1:]):\n", " shapes.append((in_dim, out_dim))\n", " return shapes\n", "\n", "weight_shapes = build_shapes(layer_sizes)\n", "bias_shapes = [(out_dim,) for out_dim in layer_sizes[1:]]\n", "\n", "def total_params_count(weight_shapes, bias_shapes):\n", " return int(sum(np.prod(s) for s in weight_shapes) + sum(np.prod(s) for s in bias_shapes))\n", "\n", "param_size = total_params_count(weight_shapes, bias_shapes)\n", "print(\"Число параметров сети:\", param_size)\n", "\n", "def unpack_params(vector, weight_shapes, bias_shapes):\n", " pos = 0\n", " weights = []\n", " biases = []\n", "\n", " for shape in weight_shapes:\n", " size = int(np.prod(shape))\n", " w = vector[pos:pos+size].reshape(shape)\n", " weights.append(w)\n", " pos += size\n", "\n", " for shape in bias_shapes:\n", " size = int(np.prod(shape))\n", " b = vector[pos:pos+size].reshape(shape)\n", " biases.append(b)\n", " pos += size\n", "\n", " return weights, biases\n", "\n", "def forward_proba(X_array, vector, weight_shapes, bias_shapes, is_binary=False):\n", " weights, biases = unpack_params(vector, weight_shapes, bias_shapes)\n", " h = X_array\n", " for idx, (w, b) in enumerate(zip(weights, biases)):\n", " h = h @ w + b\n", " if idx < len(weights) - 1:\n", " h = relu(h)\n", " if is_binary:\n", " proba_pos = sigmoid(h).reshape(-1)\n", " return proba_pos\n", " return softmax(h)\n", "\n", "def cross_entropy_loss(y_true, proba, is_binary=False):\n", " eps = 1e-12\n", " if is_binary:\n", " p = np.clip(proba.reshape(-1), eps, 1.0 - eps)\n", " y_true = y_true.astype(np.float64)\n", " return -np.mean(y_true * np.log(p) + (1.0 - y_true) * np.log(1.0 - p))\n", " p = np.clip(proba[np.arange(len(y_true)), y_true], eps, 1.0)\n", " return -np.mean(np.log(p))\n", "\n", "def metrics_from_proba(y_true, proba, is_binary=False, threshold=0.5):\n", " if is_binary:\n", " proba_pos = np.asarray(proba).reshape(-1)\n", " y_pred = (proba_pos >= threshold).astype(int)\n", " res = {\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", " res[\"roc_auc\"] = roc_auc_score(y_true, proba_pos)\n", " except Exception:\n", " res[\"roc_auc\"] = np.nan\n", " return res, y_pred\n", " y_pred = np.argmax(proba, axis=1)\n", " res = {\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", " res[\"roc_auc\"] = roc_auc_score(y_true, proba, multi_class=\"ovr\", average=\"macro\")\n", " except Exception:\n", " res[\"roc_auc\"] = np.nan\n", " return res, y_pred\n" ] }, { "cell_type": "code", "execution_count": 12, "id": "7eedf9b2", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Размер подвыборки для PSO: (4000, 1795)\n" ] } ], "source": [ "\n", "# Подвыборка для PSO, чтобы ограничить время работы на больших данных\n", "rng = np.random.default_rng(RANDOM_STATE)\n", "\n", "max_pso_samples = min(4000, len(X_train_bal))\n", "if len(X_train_bal) > max_pso_samples:\n", " pso_idx = rng.choice(len(X_train_bal), size=max_pso_samples, replace=False)\n", " X_pso = X_train_bal[pso_idx]\n", " y_pso = y_train_bal[pso_idx]\n", "else:\n", " X_pso = X_train_bal.copy()\n", " y_pso = y_train_bal.copy()\n", "\n", "print(\"Размер подвыборки для PSO:\", X_pso.shape)\n" ] }, { "cell_type": "code", "execution_count": 13, "id": "3985e855", "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", "
iterationbest_scorebest_train_lossbest_val_lossbest_val_f1_macromean_particle_score
560.7357610.6250300.6440120.7378600.836567
670.7357610.6250300.6440120.737860NaN
780.6936880.5985010.6054140.7477880.834892
890.682013NaN0.5884100.732563NaN
9100.6743450.5738700.5821760.736661NaN
10110.6743450.5738700.5821760.736661NaN
11120.6604000.5625790.5943420.811262NaN
12130.6604000.5625790.5943420.811262NaN
13140.6408240.5712020.5676750.7910040.763580
14150.6113860.5513050.5440890.807722NaN
\n", "
" ], "text/plain": [ " iteration best_score best_train_loss best_val_loss best_val_f1_macro \\\n", "5 6 0.735761 0.625030 0.644012 0.737860 \n", "6 7 0.735761 0.625030 0.644012 0.737860 \n", "7 8 0.693688 0.598501 0.605414 0.747788 \n", "8 9 0.682013 NaN 0.588410 0.732563 \n", "9 10 0.674345 0.573870 0.582176 0.736661 \n", "10 11 0.674345 0.573870 0.582176 0.736661 \n", "11 12 0.660400 0.562579 0.594342 0.811262 \n", "12 13 0.660400 0.562579 0.594342 0.811262 \n", "13 14 0.640824 0.571202 0.567675 0.791004 \n", "14 15 0.611386 0.551305 0.544089 0.807722 \n", "\n", " mean_particle_score \n", "5 0.836567 \n", "6 NaN \n", "7 0.834892 \n", "8 NaN \n", "9 NaN \n", "10 NaN \n", "11 NaN \n", "12 NaN \n", "13 0.763580 \n", "14 NaN " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "# PSO\n", "n_particles = 8\n", "n_iterations = 15\n", "w_inertia = 0.72\n", "c1 = 1.45\n", "c2 = 1.45\n", "init_scale = 0.08\n", "velocity_scale = 0.02\n", "\n", "particles = rng.normal(0.0, init_scale, size=(n_particles, param_size)).astype(np.float32)\n", "velocities = rng.normal(0.0, velocity_scale, size=(n_particles, param_size)).astype(np.float32)\n", "\n", "personal_best = particles.copy()\n", "personal_best_scores = np.full(n_particles, np.inf, dtype=np.float64)\n", "\n", "global_best = None\n", "global_best_score = np.inf\n", "global_best_info = {}\n", "\n", "pso_history = []\n", "\n", "for iteration in range(n_iterations):\n", " particle_scores = []\n", "\n", " for p in range(n_particles):\n", " vec = particles[p]\n", "\n", " train_proba = forward_proba(X_pso, vec, weight_shapes, bias_shapes, is_binary=is_binary)\n", " val_proba = forward_proba(X_val_np, vec, weight_shapes, bias_shapes, is_binary=is_binary)\n", "\n", " train_loss = cross_entropy_loss(y_pso, train_proba, is_binary=is_binary)\n", " val_loss = cross_entropy_loss(y_val, val_proba, is_binary=is_binary)\n", "\n", " val_metrics, _ = metrics_from_proba(y_val, val_proba, is_binary=is_binary)\n", " val_f1 = val_metrics[\"f1_macro\"]\n", "\n", " # минимизируем score: меньше loss и выше macro-F1\n", " score = val_loss + 0.35 * (1.0 - val_f1)\n", "\n", " particle_scores.append(score)\n", "\n", " if score < personal_best_scores[p]:\n", " personal_best_scores[p] = score\n", " personal_best[p] = vec.copy()\n", "\n", " if score < global_best_score:\n", " global_best_score = score\n", " global_best = vec.copy()\n", " global_best_info = {\n", " \"best_train_loss\": float(train_loss),\n", " \"best_val_loss\": float(val_loss),\n", " \"best_val_f1_macro\": float(val_f1),\n", " }\n", "\n", " # обновление скоростей и частиц\n", " for p in range(n_particles):\n", " r1 = rng.random(param_size, dtype=np.float32)\n", " r2 = rng.random(param_size, dtype=np.float32)\n", "\n", " velocities[p] = (\n", " w_inertia * velocities[p]\n", " + c1 * r1 * (personal_best[p] - particles[p])\n", " + c2 * r2 * (global_best - particles[p])\n", " )\n", "\n", " particles[p] = particles[p] + velocities[p]\n", " particles[p] = np.clip(particles[p], -2.5, 2.5)\n", "\n", " pso_history.append({\n", " \"iteration\": iteration + 1,\n", " \"best_score\": float(global_best_score),\n", " \"best_train_loss\": global_best_info.get(\"best_train_loss\", np.nan),\n", " \"best_val_loss\": global_best_info.get(\"best_val_loss\", np.nan),\n", " \"best_val_f1_macro\": global_best_info.get(\"best_val_f1_macro\", np.nan),\n", " \"mean_particle_score\": float(np.mean(particle_scores)),\n", " })\n", "\n", "pso_history_df = pd.DataFrame(pso_history)\n", "display(pso_history_df.tail(10))\n" ] }, { "cell_type": "code", "execution_count": 14, "id": "fca1e0f0", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Метрики модели после PSO на 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", "
accuracybalanced_accuracyprecision_macrorecall_macrof1_macroroc_auc
00.8099280.8086110.8214790.8086110.8077220.887868
\n", "
" ], "text/plain": [ " accuracy balanced_accuracy precision_macro recall_macro f1_macro \\\n", "0 0.809928 0.808611 0.821479 0.808611 0.807722 \n", "\n", " roc_auc \n", "0 0.887868 " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Метрики модели после PSO 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", "
accuracybalanced_accuracyprecision_macrorecall_macrof1_macroroc_auc
00.7961320.7946740.8088730.7946740.7934240.882042
\n", "
" ], "text/plain": [ " accuracy balanced_accuracy precision_macro recall_macro f1_macro \\\n", "0 0.796132 0.794674 0.808873 0.794674 0.793424 \n", "\n", " roc_auc \n", "0 0.882042 " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "# Качество сети сразу после PSO\n", "pso_val_proba = forward_proba(X_val_np, global_best, weight_shapes, bias_shapes, is_binary=is_binary)\n", "pso_test_proba = forward_proba(X_test_np, global_best, weight_shapes, bias_shapes, is_binary=is_binary)\n", "\n", "pso_val_metrics, pso_val_pred = metrics_from_proba(y_val, pso_val_proba, is_binary=is_binary)\n", "pso_test_metrics, pso_test_pred = metrics_from_proba(y_test, pso_test_proba, is_binary=is_binary)\n", "\n", "print(\"Метрики модели после PSO на validation:\")\n", "display(pd.DataFrame([pso_val_metrics]))\n", "print(\"Метрики модели после PSO на test:\")\n", "display(pd.DataFrame([pso_test_metrics]))\n" ] }, { "cell_type": "code", "execution_count": 15, "id": "6544f719", "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", "
epochtrain_lossval_accuracyval_balanced_accuracyval_f1_macroval_roc_auc
10110.0023940.9967340.9967430.9967340.997418
11120.0022040.9973870.9974050.9973870.997418
12130.0020660.9973870.9974050.9973870.997418
13140.0019550.9973870.9974050.9973870.997418
14150.0018640.9980400.9980670.9980400.997418
15160.0017870.9980400.9980670.9980400.997419
16170.0017220.9980400.9980670.9980400.997419
17180.0016650.9980400.9980670.9980400.997419
18190.0016150.9980400.9980670.9980400.997419
19200.0015700.9980400.9980670.9980400.997419
\n", "
" ], "text/plain": [ " epoch train_loss val_accuracy val_balanced_accuracy val_f1_macro \\\n", "10 11 0.002394 0.996734 0.996743 0.996734 \n", "11 12 0.002204 0.997387 0.997405 0.997387 \n", "12 13 0.002066 0.997387 0.997405 0.997387 \n", "13 14 0.001955 0.997387 0.997405 0.997387 \n", "14 15 0.001864 0.998040 0.998067 0.998040 \n", "15 16 0.001787 0.998040 0.998067 0.998040 \n", "16 17 0.001722 0.998040 0.998067 0.998040 \n", "17 18 0.001665 0.998040 0.998067 0.998040 \n", "18 19 0.001615 0.998040 0.998067 0.998040 \n", "19 20 0.001570 0.998040 0.998067 0.998040 \n", "\n", " val_roc_auc \n", "10 0.997418 \n", "11 0.997418 \n", "12 0.997418 \n", "13 0.997418 \n", "14 0.997418 \n", "15 0.997419 \n", "16 0.997419 \n", "17 0.997419 \n", "18 0.997419 \n", "19 0.997419 " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "# Инициализация и дообучение MLPClassifier весами из PSO\n", "mlp = MLPClassifier(\n", " hidden_layer_sizes=hidden_layers,\n", " activation=\"relu\",\n", " solver=\"adam\",\n", " alpha=1e-4,\n", " batch_size=256,\n", " learning_rate_init=1e-3,\n", " max_iter=1,\n", " warm_start=True,\n", " shuffle=True,\n", " random_state=RANDOM_STATE\n", ")\n", "\n", "# первый вызов partial_fit создаёт внутренние структуры модели\n", "mlp.partial_fit(X_train_bal, y_train_bal, classes=np.arange(n_classes))\n", "\n", "# перенос весов из PSO\n", "weights_best, biases_best = unpack_params(global_best, weight_shapes, bias_shapes)\n", "mlp.coefs_ = [w.astype(np.float64) for w in weights_best]\n", "mlp.intercepts_ = [b.astype(np.float64) for b in biases_best]\n", "\n", "fine_tune_history = []\n", "best_state = {\n", " \"coefs_\": [w.copy() for w in mlp.coefs_],\n", " \"intercepts_\": [b.copy() for b in mlp.intercepts_],\n", " \"best_val_f1_macro\": -np.inf\n", "}\n", "\n", "n_epochs = 20\n", "for epoch in range(n_epochs):\n", " mlp.partial_fit(X_train_bal, y_train_bal)\n", "\n", " val_pred = np.asarray(mlp.predict(X_val_np)).reshape(-1)\n", " val_proba_raw = mlp.predict_proba(X_val_np)\n", "\n", " if is_binary:\n", " if val_proba_raw.ndim == 2:\n", " val_proba_for_metric = val_proba_raw[:, 1]\n", " else:\n", " val_proba_for_metric = np.asarray(val_proba_raw).reshape(-1)\n", " else:\n", " val_proba_for_metric = val_proba_raw\n", "\n", " row = {\n", " \"epoch\": epoch + 1,\n", " \"train_loss\": float(mlp.loss_),\n", " \"val_accuracy\": accuracy_score(y_val, val_pred),\n", " \"val_balanced_accuracy\": balanced_accuracy_score(y_val, val_pred),\n", " \"val_f1_macro\": f1_score(y_val, val_pred, average=\"macro\", zero_division=0),\n", " }\n", " try:\n", " if is_binary:\n", " row[\"val_roc_auc\"] = roc_auc_score(y_val, val_proba_for_metric)\n", " else:\n", " row[\"val_roc_auc\"] = roc_auc_score(y_val, val_proba_for_metric, multi_class=\"ovr\", average=\"macro\")\n", " except Exception:\n", " row[\"val_roc_auc\"] = np.nan\n", "\n", " fine_tune_history.append(row)\n", "\n", " if row[\"val_f1_macro\"] > best_state[\"best_val_f1_macro\"]:\n", " best_state = {\n", " \"coefs_\": [w.copy() for w in mlp.coefs_],\n", " \"intercepts_\": [b.copy() for b in mlp.intercepts_],\n", " \"best_val_f1_macro\": row[\"val_f1_macro\"]\n", " }\n", "\n", "fine_tune_history_df = pd.DataFrame(fine_tune_history)\n", "display(fine_tune_history_df.tail(10))\n", "\n", "# восстановление лучшей версии модели по macro-F1 на validation\n", "mlp.coefs_ = [w.copy() for w in best_state[\"coefs_\"]]\n", "mlp.intercepts_ = [b.copy() for b in best_state[\"intercepts_\"]]\n" ] }, { "cell_type": "code", "execution_count": 16, "id": "41785391", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Метрики финальной модели на 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", "
accuracybalanced_accuracyprecision_macrorecall_macrof1_macroroc_auc
00.998040.9980670.9980210.9980670.998040.997418
\n", "
" ], "text/plain": [ " accuracy balanced_accuracy precision_macro recall_macro f1_macro \\\n", "0 0.99804 0.998067 0.998021 0.998067 0.99804 \n", "\n", " roc_auc \n", "0 0.997418 " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Метрики финальной модели 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", "
accuracybalanced_accuracyprecision_macrorecall_macrof1_macroroc_auc
00.9968640.9968630.9968630.9968630.9968630.998922
\n", "
" ], "text/plain": [ " accuracy balanced_accuracy precision_macro recall_macro f1_macro \\\n", "0 0.996864 0.996863 0.996863 0.996863 0.996863 \n", "\n", " roc_auc \n", "0 0.998922 " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "# Финальная оценка\n", "ft_val_pred = np.asarray(mlp.predict(X_val_np)).reshape(-1)\n", "ft_val_proba_raw = mlp.predict_proba(X_val_np)\n", "\n", "ft_test_pred = np.asarray(mlp.predict(X_test_np)).reshape(-1)\n", "ft_test_proba_raw = mlp.predict_proba(X_test_np)\n", "\n", "if is_binary:\n", " ft_val_proba = ft_val_proba_raw[:, 1] if np.ndim(ft_val_proba_raw) == 2 else np.asarray(ft_val_proba_raw).reshape(-1)\n", " ft_test_proba = ft_test_proba_raw[:, 1] if np.ndim(ft_test_proba_raw) == 2 else np.asarray(ft_test_proba_raw).reshape(-1)\n", "else:\n", " ft_val_proba = ft_val_proba_raw\n", " ft_test_proba = ft_test_proba_raw\n", "\n", "def collect_metrics(y_true, y_pred, y_proba, is_binary=False):\n", " res = {\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", " if is_binary:\n", " res[\"roc_auc\"] = roc_auc_score(y_true, y_proba)\n", " else:\n", " res[\"roc_auc\"] = roc_auc_score(y_true, y_proba, multi_class=\"ovr\", average=\"macro\")\n", " except Exception:\n", " res[\"roc_auc\"] = np.nan\n", " return res\n", "\n", "ft_val_metrics = collect_metrics(y_val, ft_val_pred, ft_val_proba, is_binary=is_binary)\n", "ft_test_metrics = collect_metrics(y_test, ft_test_pred, ft_test_proba, is_binary=is_binary)\n", "\n", "print(\"Метрики финальной модели на validation:\")\n", "display(pd.DataFrame([ft_val_metrics]))\n", "print(\"Метрики финальной модели на test:\")\n", "display(pd.DataFrame([ft_test_metrics]))\n" ] }, { "cell_type": "code", "execution_count": 17, "id": "54770815", "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", "
modelaccuracybalanced_accuracyprecision_macrorecall_macrof1_macroroc_aucn_original_featuresn_processed_featureshidden_layers
0Глубокая нейросеть: PSO + дообучение0.9968640.9968630.9968630.9968630.9968630.998922452240(128, 64, 32, 16)
1Глубокая нейросеть после PSO0.7961320.7946740.8088730.7946740.7934240.882042452240(128, 64, 32, 16)
\n", "
" ], "text/plain": [ " model accuracy balanced_accuracy \\\n", "0 Глубокая нейросеть: PSO + дообучение 0.996864 0.996863 \n", "1 Глубокая нейросеть после PSO 0.796132 0.794674 \n", "\n", " precision_macro recall_macro f1_macro roc_auc n_original_features \\\n", "0 0.996863 0.996863 0.996863 0.998922 45 \n", "1 0.808873 0.794674 0.793424 0.882042 45 \n", "\n", " n_processed_features hidden_layers \n", "0 2240 (128, 64, 32, 16) \n", "1 2240 (128, 64, 32, 16) " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Лучшая модель:\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", "
modelaccuracybalanced_accuracyprecision_macrorecall_macrof1_macroroc_aucn_original_featuresn_processed_featureshidden_layers
0Глубокая нейросеть: PSO + дообучение0.9968640.9968630.9968630.9968630.9968630.998922452240(128, 64, 32, 16)
\n", "
" ], "text/plain": [ " model accuracy balanced_accuracy \\\n", "0 Глубокая нейросеть: PSO + дообучение 0.996864 0.996863 \n", "\n", " precision_macro recall_macro f1_macro roc_auc n_original_features \\\n", "0 0.996863 0.996863 0.996863 0.998922 45 \n", "\n", " n_processed_features hidden_layers \n", "0 2240 (128, 64, 32, 16) " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "# Сравнение моделей\n", "results_df = pd.DataFrame([\n", " {\n", " \"model\": \"Глубокая нейросеть после PSO\",\n", " \"accuracy\": pso_test_metrics[\"accuracy\"],\n", " \"balanced_accuracy\": pso_test_metrics[\"balanced_accuracy\"],\n", " \"precision_macro\": pso_test_metrics[\"precision_macro\"],\n", " \"recall_macro\": pso_test_metrics[\"recall_macro\"],\n", " \"f1_macro\": pso_test_metrics[\"f1_macro\"],\n", " \"roc_auc\": pso_test_metrics[\"roc_auc\"],\n", " \"n_original_features\": X.shape[1],\n", " \"n_processed_features\": len(feature_names),\n", " \"hidden_layers\": str(hidden_layers),\n", " },\n", " {\n", " \"model\": \"Глубокая нейросеть: PSO + дообучение\",\n", " \"accuracy\": ft_test_metrics[\"accuracy\"],\n", " \"balanced_accuracy\": ft_test_metrics[\"balanced_accuracy\"],\n", " \"precision_macro\": ft_test_metrics[\"precision_macro\"],\n", " \"recall_macro\": ft_test_metrics[\"recall_macro\"],\n", " \"f1_macro\": ft_test_metrics[\"f1_macro\"],\n", " \"roc_auc\": ft_test_metrics[\"roc_auc\"],\n", " \"n_original_features\": X.shape[1],\n", " \"n_processed_features\": len(feature_names),\n", " \"hidden_layers\": str(hidden_layers),\n", " }\n", "]).sort_values([\"f1_macro\", \"balanced_accuracy\", \"accuracy\"], ascending=False).reset_index(drop=True)\n", "\n", "display(results_df)\n", "best_model_df = results_df.head(1)\n", "print(\"Лучшая модель:\")\n", "display(best_model_df)\n" ] }, { "cell_type": "code", "execution_count": 18, "id": "434faa41", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Матрица ошибок для модели после PSO:\n", "[[652 291]\n", " [ 99 871]]\n", "\n", "Матрица ошибок для финальной модели:\n", "[[940 3]\n", " [ 3 967]]\n", "\n", "Классы:\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", "
class_indexclass_label
00CANDIDATE
11FALSE POSITIVE
\n", "
" ], "text/plain": [ " class_index class_label\n", "0 0 CANDIDATE\n", "1 1 FALSE POSITIVE" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "\n", "Классификационный отчёт для финальной модели:\n", " precision recall f1-score support\n", "\n", " CANDIDATE 1.00 1.00 1.00 943\n", "FALSE POSITIVE 1.00 1.00 1.00 970\n", "\n", " accuracy 1.00 1913\n", " macro avg 1.00 1.00 1.00 1913\n", " weighted avg 1.00 1.00 1.00 1913\n", "\n" ] } ], "source": [ "\n", "# Матрицы ошибок и короткие отчёты\n", "cm_pso = confusion_matrix(y_test, pso_test_pred)\n", "cm_ft = confusion_matrix(y_test, ft_test_pred)\n", "\n", "print(\"Матрица ошибок для модели после PSO:\")\n", "print(cm_pso)\n", "\n", "print(\"\\nМатрица ошибок для финальной модели:\")\n", "print(cm_ft)\n", "\n", "print(\"\\nКлассы:\")\n", "display(class_mapping)\n", "\n", "print(\"\\nКлассификационный отчёт для финальной модели:\")\n", "print(classification_report(y_test, ft_test_pred, target_names=label_encoder.classes_, zero_division=0))\n" ] }, { "cell_type": "code", "execution_count": 19, "id": "9e10c041", "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", "
y_truey_pred_psoy_pred_final
0CANDIDATECANDIDATECANDIDATE
1FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
2FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
3FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
4FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
5CANDIDATECANDIDATECANDIDATE
6FALSE POSITIVECANDIDATEFALSE POSITIVE
7CANDIDATECANDIDATECANDIDATE
8FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
9FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
\n", "
" ], "text/plain": [ " y_true y_pred_pso y_pred_final\n", "0 CANDIDATE CANDIDATE CANDIDATE\n", "1 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE\n", "2 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE\n", "3 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE\n", "4 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE\n", "5 CANDIDATE CANDIDATE CANDIDATE\n", "6 FALSE POSITIVE CANDIDATE FALSE POSITIVE\n", "7 CANDIDATE CANDIDATE CANDIDATE\n", "8 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE\n", "9 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "# Первые 10 фактических и предсказанных значений\n", "pred_preview = pd.DataFrame({\n", " \"y_true\": label_encoder.inverse_transform(y_test[:10]),\n", " \"y_pred_pso\": label_encoder.inverse_transform(pso_test_pred[:10]),\n", " \"y_pred_final\": label_encoder.inverse_transform(ft_test_pred[:10]),\n", "})\n", "display(pred_preview)\n" ] }, { "cell_type": "code", "execution_count": 20, "id": "43e463ea", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "=== КЛЮЧЕВАЯ ИНФОРМАЦИЯ ДЛЯ ВЫВОДА ===\n", "Зависимая переменная: koi_pdisposition\n", "\n", "Число исходных признаков: 45\n", "Число признаков после кодирования: 2240\n", "Архитектура скрытых слоёв: (128, 64, 32, 16)\n", "\n", "Распределение классов:\n" ] }, { "data": { "text/plain": [ "koi_pdisposition\n", "CANDIDATE 0.493204\n", "FALSE POSITIVE 0.506796\n", "Name: share, dtype: float64" ] }, "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", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \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
0kepid
1kepler_name
2koi_disposition
3koi_score
4koi_fpflag_nt
5koi_fpflag_ss
6koi_fpflag_co
7koi_fpflag_ec
8koi_period
9koi_period_err1
10koi_period_err2
11koi_time0bk
12koi_time0bk_err1
13koi_time0bk_err2
14koi_impact
15koi_impact_err1
16koi_impact_err2
17koi_duration
18koi_duration_err1
19koi_duration_err2
20koi_depth
21koi_depth_err1
22koi_depth_err2
23koi_prad
24koi_prad_err1
25koi_prad_err2
26koi_teq
27koi_insol
28koi_insol_err1
29koi_insol_err2
30koi_model_snr
31koi_tce_plnt_num
32koi_tce_delivname
33koi_steff
34koi_steff_err1
35koi_steff_err2
36koi_slogg
37koi_slogg_err1
38koi_slogg_err2
39koi_srad
40koi_srad_err1
41koi_srad_err2
42ra
43dec
44koi_kepmag
\n", "
" ], "text/plain": [ " feature\n", "0 kepid\n", "1 kepler_name\n", "2 koi_disposition\n", "3 koi_score\n", "4 koi_fpflag_nt\n", "5 koi_fpflag_ss\n", "6 koi_fpflag_co\n", "7 koi_fpflag_ec\n", "8 koi_period\n", "9 koi_period_err1\n", "10 koi_period_err2\n", "11 koi_time0bk\n", "12 koi_time0bk_err1\n", "13 koi_time0bk_err2\n", "14 koi_impact\n", "15 koi_impact_err1\n", "16 koi_impact_err2\n", "17 koi_duration\n", "18 koi_duration_err1\n", "19 koi_duration_err2\n", "20 koi_depth\n", "21 koi_depth_err1\n", "22 koi_depth_err2\n", "23 koi_prad\n", "24 koi_prad_err1\n", "25 koi_prad_err2\n", "26 koi_teq\n", "27 koi_insol\n", "28 koi_insol_err1\n", "29 koi_insol_err2\n", "30 koi_model_snr\n", "31 koi_tce_plnt_num\n", "32 koi_tce_delivname\n", "33 koi_steff\n", "34 koi_steff_err1\n", "35 koi_steff_err2\n", "36 koi_slogg\n", "37 koi_slogg_err1\n", "38 koi_slogg_err2\n", "39 koi_srad\n", "40 koi_srad_err1\n", "41 koi_srad_err2\n", "42 ra\n", "43 dec\n", "44 koi_kepmag" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Метрики моделей на тестовой выборке:\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", "
modelaccuracybalanced_accuracyprecision_macrorecall_macrof1_macroroc_aucn_original_featuresn_processed_featureshidden_layers
0Глубокая нейросеть: PSO + дообучение0.9968640.9968630.9968630.9968630.9968630.998922452240(128, 64, 32, 16)
1Глубокая нейросеть после PSO0.7961320.7946740.8088730.7946740.7934240.882042452240(128, 64, 32, 16)
\n", "
" ], "text/plain": [ " model accuracy balanced_accuracy \\\n", "0 Глубокая нейросеть: PSO + дообучение 0.996864 0.996863 \n", "1 Глубокая нейросеть после PSO 0.796132 0.794674 \n", "\n", " precision_macro recall_macro f1_macro roc_auc n_original_features \\\n", "0 0.996863 0.996863 0.996863 0.998922 45 \n", "1 0.808873 0.794674 0.793424 0.882042 45 \n", "\n", " n_processed_features hidden_layers \n", "0 2240 (128, 64, 32, 16) \n", "1 2240 (128, 64, 32, 16) " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Лучшая модель:\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", "
modelaccuracybalanced_accuracyprecision_macrorecall_macrof1_macroroc_aucn_original_featuresn_processed_featureshidden_layers
0Глубокая нейросеть: PSO + дообучение0.9968640.9968630.9968630.9968630.9968630.998922452240(128, 64, 32, 16)
\n", "
" ], "text/plain": [ " model accuracy balanced_accuracy \\\n", "0 Глубокая нейросеть: PSO + дообучение 0.996864 0.996863 \n", "\n", " precision_macro recall_macro f1_macro roc_auc n_original_features \\\n", "0 0.996863 0.996863 0.996863 0.998922 45 \n", "\n", " n_processed_features hidden_layers \n", "0 2240 (128, 64, 32, 16) " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Последние итерации PSO:\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", "
iterationbest_scorebest_train_lossbest_val_lossbest_val_f1_macromean_particle_score
560.7357610.6250300.6440120.7378600.836567
670.7357610.6250300.6440120.737860NaN
780.6936880.5985010.6054140.7477880.834892
890.682013NaN0.5884100.732563NaN
9100.6743450.5738700.5821760.736661NaN
10110.6743450.5738700.5821760.736661NaN
11120.6604000.5625790.5943420.811262NaN
12130.6604000.5625790.5943420.811262NaN
13140.6408240.5712020.5676750.7910040.763580
14150.6113860.5513050.5440890.807722NaN
\n", "
" ], "text/plain": [ " iteration best_score best_train_loss best_val_loss best_val_f1_macro \\\n", "5 6 0.735761 0.625030 0.644012 0.737860 \n", "6 7 0.735761 0.625030 0.644012 0.737860 \n", "7 8 0.693688 0.598501 0.605414 0.747788 \n", "8 9 0.682013 NaN 0.588410 0.732563 \n", "9 10 0.674345 0.573870 0.582176 0.736661 \n", "10 11 0.674345 0.573870 0.582176 0.736661 \n", "11 12 0.660400 0.562579 0.594342 0.811262 \n", "12 13 0.660400 0.562579 0.594342 0.811262 \n", "13 14 0.640824 0.571202 0.567675 0.791004 \n", "14 15 0.611386 0.551305 0.544089 0.807722 \n", "\n", " mean_particle_score \n", "5 0.836567 \n", "6 NaN \n", "7 0.834892 \n", "8 NaN \n", "9 NaN \n", "10 NaN \n", "11 NaN \n", "12 NaN \n", "13 0.763580 \n", "14 NaN " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Последние эпохи дообучения:\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", "
epochtrain_lossval_accuracyval_balanced_accuracyval_f1_macroval_roc_auc
10110.0023940.9967340.9967430.9967340.997418
11120.0022040.9973870.9974050.9973870.997418
12130.0020660.9973870.9974050.9973870.997418
13140.0019550.9973870.9974050.9973870.997418
14150.0018640.9980400.9980670.9980400.997418
15160.0017870.9980400.9980670.9980400.997419
16170.0017220.9980400.9980670.9980400.997419
17180.0016650.9980400.9980670.9980400.997419
18190.0016150.9980400.9980670.9980400.997419
19200.0015700.9980400.9980670.9980400.997419
\n", "
" ], "text/plain": [ " epoch train_loss val_accuracy val_balanced_accuracy val_f1_macro \\\n", "10 11 0.002394 0.996734 0.996743 0.996734 \n", "11 12 0.002204 0.997387 0.997405 0.997387 \n", "12 13 0.002066 0.997387 0.997405 0.997387 \n", "13 14 0.001955 0.997387 0.997405 0.997387 \n", "14 15 0.001864 0.998040 0.998067 0.998040 \n", "15 16 0.001787 0.998040 0.998067 0.998040 \n", "16 17 0.001722 0.998040 0.998067 0.998040 \n", "17 18 0.001665 0.998040 0.998067 0.998040 \n", "18 19 0.001615 0.998040 0.998067 0.998040 \n", "19 20 0.001570 0.998040 0.998067 0.998040 \n", "\n", " val_roc_auc \n", "10 0.997418 \n", "11 0.997418 \n", "12 0.997418 \n", "13 0.997418 \n", "14 0.997418 \n", "15 0.997419 \n", "16 0.997419 \n", "17 0.997419 \n", "18 0.997419 \n", "19 0.997419 " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Первые 10 фактических и предсказанных значений на тесте для финальной 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", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
y_truey_pred_psoy_pred_final
0CANDIDATECANDIDATECANDIDATE
1FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
2FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
3FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
4FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
5CANDIDATECANDIDATECANDIDATE
6FALSE POSITIVECANDIDATEFALSE POSITIVE
7CANDIDATECANDIDATECANDIDATE
8FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
9FALSE POSITIVEFALSE POSITIVEFALSE POSITIVE
\n", "
" ], "text/plain": [ " y_true y_pred_pso y_pred_final\n", "0 CANDIDATE CANDIDATE CANDIDATE\n", "1 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE\n", "2 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE\n", "3 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE\n", "4 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE\n", "5 CANDIDATE CANDIDATE CANDIDATE\n", "6 FALSE POSITIVE CANDIDATE FALSE POSITIVE\n", "7 CANDIDATE CANDIDATE CANDIDATE\n", "8 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE\n", "9 FALSE POSITIVE FALSE POSITIVE FALSE POSITIVE" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Матрица ошибок для модели после PSO:\n", "[[652 291]\n", " [ 99 871]]\n", "Матрица ошибок для финальной модели:\n", "[[940 3]\n", " [ 3 967]]\n" ] } ], "source": [ "\n", "print(\"=== КЛЮЧЕВАЯ ИНФОРМАЦИЯ ДЛЯ ВЫВОДА ===\")\n", "print(\"Зависимая переменная:\", target_col)\n", "print()\n", "print(\"Число исходных признаков:\", X.shape[1])\n", "print(\"Число признаков после кодирования:\", len(feature_names))\n", "print(\"Архитектура скрытых слоёв:\", hidden_layers)\n", "print()\n", "print(\"Распределение классов:\")\n", "display(class_share)\n", "print(\"Использованные исходные признаки:\")\n", "display(pd.DataFrame({\"feature\": X.columns}))\n", "print(\"Метрики моделей на тестовой выборке:\")\n", "display(results_df)\n", "print(\"Лучшая модель:\")\n", "display(best_model_df)\n", "print(\"Последние итерации PSO:\")\n", "display(pso_history_df.tail(10))\n", "print(\"Последние эпохи дообучения:\")\n", "display(fine_tune_history_df.tail(10))\n", "print(\"Первые 10 фактических и предсказанных значений на тесте для финальной модели:\")\n", "display(pred_preview)\n", "print(\"Матрица ошибок для модели после PSO:\")\n", "print(cm_pso)\n", "print(\"Матрица ошибок для финальной модели:\")\n", "print(cm_ft)\n" ] }, { "cell_type": "markdown", "id": "3f6e90a3", "metadata": {}, "source": [ "Итог\n", "\n", "Зависимой переменной выбрана koi_pdisposition. Построена глубокая нейросеть с архитектурой (128, 64, 32, 16), сначала обученная методом PSO, затем дообученная градиентным методом.\n", "\n", "Лучшая модель — PSO + дообучение: на тестовой выборке получено accuracy = 0.9969, balanced accuracy = 0.9969, F1-macro = 0.9969, ROC-AUC = 0.9989. По сравнению с моделью только после PSO качество резко улучшилось (accuracy: 0.796 → 0.997).\n", "\n", "Следовательно, дообучение после весовой эволюции существенно повышает точность классификации, и лучшей является финальная модель PSO + fine-tuning." ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "name": "python" } }, "nbformat": 4, "nbformat_minor": 5 }