{ "cells": [ { "cell_type": "markdown", "id": "ea8251c3", "metadata": {}, "source": [ "\n", "# Регрессия нейросетями и эволюционный поиск архитектуры\n", "\n", "В работе автоматически выбирается ключевая зависимая переменная, строится базовая нейросеть по всем признакам, затем определяется важность признаков и формируется сокращённый набор переменных. После этого по выбранным признакам обучается вторая нейросеть, архитектура которой подбирается методом эволюционного поиска. В конце выводятся все метрики и таблицы, необходимые для краткого итогового вывода.\n" ] }, { "cell_type": "code", "execution_count": 1, "id": "4a61905f", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Используемый файл: Mental_Health_Lifestyle_Dataset(1).csv\n", "Форма датасета: (3000, 12)\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", "
CountryAgeGenderExercise LevelDiet TypeSleep HoursStress LevelMental Health ConditionWork Hours per WeekScreen Time per Day (Hours)Social Interaction ScoreHappiness Score
0Brazil48MaleLowVegetarian6.3LowNaN214.07.86.5
1Australia31MaleModerateVegan4.9LowPTSD485.28.26.8
2Japan37FemaleLowVegetarian7.2HighNaN434.79.69.7
3Brazil35MaleLowVegan7.2LowDepression432.28.26.6
4Germany46MaleLowBalanced7.3LowAnxiety353.64.74.4
\n", "
" ], "text/plain": [ " Country Age Gender Exercise Level Diet Type Sleep Hours \\\n", "0 Brazil 48 Male Low Vegetarian 6.3 \n", "1 Australia 31 Male Moderate Vegan 4.9 \n", "2 Japan 37 Female Low Vegetarian 7.2 \n", "3 Brazil 35 Male Low Vegan 7.2 \n", "4 Germany 46 Male Low Balanced 7.3 \n", "\n", " Stress Level Mental Health Condition Work Hours per Week \\\n", "0 Low NaN 21 \n", "1 Low PTSD 48 \n", "2 High NaN 43 \n", "3 Low Depression 43 \n", "4 Low Anxiety 35 \n", "\n", " Screen Time per Day (Hours) Social Interaction Score Happiness Score \n", "0 4.0 7.8 6.5 \n", "1 5.2 8.2 6.8 \n", "2 4.7 9.6 9.7 \n", "3 2.2 8.2 6.6 \n", "4 3.6 4.7 4.4 " ] }, "metadata": {}, "output_type": "display_data" }, { "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", "
original_namenormalized_name
0Countrycountry
1Ageage
2Gendergender
3Exercise Levelexercise_level
4Diet Typediet_type
5Sleep Hourssleep_hours
6Stress Levelstress_level
7Mental Health Conditionmental_health_condition
8Work Hours per Weekwork_hours_per_week
9Screen Time per Day (Hours)screen_time_per_day_hours
10Social Interaction Scoresocial_interaction_score
11Happiness Scorehappiness_score
\n", "
" ], "text/plain": [ " original_name normalized_name\n", "0 Country country\n", "1 Age age\n", "2 Gender gender\n", "3 Exercise Level exercise_level\n", "4 Diet Type diet_type\n", "5 Sleep Hours sleep_hours\n", "6 Stress Level stress_level\n", "7 Mental Health Condition mental_health_condition\n", "8 Work Hours per Week work_hours_per_week\n", "9 Screen Time per Day (Hours) screen_time_per_day_hours\n", "10 Social Interaction Score social_interaction_score\n", "11 Happiness Score happiness_score" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "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", "
countryagegenderexercise_leveldiet_typesleep_hoursstress_levelmental_health_conditionwork_hours_per_weekscreen_time_per_day_hourssocial_interaction_scorehappiness_score
dtypeobjectint64objectobjectobjectfloat64objectobjectint64float64float64float64
\n", "
" ], "text/plain": [ " country age gender exercise_level diet_type sleep_hours \\\n", "dtype object int64 object object object float64 \n", "\n", " stress_level mental_health_condition work_hours_per_week \\\n", "dtype object object int64 \n", "\n", " screen_time_per_day_hours social_interaction_score happiness_score \n", "dtype float64 float64 float64 " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "import os\n", "import re\n", "import math\n", "import json\n", "import warnings\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.neural_network import MLPRegressor\n", "from sklearn.inspection import permutation_importance\n", "from sklearn.metrics import r2_score, mean_squared_error, mean_absolute_error\n", "\n", "warnings.filterwarnings(\"ignore\")\n", "RANDOM_STATE = 42\n", "np.random.seed(RANDOM_STATE)\n", "\n", "def find_file(candidates):\n", " for path in candidates:\n", " if os.path.exists(path):\n", " return path\n", " data_dir = \"/mnt/data\"\n", " if os.path.isdir(data_dir):\n", " files = os.listdir(data_dir)\n", " for name in files:\n", " low = name.lower()\n", " if \"mental\" in low and \"lifestyle\" in low and low.endswith(\".csv\"):\n", " return os.path.join(data_dir, name)\n", " if \"anxiety\" in low and low.endswith(\".csv\"):\n", " return os.path.join(data_dir, name)\n", " raise FileNotFoundError(\"CSV-файл не найден. Поместите его рядом с ноутбуком или укажите корректное имя.\")\n", "\n", "file_candidates = [\n", " \"Mental_Health_Lifestyle_Dataset(1).csv\",\n", " \"Mental_Health_Lifestyle_Dataset.csv\",\n", " \"mental_health_lifestyle_dataset.csv\",\n", " \"mental_health_lifestyle.csv\",\n", " \"/mnt/data/Mental_Health_Lifestyle_Dataset(1).csv\",\n", " \"/mnt/data/Mental_Health_Lifestyle_Dataset.csv\",\n", " \"/mnt/data/mental_health_lifestyle_dataset.csv\",\n", "]\n", "file_path = find_file(file_candidates)\n", "print(\"Используемый файл:\", file_path)\n", "\n", "df = pd.read_csv(file_path)\n", "print(\"Форма датасета:\", df.shape)\n", "display(df.head())\n", "\n", "def normalize_col(col):\n", " col = str(col).strip().lower()\n", " col = col.replace(\"%\", \" pct \")\n", " col = re.sub(r\"[^a-zа-я0-9]+\", \"_\", col)\n", " col = re.sub(r\"_+\", \"_\", col).strip(\"_\")\n", " return col\n", "\n", "orig_cols = df.columns.tolist()\n", "df.columns = [normalize_col(c) for c in df.columns]\n", "display(pd.DataFrame({\"original_name\": orig_cols, \"normalized_name\": df.columns}))\n", "\n", "print(\"Типы данных:\")\n", "display(df.dtypes.to_frame(\"dtype\").T)\n" ] }, { "cell_type": "code", "execution_count": 2, "id": "0cfbe798", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Выбранная зависимая переменная: happiness_score\n", "Найденные идентификаторы / потенциально несодержательные поля: []\n", "Число объектов после удаления пропусков по целевой переменной: 3000\n", "Число исходных признаков: 11\n", "Числовые признаки: ['age', 'sleep_hours', 'work_hours_per_week', 'screen_time_per_day_hours', 'social_interaction_score']\n", "Категориальные признаки: ['country', 'gender', 'exercise_level', 'diet_type', 'stress_level', 'mental_health_condition']\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", "
countryagegenderexercise_leveldiet_typesleep_hoursstress_levelmental_health_conditionwork_hours_per_weekscreen_time_per_day_hourssocial_interaction_score
0Brazil48MaleLowVegetarian6.3LowNaN214.07.8
1Australia31MaleModerateVegan4.9LowPTSD485.28.2
2Japan37FemaleLowVegetarian7.2HighNaN434.79.6
3Brazil35MaleLowVegan7.2LowDepression432.28.2
4Germany46MaleLowBalanced7.3LowAnxiety353.64.7
\n", "
" ], "text/plain": [ " country age gender exercise_level diet_type sleep_hours \\\n", "0 Brazil 48 Male Low Vegetarian 6.3 \n", "1 Australia 31 Male Moderate Vegan 4.9 \n", "2 Japan 37 Female Low Vegetarian 7.2 \n", "3 Brazil 35 Male Low Vegan 7.2 \n", "4 Germany 46 Male Low Balanced 7.3 \n", "\n", " stress_level mental_health_condition work_hours_per_week \\\n", "0 Low NaN 21 \n", "1 Low PTSD 48 \n", "2 High NaN 43 \n", "3 Low Depression 43 \n", "4 Low Anxiety 35 \n", "\n", " screen_time_per_day_hours social_interaction_score \n", "0 4.0 7.8 \n", "1 5.2 8.2 \n", "2 4.7 9.6 \n", "3 2.2 8.2 \n", "4 3.6 4.7 " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "def choose_target(dataframe):\n", " cols = list(dataframe.columns)\n", " preferred = [\n", " \"happiness_score\",\n", " \"score\",\n", " \"target\",\n", " \"price\",\n", " \"amount\",\n", " \"depression_score\",\n", " \"sales\",\n", " \"profit\",\n", " ]\n", " for col in preferred:\n", " if col in cols:\n", " ser = pd.to_numeric(dataframe[col], errors=\"coerce\")\n", " if ser.notna().mean() >= 0.7:\n", " return col\n", "\n", " numeric_candidates = []\n", " for col in cols:\n", " ser = pd.to_numeric(dataframe[col], errors=\"coerce\")\n", " nunique = ser.nunique(dropna=True)\n", " if ser.notna().mean() >= 0.7 and nunique > 10:\n", " numeric_candidates.append((col, nunique))\n", " if numeric_candidates:\n", " return numeric_candidates[-1][0]\n", " raise ValueError(\"Не удалось автоматически определить числовую зависимую переменную.\")\n", "\n", "target_col = choose_target(df)\n", "print(\"Выбранная зависимая переменная:\", target_col)\n", "\n", "id_like_cols = [c for c in df.columns if any(tok in c for tok in [\"id\", \"name\"]) and c != target_col]\n", "print(\"Найденные идентификаторы / потенциально несодержательные поля:\", id_like_cols)\n", "\n", "y = pd.to_numeric(df[target_col], errors=\"coerce\")\n", "X = df.drop(columns=[target_col] + id_like_cols).copy()\n", "\n", "# приведение числовых столбцов\n", "for col in X.columns:\n", " converted = pd.to_numeric(X[col], errors=\"ignore\")\n", " X[col] = converted\n", "\n", "# удаляем строки без цели\n", "mask = y.notna()\n", "X = X.loc[mask].reset_index(drop=True)\n", "y = y.loc[mask].reset_index(drop=True)\n", "\n", "print(\"Число объектов после удаления пропусков по целевой переменной:\", len(y))\n", "print(\"Число исходных признаков:\", X.shape[1])\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(\"Числовые признаки:\", numeric_features)\n", "print(\"Категориальные признаки:\", categorical_features)\n", "\n", "display(X.head())\n" ] }, { "cell_type": "code", "execution_count": 3, "id": "7807e1b0", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Размеры выборок:\n", "train: (1800, 11) (1800,)\n", "val: (600, 11) (600,)\n", "test: (600, 11) (600,)\n", "Число признаков после кодирования: 30\n", "Первые 20 преобразованных признаков:\n", "['num__age', 'num__sleep_hours', 'num__work_hours_per_week', 'num__screen_time_per_day_hours', 'num__social_interaction_score', 'cat__country_Australia', 'cat__country_Brazil', 'cat__country_Canada', 'cat__country_Germany', 'cat__country_India', 'cat__country_Japan', 'cat__country_USA', 'cat__gender_Female', 'cat__gender_Male', 'cat__gender_Other', 'cat__exercise_level_High', 'cat__exercise_level_Low', 'cat__exercise_level_Moderate', 'cat__diet_type_Balanced', 'cat__diet_type_Junk Food']\n" ] } ], "source": [ "\n", "X_trainval, X_test, y_trainval, y_test = train_test_split(\n", " X, y, test_size=0.2, random_state=RANDOM_STATE\n", ")\n", "X_train, X_val, y_train, y_val = train_test_split(\n", " X_trainval, y_trainval, test_size=0.25, random_state=RANDOM_STATE\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)\n", "\n", "def make_ohe():\n", " try:\n", " return OneHotEncoder(handle_unknown=\"ignore\", sparse_output=False)\n", " except TypeError:\n", " return OneHotEncoder(handle_unknown=\"ignore\", sparse=False)\n", "\n", "preprocessor = ColumnTransformer(\n", " transformers=[\n", " (\"num\", Pipeline([\n", " (\"imputer\", SimpleImputer(strategy=\"median\")),\n", " (\"scaler\", StandardScaler()),\n", " ]), numeric_features),\n", " (\"cat\", Pipeline([\n", " (\"imputer\", SimpleImputer(strategy=\"most_frequent\")),\n", " (\"ohe\", make_ohe()),\n", " ]), 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_from_trainval_proc = preprocessor.transform(X_test)\n", "\n", "try:\n", " feature_names = preprocessor.get_feature_names_out()\n", "except Exception:\n", " feature_names = [f\"feature_{i}\" for i in range(X_train_proc.shape[1])]\n", "\n", "print(\"Число признаков после кодирования:\", len(feature_names))\n", "print(\"Первые 20 преобразованных признаков:\")\n", "print(list(feature_names[:20]))\n" ] }, { "cell_type": "code", "execution_count": 4, "id": "8f4120a2", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Метрики базовой нейросети:\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", "
splitr2rmsemae
0train0.0516862.4674462.120161
1validation-0.0717692.6725842.290122
2test-0.0254782.6347322.256196
\n", "
" ], "text/plain": [ " split r2 rmse mae\n", "0 train 0.051686 2.467446 2.120161\n", "1 validation -0.071769 2.672584 2.290122\n", "2 test -0.025478 2.634732 2.256196" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "image/png": "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", "text/plain": [ "
" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "def regression_metrics(y_true, y_pred):\n", " return {\n", " \"r2\": float(r2_score(y_true, y_pred)),\n", " \"rmse\": float(np.sqrt(mean_squared_error(y_true, y_pred))),\n", " \"mae\": float(mean_absolute_error(y_true, y_pred)),\n", " }\n", "\n", "baseline_hidden = (64, 32)\n", "baseline_model = MLPRegressor(\n", " hidden_layer_sizes=baseline_hidden,\n", " activation=\"relu\",\n", " solver=\"adam\",\n", " alpha=1e-4,\n", " learning_rate_init=1e-3,\n", " max_iter=400,\n", " early_stopping=True,\n", " validation_fraction=0.15,\n", " n_iter_no_change=20,\n", " random_state=RANDOM_STATE\n", ")\n", "baseline_model.fit(X_train_proc, y_train)\n", "\n", "pred_train_baseline = baseline_model.predict(X_train_proc)\n", "pred_val_baseline = baseline_model.predict(X_val_proc)\n", "pred_test_baseline = baseline_model.predict(X_test_proc)\n", "\n", "baseline_metrics = pd.DataFrame([\n", " {\"split\": \"train\", **regression_metrics(y_train, pred_train_baseline)},\n", " {\"split\": \"validation\", **regression_metrics(y_val, pred_val_baseline)},\n", " {\"split\": \"test\", **regression_metrics(y_test, pred_test_baseline)},\n", "])\n", "\n", "print(\"Метрики базовой нейросети:\")\n", "display(baseline_metrics)\n", "\n", "plt.figure(figsize=(8, 4))\n", "plt.plot(baseline_model.loss_curve_)\n", "plt.title(\"Кривая обучения базовой нейросети\")\n", "plt.xlabel(\"Итерация\")\n", "plt.ylabel(\"Loss\")\n", "plt.grid(True)\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": 5, "id": "c0c544b1", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Топ-10 признаков по перестановочной 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
0exercise_level0.0228010.011556
1stress_level0.0218120.014420
2diet_type0.0191730.008693
3mental_health_condition0.0107480.005653
4gender0.0078880.005160
5country0.0074680.007103
6work_hours_per_week0.0069360.006448
7age-0.0129040.008048
8sleep_hours-0.0130750.008175
9social_interaction_score-0.0164820.006457
\n", "
" ], "text/plain": [ " feature permutation_importance_mean \\\n", "0 exercise_level 0.022801 \n", "1 stress_level 0.021812 \n", "2 diet_type 0.019173 \n", "3 mental_health_condition 0.010748 \n", "4 gender 0.007888 \n", "5 country 0.007468 \n", "6 work_hours_per_week 0.006936 \n", "7 age -0.012904 \n", "8 sleep_hours -0.013075 \n", "9 social_interaction_score -0.016482 \n", "\n", " permutation_importance_std \n", "0 0.011556 \n", "1 0.014420 \n", "2 0.008693 \n", "3 0.005653 \n", "4 0.005160 \n", "5 0.007103 \n", "6 0.006448 \n", "7 0.008048 \n", "8 0.008175 \n", "9 0.006457 " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Отобранные признаки:\n", "['exercise_level', 'stress_level', 'diet_type', 'mental_health_condition', 'gender', 'country', 'work_hours_per_week']\n", "Число отобранных признаков: 7\n" ] } ], "source": [ "\n", "# Важность признаков по перестановочному подходу на валидационной выборке\n", "perm = permutation_importance(\n", " estimator=baseline_model,\n", " X=X_val_proc,\n", " y=y_val,\n", " scoring=\"neg_root_mean_squared_error\",\n", " n_repeats=10,\n", " random_state=RANDOM_STATE,\n", ")\n", "\n", "# Нужно агрегировать важности по исходным признакам\n", "proc_feature_names = np.array(feature_names)\n", "agg_rows = []\n", "for original_feature in X.columns:\n", " mask = np.array([name.startswith(f\"num__{original_feature}\") or name.startswith(f\"cat__{original_feature}_\") for name in proc_feature_names])\n", " if mask.sum() == 0:\n", " continue\n", " agg_rows.append({\n", " \"feature\": original_feature,\n", " \"permutation_importance_mean\": float(np.mean(perm.importances_mean[mask])),\n", " \"permutation_importance_std\": float(np.sqrt(np.mean(np.square(perm.importances_std[mask])))),\n", " })\n", "\n", "perm_df = pd.DataFrame(agg_rows).sort_values(\"permutation_importance_mean\", ascending=False).reset_index(drop=True)\n", "print(\"Топ-10 признаков по перестановочной важности:\")\n", "display(perm_df.head(10))\n", "\n", "positive_features = perm_df.loc[perm_df[\"permutation_importance_mean\"] > 0, \"feature\"].tolist()\n", "if len(positive_features) >= 4:\n", " selected_features = positive_features[:min(8, len(positive_features))]\n", "else:\n", " selected_features = perm_df[\"feature\"].head(min(6, len(perm_df))).tolist()\n", "\n", "print(\"Отобранные признаки:\")\n", "print(selected_features)\n", "print(\"Число отобранных признаков:\", len(selected_features))\n" ] }, { "cell_type": "code", "execution_count": 6, "id": "aa552126", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Число признаков после кодирования для второй модели: 26\n" ] } ], "source": [ "\n", "X_sel = X[selected_features].copy()\n", "\n", "numeric_features_sel = X_sel.select_dtypes(include=[np.number]).columns.tolist()\n", "categorical_features_sel = [c for c in X_sel.columns if c not in numeric_features_sel]\n", "\n", "preprocessor_sel = ColumnTransformer(\n", " transformers=[\n", " (\"num\", Pipeline([\n", " (\"imputer\", SimpleImputer(strategy=\"median\")),\n", " (\"scaler\", StandardScaler()),\n", " ]), numeric_features_sel),\n", " (\"cat\", Pipeline([\n", " (\"imputer\", SimpleImputer(strategy=\"most_frequent\")),\n", " (\"ohe\", make_ohe()),\n", " ]), categorical_features_sel),\n", " ],\n", " remainder=\"drop\"\n", ")\n", "\n", "X_sel_trainval = X_trainval[selected_features].copy()\n", "X_sel_test = X_test[selected_features].copy()\n", "X_sel_train = X_train[selected_features].copy()\n", "X_sel_val = X_val[selected_features].copy()\n", "\n", "X_sel_train_proc = preprocessor_sel.fit_transform(X_sel_train)\n", "X_sel_val_proc = preprocessor_sel.transform(X_sel_val)\n", "X_sel_trainval_proc = preprocessor_sel.fit_transform(X_sel_trainval)\n", "X_sel_test_proc = preprocessor_sel.transform(X_sel_test)\n", "\n", "try:\n", " feature_names_sel = preprocessor_sel.get_feature_names_out()\n", "except Exception:\n", " feature_names_sel = [f\"feature_{i}\" for i in range(X_sel_train_proc.shape[1])]\n", "\n", "print(\"Число признаков после кодирования для второй модели:\", len(feature_names_sel))\n" ] }, { "cell_type": "code", "execution_count": 7, "id": "ddd683b4", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Топ-10 конфигураций эволюционного поиска:\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", "
hidden_layersalphalearning_rate_initrmse_valmae_valr2_val
0(56,)0.0000810.0006602.5778212.2209580.002888
1(96,)0.0000060.0012742.5809102.2211980.000497
2(58, 29)0.0001450.0039172.5809382.2272310.000475
3(54, 27)0.0000030.0019582.5809522.2177380.000464
4(104, 52, 26)0.0010170.0001602.5855072.228321-0.003067
5(93,)0.0001850.0003492.5859202.224106-0.003387
6(126, 63, 32, 16)0.0048600.0019042.5866282.233122-0.003937
7(104, 52, 26)0.0000020.0003962.5871822.227062-0.004367
8(105, 52, 26)0.0016840.0001432.5873072.224707-0.004464
9(29, 14, 8)0.0009160.0014462.5879842.223008-0.004990
\n", "
" ], "text/plain": [ " hidden_layers alpha learning_rate_init rmse_val mae_val \\\n", "0 (56,) 0.000081 0.000660 2.577821 2.220958 \n", "1 (96,) 0.000006 0.001274 2.580910 2.221198 \n", "2 (58, 29) 0.000145 0.003917 2.580938 2.227231 \n", "3 (54, 27) 0.000003 0.001958 2.580952 2.217738 \n", "4 (104, 52, 26) 0.001017 0.000160 2.585507 2.228321 \n", "5 (93,) 0.000185 0.000349 2.585920 2.224106 \n", "6 (126, 63, 32, 16) 0.004860 0.001904 2.586628 2.233122 \n", "7 (104, 52, 26) 0.000002 0.000396 2.587182 2.227062 \n", "8 (105, 52, 26) 0.001684 0.000143 2.587307 2.224707 \n", "9 (29, 14, 8) 0.000916 0.001446 2.587984 2.223008 \n", "\n", " r2_val \n", "0 0.002888 \n", "1 0.000497 \n", "2 0.000475 \n", "3 0.000464 \n", "4 -0.003067 \n", "5 -0.003387 \n", "6 -0.003937 \n", "7 -0.004367 \n", "8 -0.004464 \n", "9 -0.004990 " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Лучшая архитектура, найденная эволюционным поиском: (56,)\n", "alpha: 8.067758731492677e-05\n", "learning_rate_init: 0.0006600577523657266\n" ] } ], "source": [ "\n", "search_results = []\n", "\n", "def decode_candidate(vec):\n", " # [layers_raw, units_raw, alpha_raw, lr_raw]\n", " layers = int(np.clip(round(vec[0]), 1, 4))\n", " base_units = int(np.clip(round(vec[1]), 16, 128))\n", " alpha = float(10 ** np.clip(vec[2], -6, -2))\n", " lr = float(10 ** np.clip(vec[3], -4, -2))\n", "\n", " arch = []\n", " units = base_units\n", " for _ in range(layers):\n", " arch.append(max(8, int(round(units))))\n", " units = max(8, units / 2)\n", " return tuple(arch), alpha, lr\n", "\n", "_eval_cache = {}\n", "\n", "def evaluate_candidate(vec):\n", " arch, alpha, lr = decode_candidate(vec)\n", " key = (arch, round(alpha, 8), round(lr, 8))\n", " if key in _eval_cache:\n", " return _eval_cache[key]\n", "\n", " model = MLPRegressor(\n", " hidden_layer_sizes=arch,\n", " activation=\"relu\",\n", " solver=\"adam\",\n", " alpha=alpha,\n", " learning_rate_init=lr,\n", " max_iter=250,\n", " early_stopping=True,\n", " validation_fraction=0.15,\n", " n_iter_no_change=15,\n", " random_state=RANDOM_STATE\n", " )\n", " model.fit(X_sel_train_proc, y_train)\n", " pred_val = model.predict(X_sel_val_proc)\n", " val_rmse = float(np.sqrt(mean_squared_error(y_val, pred_val)))\n", " val_mae = float(mean_absolute_error(y_val, pred_val))\n", " val_r2 = float(r2_score(y_val, pred_val))\n", "\n", " row = {\n", " \"hidden_layers\": arch,\n", " \"alpha\": alpha,\n", " \"learning_rate_init\": lr,\n", " \"rmse_val\": val_rmse,\n", " \"mae_val\": val_mae,\n", " \"r2_val\": val_r2,\n", " }\n", " search_results.append(row)\n", " _eval_cache[key] = val_rmse\n", " return val_rmse\n", "\n", "use_scipy = True\n", "try:\n", " from scipy.optimize import differential_evolution\n", "except Exception:\n", " use_scipy = False\n", "\n", "bounds = [\n", " (1, 4), # число слоёв\n", " (16, 128), # базовое число нейронов\n", " (-6, -2), # log10(alpha)\n", " (-4, -2), # log10(lr)\n", "]\n", "\n", "if use_scipy:\n", " result = differential_evolution(\n", " func=evaluate_candidate,\n", " bounds=bounds,\n", " strategy=\"best1bin\",\n", " maxiter=6,\n", " popsize=6,\n", " tol=0.01,\n", " polish=False,\n", " seed=RANDOM_STATE,\n", " updating=\"deferred\",\n", " workers=1,\n", " )\n", " best_arch, best_alpha, best_lr = decode_candidate(result.x)\n", "else:\n", " # упрощённый эволюционный поиск без scipy\n", " population = [np.array([\n", " np.random.uniform(1, 4),\n", " np.random.uniform(16, 128),\n", " np.random.uniform(-6, -2),\n", " np.random.uniform(-4, -2)\n", " ]) for _ in range(18)]\n", "\n", " for gen in range(8):\n", " scores = np.array([evaluate_candidate(ind) for ind in population])\n", " order = np.argsort(scores)\n", " population = [population[i] for i in order[:9]]\n", " new_population = population.copy()\n", " while len(new_population) < 18:\n", " a, b = np.random.choice(len(population), 2, replace=False)\n", " child = 0.5 * population[a] + 0.5 * population[b]\n", " child = child + np.random.normal(0, [0.4, 8, 0.4, 0.3], size=4)\n", " child[0] = np.clip(child[0], 1, 4)\n", " child[1] = np.clip(child[1], 16, 128)\n", " child[2] = np.clip(child[2], -6, -2)\n", " child[3] = np.clip(child[3], -4, -2)\n", " new_population.append(child)\n", " population = new_population\n", " best_vec = min(population, key=evaluate_candidate)\n", " best_arch, best_alpha, best_lr = decode_candidate(best_vec)\n", "\n", "search_df = pd.DataFrame(search_results).sort_values([\"rmse_val\", \"mae_val\"]).drop_duplicates().reset_index(drop=True)\n", "print(\"Топ-10 конфигураций эволюционного поиска:\")\n", "display(search_df.head(10))\n", "\n", "print(\"Лучшая архитектура, найденная эволюционным поиском:\", best_arch)\n", "print(\"alpha:\", best_alpha)\n", "print(\"learning_rate_init:\", best_lr)\n" ] }, { "cell_type": "code", "execution_count": 8, "id": "687ddc75", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Метрики второй модели (по выбранным признакам):\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", "
splitr2rmsemae
0trainval0.0081972.535432.192504
1test-0.0115362.616762.264249
\n", "
" ], "text/plain": [ " split r2 rmse mae\n", "0 trainval 0.008197 2.53543 2.192504\n", "1 test -0.011536 2.61676 2.264249" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "image/png": "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", "text/plain": [ "
" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "best_model = MLPRegressor(\n", " hidden_layer_sizes=best_arch,\n", " activation=\"relu\",\n", " solver=\"adam\",\n", " alpha=best_alpha,\n", " learning_rate_init=best_lr,\n", " max_iter=500,\n", " early_stopping=True,\n", " validation_fraction=0.15,\n", " n_iter_no_change=20,\n", " random_state=RANDOM_STATE\n", ")\n", "best_model.fit(X_sel_trainval_proc, y_trainval)\n", "\n", "pred_trainval_best = best_model.predict(X_sel_trainval_proc)\n", "pred_test_best = best_model.predict(X_sel_test_proc)\n", "\n", "best_metrics = pd.DataFrame([\n", " {\"split\": \"trainval\", **regression_metrics(y_trainval, pred_trainval_best)},\n", " {\"split\": \"test\", **regression_metrics(y_test, pred_test_best)},\n", "])\n", "\n", "print(\"Метрики второй модели (по выбранным признакам):\")\n", "display(best_metrics)\n", "\n", "plt.figure(figsize=(8, 4))\n", "plt.plot(best_model.loss_curve_)\n", "plt.title(\"Кривая обучения второй нейросети\")\n", "plt.xlabel(\"Итерация\")\n", "plt.ylabel(\"Loss\")\n", "plt.grid(True)\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": 9, "id": "63502f80", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Сравнение моделей:\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", "
modelr2_testrmse_testmae_testr2_valrmse_valmae_valn_original_featuresn_processed_featureshidden_layers
0Нейросеть по отобранным признакам + эволюционн...-0.0115362.6167602.2642490.0106862.5677212.214208726(56,)
1Базовая нейросеть по всем признакам-0.0254782.6347322.256196-0.0717692.6725842.2901221130(64, 32)
\n", "
" ], "text/plain": [ " model r2_test rmse_test \\\n", "0 Нейросеть по отобранным признакам + эволюционн... -0.011536 2.616760 \n", "1 Базовая нейросеть по всем признакам -0.025478 2.634732 \n", "\n", " mae_test r2_val rmse_val mae_val n_original_features \\\n", "0 2.264249 0.010686 2.567721 2.214208 7 \n", "1 2.256196 -0.071769 2.672584 2.290122 11 \n", "\n", " n_processed_features hidden_layers \n", "0 26 (56,) \n", "1 30 (64, 32) " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "comparison_df = pd.DataFrame([\n", " {\n", " \"model\": \"Базовая нейросеть по всем признакам\",\n", " \"r2_test\": regression_metrics(y_test, pred_test_baseline)[\"r2\"],\n", " \"rmse_test\": regression_metrics(y_test, pred_test_baseline)[\"rmse\"],\n", " \"mae_test\": regression_metrics(y_test, pred_test_baseline)[\"mae\"],\n", " \"r2_val\": regression_metrics(y_val, pred_val_baseline)[\"r2\"],\n", " \"rmse_val\": regression_metrics(y_val, pred_val_baseline)[\"rmse\"],\n", " \"mae_val\": regression_metrics(y_val, pred_val_baseline)[\"mae\"],\n", " \"n_original_features\": X.shape[1],\n", " \"n_processed_features\": len(feature_names),\n", " \"hidden_layers\": str(baseline_hidden),\n", " },\n", " {\n", " \"model\": \"Нейросеть по отобранным признакам + эволюционный поиск архитектуры\",\n", " \"r2_test\": regression_metrics(y_test, pred_test_best)[\"r2\"],\n", " \"rmse_test\": regression_metrics(y_test, pred_test_best)[\"rmse\"],\n", " \"mae_test\": regression_metrics(y_test, pred_test_best)[\"mae\"],\n", " \"r2_val\": regression_metrics(y_val, best_model.predict(X_sel_val_proc))[\"r2\"],\n", " \"rmse_val\": regression_metrics(y_val, best_model.predict(X_sel_val_proc))[\"rmse\"],\n", " \"mae_val\": regression_metrics(y_val, best_model.predict(X_sel_val_proc))[\"mae\"],\n", " \"n_original_features\": len(selected_features),\n", " \"n_processed_features\": len(feature_names_sel),\n", " \"hidden_layers\": str(best_arch),\n", " },\n", "]).sort_values([\"rmse_test\", \"mae_test\"]).reset_index(drop=True)\n", "\n", "print(\"Сравнение моделей:\")\n", "display(comparison_df)\n" ] }, { "cell_type": "code", "execution_count": 10, "id": "662ca5d8", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "=== КЛЮЧЕВАЯ ИНФОРМАЦИЯ ДЛЯ ВЫВОДА ===\n", "Зависимая переменная: happiness_score\n", "\n", "Число исходных признаков: 11\n", "Число признаков после кодирования (все признаки): 30\n", "Число отобранных исходных признаков: 7\n", "Число признаков после кодирования (отобранные признаки): 26\n", "Фиксированная архитектура первой модели: (64, 32)\n", "Архитектура, найденная эволюционным поиском: (56,)\n", "alpha найденной модели: 8.067758731492677e-05\n", "learning_rate_init найденной модели: 0.0006600577523657266\n", "\n", "Отобранные признаки:\n", "['exercise_level', 'stress_level', 'diet_type', 'mental_health_condition', 'gender', 'country', 'work_hours_per_week']\n", "\n", "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", "
modelr2_testrmse_testmae_testr2_valrmse_valmae_valn_original_featuresn_processed_featureshidden_layers
0Нейросеть по отобранным признакам + эволюционн...-0.0115362.6167602.2642490.0106862.5677212.214208726(56,)
1Базовая нейросеть по всем признакам-0.0254782.6347322.256196-0.0717692.6725842.2901221130(64, 32)
\n", "
" ], "text/plain": [ " model r2_test rmse_test \\\n", "0 Нейросеть по отобранным признакам + эволюционн... -0.011536 2.616760 \n", "1 Базовая нейросеть по всем признакам -0.025478 2.634732 \n", "\n", " mae_test r2_val rmse_val mae_val n_original_features \\\n", "0 2.264249 0.010686 2.567721 2.214208 7 \n", "1 2.256196 -0.071769 2.672584 2.290122 11 \n", "\n", " n_processed_features hidden_layers \n", "0 26 (56,) \n", "1 30 (64, 32) " ] }, "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", "
modelr2_testrmse_testmae_testr2_valrmse_valmae_valn_original_featuresn_processed_featureshidden_layers
0Нейросеть по отобранным признакам + эволюционн...-0.0115362.616762.2642490.0106862.5677212.214208726(56,)
\n", "
" ], "text/plain": [ " model r2_test rmse_test \\\n", "0 Нейросеть по отобранным признакам + эволюционн... -0.011536 2.61676 \n", "\n", " mae_test r2_val rmse_val mae_val n_original_features \\\n", "0 2.264249 0.010686 2.567721 2.214208 7 \n", "\n", " n_processed_features hidden_layers \n", "0 26 (56,) " ] }, "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
0exercise_level0.0228010.011556
1stress_level0.0218120.014420
2diet_type0.0191730.008693
3mental_health_condition0.0107480.005653
4gender0.0078880.005160
5country0.0074680.007103
6work_hours_per_week0.0069360.006448
7age-0.0129040.008048
8sleep_hours-0.0130750.008175
9social_interaction_score-0.0164820.006457
\n", "
" ], "text/plain": [ " feature permutation_importance_mean \\\n", "0 exercise_level 0.022801 \n", "1 stress_level 0.021812 \n", "2 diet_type 0.019173 \n", "3 mental_health_condition 0.010748 \n", "4 gender 0.007888 \n", "5 country 0.007468 \n", "6 work_hours_per_week 0.006936 \n", "7 age -0.012904 \n", "8 sleep_hours -0.013075 \n", "9 social_interaction_score -0.016482 \n", "\n", " permutation_importance_std \n", "0 0.011556 \n", "1 0.014420 \n", "2 0.008693 \n", "3 0.005653 \n", "4 0.005160 \n", "5 0.007103 \n", "6 0.006448 \n", "7 0.008048 \n", "8 0.008175 \n", "9 0.006457 " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Топ-10 конфигураций эволюционного поиска:\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", "
hidden_layersalphalearning_rate_initrmse_valmae_valr2_val
0(56,)0.0000810.0006602.5778212.2209580.002888
1(96,)0.0000060.0012742.5809102.2211980.000497
2(58, 29)0.0001450.0039172.5809382.2272310.000475
3(54, 27)0.0000030.0019582.5809522.2177380.000464
4(104, 52, 26)0.0010170.0001602.5855072.228321-0.003067
5(93,)0.0001850.0003492.5859202.224106-0.003387
6(126, 63, 32, 16)0.0048600.0019042.5866282.233122-0.003937
7(104, 52, 26)0.0000020.0003962.5871822.227062-0.004367
8(105, 52, 26)0.0016840.0001432.5873072.224707-0.004464
9(29, 14, 8)0.0009160.0014462.5879842.223008-0.004990
\n", "
" ], "text/plain": [ " hidden_layers alpha learning_rate_init rmse_val mae_val \\\n", "0 (56,) 0.000081 0.000660 2.577821 2.220958 \n", "1 (96,) 0.000006 0.001274 2.580910 2.221198 \n", "2 (58, 29) 0.000145 0.003917 2.580938 2.227231 \n", "3 (54, 27) 0.000003 0.001958 2.580952 2.217738 \n", "4 (104, 52, 26) 0.001017 0.000160 2.585507 2.228321 \n", "5 (93,) 0.000185 0.000349 2.585920 2.224106 \n", "6 (126, 63, 32, 16) 0.004860 0.001904 2.586628 2.233122 \n", "7 (104, 52, 26) 0.000002 0.000396 2.587182 2.227062 \n", "8 (105, 52, 26) 0.001684 0.000143 2.587307 2.224707 \n", "9 (29, 14, 8) 0.000916 0.001446 2.587984 2.223008 \n", "\n", " r2_val \n", "0 0.002888 \n", "1 0.000497 \n", "2 0.000475 \n", "3 0.000464 \n", "4 -0.003067 \n", "5 -0.003387 \n", "6 -0.003937 \n", "7 -0.004367 \n", "8 -0.004464 \n", "9 -0.004990 " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Последние эпохи обучения базовой нейросети:\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", "
epochtrain_loss
53542.892276
54552.885353
55562.878268
56572.873679
57582.862408
58592.860602
59602.849065
60612.841779
61622.832322
62632.824808
\n", "
" ], "text/plain": [ " epoch train_loss\n", "53 54 2.892276\n", "54 55 2.885353\n", "55 56 2.878268\n", "56 57 2.873679\n", "57 58 2.862408\n", "58 59 2.860602\n", "59 60 2.849065\n", "60 61 2.841779\n", "61 62 2.832322\n", "62 63 2.824808" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Последние эпохи обучения модели с найденной архитектурой:\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", "
epochtrain_loss
61623.156066
62633.156377
63643.154805
64653.154852
65663.153527
66673.154119
67683.153056
68693.152203
69703.151758
70713.151723
\n", "
" ], "text/plain": [ " epoch train_loss\n", "61 62 3.156066\n", "62 63 3.156377\n", "63 64 3.154805\n", "64 65 3.154852\n", "65 66 3.153527\n", "66 67 3.154119\n", "67 68 3.153056\n", "68 69 3.152203\n", "69 70 3.151758\n", "70 71 3.151723" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Первые 10 фактических и предсказанных значений на тесте:\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_baseliney_pred_selected_evolution
05.64.8283115.289400
19.85.6990565.482407
26.04.6198575.499895
38.65.0559704.970277
42.86.0084715.316428
55.35.4790885.267063
62.25.3146685.545767
74.55.2690715.413516
85.15.5030095.547910
94.34.8465075.304741
\n", "
" ], "text/plain": [ " y_true y_pred_baseline y_pred_selected_evolution\n", "0 5.6 4.828311 5.289400\n", "1 9.8 5.699056 5.482407\n", "2 6.0 4.619857 5.499895\n", "3 8.6 5.055970 4.970277\n", "4 2.8 6.008471 5.316428\n", "5 5.3 5.479088 5.267063\n", "6 2.2 5.314668 5.545767\n", "7 4.5 5.269071 5.413516\n", "8 5.1 5.503009 5.547910\n", "9 4.3 4.846507 5.304741" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\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 (все признаки): {len(feature_names)}\")\n", "print(f\"Число отобранных исходных признаков: {len(selected_features)}\")\n", "print(f\"Число признаков после кодирования (отобранные признаки): {len(feature_names_sel)}\")\n", "print(f\"Фиксированная архитектура первой модели: {baseline_hidden}\")\n", "print(f\"Архитектура, найденная эволюционным поиском: {best_arch}\")\n", "print(f\"alpha найденной модели: {best_alpha}\")\n", "print(f\"learning_rate_init найденной модели: {best_lr}\")\n", "print()\n", "\n", "print(\"Отобранные признаки:\")\n", "print(selected_features)\n", "print()\n", "\n", "print(\"Метрики моделей на тестовой выборке:\")\n", "display(comparison_df)\n", "\n", "print(\"Лучшая модель:\")\n", "best_row = comparison_df.sort_values([\"rmse_test\", \"mae_test\", \"r2_test\"], ascending=[True, True, False]).head(1)\n", "display(best_row)\n", "\n", "print(\"Топ-10 признаков по важности:\")\n", "display(perm_df.head(10))\n", "\n", "print(\"Топ-10 конфигураций эволюционного поиска:\")\n", "display(search_df.head(10))\n", "\n", "print(\"Последние эпохи обучения базовой нейросети:\")\n", "display(pd.DataFrame({\n", " \"epoch\": np.arange(1, len(baseline_model.loss_curve_) + 1),\n", " \"train_loss\": baseline_model.loss_curve_,\n", "}).tail(10))\n", "\n", "print(\"Последние эпохи обучения модели с найденной архитектурой:\")\n", "display(pd.DataFrame({\n", " \"epoch\": np.arange(1, len(best_model.loss_curve_) + 1),\n", " \"train_loss\": best_model.loss_curve_,\n", "}).tail(10))\n", "\n", "pred_example_df = pd.DataFrame({\n", " \"y_true\": y_test.reset_index(drop=True).head(10),\n", " \"y_pred_baseline\": pd.Series(pred_test_baseline).head(10),\n", " \"y_pred_selected_evolution\": pd.Series(pred_test_best).head(10),\n", "})\n", "print(\"Первые 10 фактических и предсказанных значений на тесте:\")\n", "display(pred_example_df)\n" ] }, { "cell_type": "markdown", "id": "e6a08f26", "metadata": {}, "source": [ "Итог\n", "\n", "Зависимой переменной выбрана happiness_score. Для прогноза были построены две нейросети: базовая модель по всем признакам и модель по 7 отобранным признакам с архитектурой, найденной эволюционным поиском.\n", "\n", "По тестовым метрикам обе модели показали неприемлемое качество: у обеих R² < 0, а предсказания слабо отражают реальную вариацию целевой переменной. Это означает, что модели фактически не дают полезного регрессионного описания данных.\n", "\n", "Наиболее важными признаками по permutation importance оказались exercise_level, stress_level, diet_type, mental_health_condition, gender, country, work_hours_per_week, однако даже их использование не привело к практически пригодному результату.\n", "\n", "Следовательно, все построенные модели в данной постановке задачи неудовлетворительны и не представляют практической ценности для использования." ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "name": "python" } }, "nbformat": 4, "nbformat_minor": 5 }