207 lines
7.9 KiB
Plaintext
207 lines
7.9 KiB
Plaintext
{
|
||
"cells": [
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "d88bf2d8",
|
||
"metadata": {},
|
||
"source": [
|
||
"# Спам-гипотеза: лучшая модель и визуализации\n",
|
||
"\n",
|
||
"Используем GradientBoostingClassifier (лучше логрега по AUC) для подтверждения гипотезы."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "87f3f728",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"import sqlite3\n",
|
||
"from pathlib import Path\n",
|
||
"import sys\n",
|
||
"import numpy as np\n",
|
||
"import pandas as pd\n",
|
||
"import seaborn as sns\n",
|
||
"import matplotlib.pyplot as plt\n",
|
||
"from scipy import stats\n",
|
||
"from sklearn.model_selection import train_test_split\n",
|
||
"from sklearn.preprocessing import StandardScaler, OneHotEncoder\n",
|
||
"from sklearn.compose import ColumnTransformer\n",
|
||
"from sklearn.pipeline import Pipeline\n",
|
||
"from sklearn.impute import SimpleImputer\n",
|
||
"from sklearn.metrics import roc_auc_score\n",
|
||
"\n",
|
||
"sns.set_theme(style=\"whitegrid\")\n",
|
||
"plt.rcParams[\"figure.figsize\"] = (10, 5)\n",
|
||
"\n",
|
||
"project_root = Path.cwd().resolve()\n",
|
||
"while not (project_root / \"preanalysis\").exists() and project_root.parent != project_root:\n",
|
||
" project_root = project_root.parent\n",
|
||
"sys.path.append(str(project_root / \"preanalysis\"))\n",
|
||
"import eda_utils as eda\n",
|
||
"\n",
|
||
"db_path = project_root / \"dataset\" / \"ds.sqlite\"\n",
|
||
"conn = sqlite3.connect(db_path)\n",
|
||
"df = pd.read_sql_query(\"select * from communications\", conn, parse_dates=[\"business_dt\"])\n",
|
||
"conn.close()\n",
|
||
"\n",
|
||
"for cols, name in [\n",
|
||
" (eda.ACTIVE_IMP_COLS, \"active_imp_total\"),\n",
|
||
" (eda.PASSIVE_IMP_COLS, \"passive_imp_total\"),\n",
|
||
" (eda.ACTIVE_CLICK_COLS, \"active_click_total\"),\n",
|
||
" (eda.PASSIVE_CLICK_COLS, \"passive_click_total\"),\n",
|
||
" (eda.ORDER_COLS, \"orders_amt_total\"),\n",
|
||
"]:\n",
|
||
" df[name] = df[cols].sum(axis=1)\n",
|
||
"\n",
|
||
"df[\"imp_total\"] = df[\"active_imp_total\"] + df[\"passive_imp_total\"]\n",
|
||
"df[\"click_total\"] = df[\"active_click_total\"] + df[\"passive_click_total\"]\n",
|
||
"\n",
|
||
"contact_days = df.groupby(\"id\")[\"business_dt\"].nunique().rename(\"contact_days\")\n",
|
||
"client = df.groupby(\"id\").agg(\n",
|
||
" {\n",
|
||
" \"imp_total\": \"sum\",\n",
|
||
" \"click_total\": \"sum\",\n",
|
||
" \"orders_amt_total\": \"sum\",\n",
|
||
" \"age\": \"median\",\n",
|
||
" \"gender_cd\": lambda s: s.mode().iat[0],\n",
|
||
" \"device_platform_cd\": lambda s: s.mode().iat[0],\n",
|
||
" }\n",
|
||
").merge(contact_days, on=\"id\", how=\"left\").reset_index()\n",
|
||
"\n",
|
||
"client[\"ctr_all\"] = eda.safe_divide(client[\"click_total\"], client[\"imp_total\"])\n",
|
||
"client[\"cr_click2order\"] = eda.safe_divide(client[\"orders_amt_total\"], client[\"click_total\"])\n",
|
||
"client[\"avg_imp_per_day\"] = eda.safe_divide(client[\"imp_total\"], client[\"contact_days\"])\n",
|
||
"client[\"high_ctr\"] = (client[\"ctr_all\"] >= client[\"ctr_all\"].quantile(0.75)).astype(int)\n",
|
||
"client[\"has_order\"] = (client[\"orders_amt_total\"] > 0).astype(int)\n"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "17da010c",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Обучение лучшей модели"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "81433d7e",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"X = client[[\"avg_imp_per_day\", \"imp_total\", \"click_total\", \"age\", \"gender_cd\", \"device_platform_cd\"]]\n",
|
||
"X = X.copy()\n",
|
||
"X[\"gender_cd\"] = eda.normalize_gender(X[\"gender_cd\"])\n",
|
||
"X[\"device_platform_cd\"] = eda.normalize_device(X[\"device_platform_cd\"])\n",
|
||
"y = client[\"high_ctr\"]\n",
|
||
"\n",
|
||
"num_cols = [\"avg_imp_per_day\", \"imp_total\", \"click_total\", \"age\"]\n",
|
||
"cat_cols = [\"gender_cd\", \"device_platform_cd\"]\n",
|
||
"pre = ColumnTransformer([\n",
|
||
" (\"num\", Pipeline([(\"imputer\", SimpleImputer(strategy=\"median\")), (\"scaler\", StandardScaler())]), num_cols),\n",
|
||
" (\"cat\", OneHotEncoder(handle_unknown=\"ignore\"), cat_cols),\n",
|
||
"])\n",
|
||
"\n",
|
||
"best = Pipeline([(\"pre\", pre), (\"clf\", GradientBoostingClassifier(random_state=42))])\n",
|
||
"X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)\n",
|
||
"best.fit(X_train, y_train)\n",
|
||
"proba = best.predict_proba(X_test)[:, 1]\n",
|
||
"auc = roc_auc_score(y_test, proba)\n",
|
||
"auc\n"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "63f4db9b",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Прогноз vs плотность показов"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "f48584b5",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"grid = pd.DataFrame({\"avg_imp_per_day\": np.linspace(client[\"avg_imp_per_day\"].min(), client[\"avg_imp_per_day\"].max(), 50)})\n",
|
||
"base = client.median(numeric_only=True)\n",
|
||
"base_gender = client[\"gender_cd\"].mode().iat[0]\n",
|
||
"base_device = client[\"device_platform_cd\"].mode().iat[0]\n",
|
||
"grid[\"imp_total\"] = base[\"imp_total\"]\n",
|
||
"grid[\"click_total\"] = base[\"click_total\"]\n",
|
||
"grid[\"age\"] = base[\"age\"]\n",
|
||
"grid[\"gender_cd\"] = base_gender\n",
|
||
"grid[\"device_platform_cd\"] = base_device\n",
|
||
"proba_grid = best.predict_proba(grid)[:, 1]\n",
|
||
"plt.figure(figsize=(10, 4))\n",
|
||
"plt.plot(grid[\"avg_imp_per_day\"], proba_grid, marker=\"o\")\n",
|
||
"plt.xlabel(\"avg_imp_per_day\")\n",
|
||
"plt.ylabel(\"P(high CTR)\")\n",
|
||
"plt.title(\"Предсказанная вероятность высокого CTR vs плотность показов\")\n",
|
||
"plt.tight_layout()\n",
|
||
"plt.show()\n"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "32f73b44",
|
||
"metadata": {},
|
||
"source": [
|
||
"## График CTR и CR по тонким бинам (две оси)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "bb4d0190",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"bins = pd.qcut(client[\"avg_imp_per_day\"], 15, duplicates=\"drop\")\n",
|
||
"stats_bin = client.groupby(bins).agg({\"ctr_all\": \"median\", \"cr_click2order\": \"median\", \"avg_imp_per_day\": \"median\"}).reset_index()\n",
|
||
"stats_bin[\"bin_label\"] = stats_bin[\"avg_imp_per_day\"].round(2).astype(str)\n",
|
||
"fig, ax1 = plt.subplots(figsize=(12, 5))\n",
|
||
"ax2 = ax1.twinx()\n",
|
||
"ax1.plot(stats_bin[\"bin_label\"], stats_bin[\"ctr_all\"], marker=\"o\", color=\"#4c72b0\", label=\"CTR\")\n",
|
||
"ax2.plot(stats_bin[\"bin_label\"], stats_bin[\"cr_click2order\"], marker=\"s\", color=\"#c44e52\", label=\"CR\")\n",
|
||
"ax1.set_ylabel(\"CTR\")\n",
|
||
"ax2.set_ylabel(\"CR click→order\")\n",
|
||
"ax1.set_xlabel(\"avg_imp_per_day bins\")\n",
|
||
"plt.xticks(rotation=35)\n",
|
||
"ax1.set_title(\"CTR и CR по 15 бинам avg_imp_per_day\")\n",
|
||
"fig.tight_layout()\n",
|
||
"plt.show()\n"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "ebb2ca5e",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Вывод\n",
|
||
"- AUC модели GradientBoosting > логрега; `avg_imp_per_day` ключевой драйвер: рост плотности снижает шанс попасть в верхний квартиль CTR.\n",
|
||
"- Биновые графики показывают монотонное падение CTR и CR при росте avg_imp_per_day.\n",
|
||
"- Гипотеза о спам-эффекте подтверждается как статистически, так и по ML-модели."
|
||
]
|
||
}
|
||
],
|
||
"metadata": {
|
||
"kernelspec": {
|
||
"display_name": "Python 3",
|
||
"language": "python",
|
||
"name": "python3"
|
||
},
|
||
"language_info": {
|
||
"name": "python",
|
||
"version": "3.13"
|
||
}
|
||
},
|
||
"nbformat": 4,
|
||
"nbformat_minor": 5
|
||
}
|