{ "cells": [ { "cell_type": "markdown", "id": "4e333c36", "metadata": {}, "source": [ "# DBbun Davis — Machine Learning Demo\n", "This notebook shows how to build ML models using the synthetic **Davis Square** dataset.\n", "\n", "We’ll do three examples:\n", "1. **Classification:** predict whether a public safety incident is **high severity** using incident, weather, and street features.\n", "2. **Regression:** predict **noise (dB)** from weather and traffic context.\n", "3. **Clustering:** group streets by **mobility signature** (mode mix) using K-Means.\n", "\n", "All **figures, tables, and models** are saved locally.\n" ] }, { "cell_type": "code", "execution_count": 1, "id": "4f2d9fff", "metadata": {}, "outputs": [], "source": [ "# --- User config ---\n", "#DATA_DIR = \"./out_dbbun_davis_medium\" # Set to your dataset folder\n", "DATA_DIR = \"./\"\n", "FIGS_DIR = \"./ml_figs\"\n", "TABLES_DIR = \"./ml_tables\"\n", "MODELS_DIR = \"./ml_models\"\n", "\n", "import os\n", "os.environ[\"OMP_NUM_THREADS\"] = \"1\"\n", "os.environ[\"MKL_NUM_THREADS\"] = \"1\" # extra-safe on Windows/MKL\n", "os.environ[\"NUMEXPR_NUM_THREADS\"] = \"1\" # extra-safe\n", "\n", "# --- Imports ---\n", "import os, json, pickle, re\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, cross_val_score\n", "from sklearn.preprocessing import OneHotEncoder, StandardScaler\n", "from sklearn.compose import ColumnTransformer\n", "from sklearn.pipeline import Pipeline\n", "from sklearn.metrics import classification_report, confusion_matrix, roc_auc_score, RocCurveDisplay, mean_absolute_error, r2_score\n", "from sklearn.linear_model import LogisticRegression, LinearRegression\n", "from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor\n", "from sklearn.cluster import KMeans\n", "\n", "os.makedirs(FIGS_DIR, exist_ok=True)\n", "os.makedirs(TABLES_DIR, exist_ok=True)\n", "os.makedirs(MODELS_DIR, exist_ok=True)\n", "\n", "def savefig(name):\n", " path = os.path.join(FIGS_DIR, name)\n", " plt.savefig(path, bbox_inches=\"tight\", dpi=144)\n", " plt.close()\n", " print(f\"Saved figure: {path}\")\n", "\n", "def savetab(df, name):\n", " path = os.path.join(TABLES_DIR, name)\n", " df.to_csv(path, index=False)\n", " print(f\"Saved table: {path}\")\n", "\n", "def savemodel(model, name):\n", " path = os.path.join(MODELS_DIR, name)\n", " with open(path, \"wb\") as f:\n", " pickle.dump(model, f)\n", " print(f\"Saved model: {path}\")" ] }, { "cell_type": "code", "execution_count": 2, "id": "0fd4f78b-44e7-424d-8f1c-3bbc29468cba", "metadata": {}, "outputs": [], "source": [ "import warnings\n", "warnings.filterwarnings(\"ignore\", category=FutureWarning)" ] }, { "cell_type": "code", "execution_count": 3, "id": "2a2efd9d-75b0-43f4-894b-f7347e3febdb", "metadata": {}, "outputs": [], "source": [ "pd.set_option('future.no_silent_downcasting', True)" ] }, { "cell_type": "code", "execution_count": 4, "id": "0ac6980b", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Loaded tables:\n", " streets: (15, 7) | safety: (12000, 8) | obs: (35000, 7) | weather: (101136, 7)\n" ] } ], "source": [ "# Load data\n", "def load_csv(name):\n", " return pd.read_csv(os.path.join(DATA_DIR, name))\n", "\n", "# Helper for WKT point -> (lon,lat)\n", "def wkt_point_to_lonlat(wkt):\n", " m = re.match(r\"POINT\\(([-0-9\\.]+)\\s+([-0-9\\.]+)\\)\", str(wkt))\n", " if not m:\n", " return np.nan, np.nan\n", " lon, lat = float(m.group(1)), float(m.group(2))\n", " return lon, lat\n", "\n", "# Tables\n", "streets = load_csv(\"geo_streets.csv\")\n", "parks = load_csv(\"geo_parks.csv\")\n", "poi = load_csv(\"poi_generic.csv\")\n", "households = load_csv(\"households.csv\")\n", "pets = load_csv(\"pets_registry.csv\")\n", "pet_inc = load_csv(\"pet_incidents.csv\")\n", "trips = load_csv(\"mobility_trips.csv\")\n", "safety = load_csv(\"public_safety.csv\")\n", "events = load_csv(\"events_civic.csv\")\n", "obs = load_csv(\"observations.csv\")\n", "lines = load_csv(\"transit_lines.csv\")\n", "stops = load_csv(\"transit_stops.csv\")\n", "ridership = load_csv(\"transit_ridership_daily.csv\")\n", "bike = load_csv(\"bike_infra.csv\")\n", "traffic = load_csv(\"traffic_counts.csv\")\n", "prices = load_csv(\"prices_index.csv\")\n", "weather = load_csv(\"weather_daily.csv\")\n", "trees = load_csv(\"trees_inventory.csv\")\n", "infra = load_csv(\"infrastructure_events.csv\")\n", "bissues = load_csv(\"building_issues.csv\")\n", "\n", "with open(os.path.join(DATA_DIR, \"DATA_DICTIONARY.json\"), \"r\", encoding=\"utf-8\") as f:\n", " dd = json.load(f)\n", "\n", "# Basic sanity\n", "print(\"Loaded tables:\")\n", "print(\" streets:\", streets.shape, \"| safety:\", safety.shape, \"| obs:\", obs.shape, \"| weather:\", weather.shape)" ] }, { "cell_type": "code", "execution_count": 5, "id": "2a457810", "metadata": {}, "outputs": [], "source": [ "def nearest_year_join(left_df, left_year_col, right_df, right_year_col, on_key_col):\n", " # For each (on_key_col, year) in left_df, find the right_df row with smallest |year diff|\n", " # This is a simple but effective approach for decimated time series.\n", " right_sorted = right_df.sort_values([on_key_col, right_year_col]).copy()\n", " out_rows = []\n", " for (k, y), block in left_df.groupby([on_key_col, left_year_col]):\n", " cand = right_sorted[right_sorted[on_key_col]==k]\n", " if cand.empty:\n", " # no match for this key\n", " tmp = block.copy()\n", " tmp[\"__ry\"] = np.nan\n", " tmp[\"__ridx\"] = -1\n", " out_rows.append(tmp)\n", " continue\n", " # pick nearest by year\n", " ridx = int((cand[right_year_col] - y).abs().idxmin())\n", " tmp = block.copy()\n", " tmp[\"__ry\"] = cand.loc[ridx, right_year_col]\n", " tmp[\"__ridx\"] = ridx\n", " out_rows.append(tmp)\n", " joined_left = pd.concat(out_rows, axis=0)\n", "\n", " # Attach right_df columns by __ridx where available\n", " joined = joined_left.merge(right_df.add_prefix(\"r__\"), left_on=\"__ridx\", right_index=True, how=\"left\")\n", " joined = joined.drop(columns=[\"__ridx\"])\n", " return joined" ] }, { "cell_type": "markdown", "id": "ff73a6ff", "metadata": {}, "source": [ "## 1) Classification: Predict high-severity public safety incidents\n", "**Target:** `severity` mapped to binary: *high* vs *not-high* \n", "**Features:** incident `category`, `reported_by`, street attributes (bike infra presence), daily weather (`temp_c`, `precip_mm`, `snow_cm`, `wind_kph`, `heatwave`)." ] }, { "cell_type": "code", "execution_count": 6, "id": "0a9bb0f2", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ " precision recall f1-score support\n", "\n", " 0 0.67 0.82 0.74 2026\n", " 1 0.30 0.16 0.21 974\n", "\n", " accuracy 0.61 3000\n", " macro avg 0.49 0.49 0.47 3000\n", "weighted avg 0.55 0.61 0.57 3000\n", "\n", "Confusion matrix:\n", " [[1671 355]\n", " [ 820 154]]\n", "ROC AUC: 0.4951799603106231\n", "Saved figure: ./ml_figs\\clf_roc_curve.png\n", "Saved model: ./ml_models\\incident_severity_rf.pkl\n", "Saved table: ./ml_tables\\clf_metrics_incident_severity.csv\n" ] } ], "source": [ "# Prepare safety data\n", "safety['date'] = pd.to_datetime(safety['date'])\n", "safety['year'] = safety['date'].dt.year\n", "safety['month'] = safety['date'].dt.month\n", "safety['day'] = safety['date'].dt.day\n", "\n", "# Target\n", "y = (safety['severity'] == 'high').astype(int)\n", "\n", "# Merge weather on date\n", "weather['date'] = pd.to_datetime(weather['date'])\n", "s_w = safety.merge(weather[['date','temp_c','precip_mm','snow_cm','wind_kph','heatwave']], on='date', how='left')\n", "\n", "# Street-level bike infra presence (binary, ever installed before safety year)\n", "bike_any = bike.copy()\n", "bike_any['has_bike_infra'] = True\n", "s_w = s_w.merge(\n", " bike_any[['street_id','has_bike_infra']].drop_duplicates('street_id'),\n", " left_on='location_street_id', right_on='street_id', how='left'\n", ")\n", "s_w['has_bike_infra'] = s_w['has_bike_infra'].fillna(False).infer_objects(copy=False)\n", "\n", "# Assemble features\n", "X = s_w[['category','reported_by','has_bike_infra','temp_c','precip_mm','snow_cm','wind_kph','heatwave','month']].copy()\n", "\n", "# Split numeric/categorical\n", "num_cols = ['temp_c','precip_mm','snow_cm','wind_kph','month']\n", "cat_cols = ['category','reported_by','heatwave','has_bike_infra']\n", "\n", "pre = ColumnTransformer([\n", " (\"num\", StandardScaler(), num_cols),\n", " (\"cat\", OneHotEncoder(handle_unknown=\"ignore\"), cat_cols)\n", "])\n", "\n", "clf = Pipeline([\n", " (\"pre\", pre),\n", " (\"rf\", RandomForestClassifier(n_estimators=200, random_state=42, class_weight=\"balanced\"))\n", "])\n", "\n", "X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=42, stratify=y)\n", "clf.fit(X_train, y_train)\n", "\n", "# Evaluate\n", "pred = clf.predict(X_test)\n", "proba = clf.predict_proba(X_test)[:,1]\n", "\n", "print(classification_report(y_test, pred))\n", "cm = confusion_matrix(y_test, pred)\n", "print(\"Confusion matrix:\\n\", cm)\n", "\n", "# ROC AUC & curve\n", "auc = roc_auc_score(y_test, proba)\n", "print(\"ROC AUC:\", auc)\n", "\n", "disp = RocCurveDisplay.from_predictions(y_test, proba)\n", "plt.title(\"ROC Curve — High Severity Incident Classifier\")\n", "savefig(\"clf_roc_curve.png\")\n", "\n", "# Save model\n", "savemodel(clf, \"incident_severity_rf.pkl\")\n", "\n", "# Save metrics as table\n", "metrics_tbl = pd.DataFrame({\n", " \"roc_auc\":[auc],\n", " \"accuracy\":[(pred==y_test).mean()],\n", " \"n_test\":[len(y_test)]\n", "})\n", "savetab(metrics_tbl, \"clf_metrics_incident_severity.csv\")" ] }, { "cell_type": "markdown", "id": "78159761", "metadata": {}, "source": [ "## 2) Regression: Predict street noise (dB)\n", "**Target:** `observations` where `metric == \"noise_db\"` → `value` (dB). \n", "**Features:** weather (temp, wind, rain, snow, heatwave), month, traffic volumes (nearest-year by street).\n" ] }, { "cell_type": "code", "execution_count": 7, "id": "bb03cb92", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "MAE: 5.521102874867018 R^2: 0.3084211030160505\n", "Saved figure: ./ml_figs\\regr_noise_pred_vs_actual.png\n", "Saved model: ./ml_models\\noise_rf.pkl\n", "Saved table: ./ml_tables\\regr_noise_metrics.csv\n" ] } ], "source": [ "# Filter observations to noise\n", "noise = obs[obs['metric']=='noise_db'].copy()\n", "noise['date'] = pd.to_datetime(noise['date'])\n", "noise['year'] = noise['date'].dt.year\n", "noise['month'] = noise['date'].dt.month\n", "\n", "# Merge weather\n", "weather['date'] = pd.to_datetime(weather['date'])\n", "nz = noise.merge(weather[['date','temp_c','precip_mm','snow_cm','wind_kph','heatwave']], on='date', how='left')\n", "\n", "# Traffic is every ~5 years; map to nearest year by street\n", "traffic['year'] = pd.to_datetime(traffic['date']).dt.year\n", "nz_key = nz[['location_street_id','year']].copy()\n", "nz_key = nz_key.rename(columns={'location_street_id':'street_id'})\n", "traffic_near = nearest_year_join(nz_key, 'year', traffic[['street_id','year','vehicles_per_day','bikes_per_day','peds_per_day']], 'year', 'street_id')\n", "\n", "# Merge traffic back by index alignment\n", "nz = nz.join(traffic_near[['r__street_id','r__year','r__vehicles_per_day','r__bikes_per_day','r__peds_per_day']].reset_index(drop=True))\n", "# If missing traffic (e.g., older years), fill with medians\n", "for c in ['r__vehicles_per_day','r__bikes_per_day','r__peds_per_day']:\n", " nz[c] = nz[c].fillna(nz[c].median())\n", "\n", "# Prepare features/target\n", "y = nz['value'].astype(float) # noise dB\n", "X = nz[['temp_c','precip_mm','snow_cm','wind_kph','heatwave','month','r__vehicles_per_day','r__bikes_per_day','r__peds_per_day']].copy()\n", "\n", "num_cols = ['temp_c','precip_mm','snow_cm','wind_kph','month','r__vehicles_per_day','r__bikes_per_day','r__peds_per_day']\n", "cat_cols = ['heatwave']\n", "\n", "pre = ColumnTransformer([\n", " (\"num\", StandardScaler(), num_cols),\n", " (\"cat\", OneHotEncoder(handle_unknown=\"ignore\"), cat_cols)\n", "])\n", "\n", "regr = Pipeline([\n", " (\"pre\", pre),\n", " (\"rf\", RandomForestRegressor(n_estimators=200, random_state=42))\n", "])\n", "\n", "X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=42)\n", "regr.fit(X_train, y_train)\n", "pred = regr.predict(X_test)\n", "\n", "mae = mean_absolute_error(y_test, pred)\n", "r2 = r2_score(y_test, pred)\n", "print(\"MAE:\", mae, \" R^2:\", r2)\n", "\n", "# Scatter: predicted vs actual\n", "plt.figure(figsize=(6,6))\n", "plt.scatter(y_test, pred, alpha=0.3)\n", "plt.xlabel(\"Actual dB\"); plt.ylabel(\"Predicted dB\")\n", "plt.title(\"Noise Regression — Predicted vs Actual\")\n", "savefig(\"regr_noise_pred_vs_actual.png\")\n", "\n", "# Save model & metrics\n", "savemodel(regr, \"noise_rf.pkl\")\n", "savetab(pd.DataFrame({\"MAE\":[mae], \"R2\":[r2], \"n_test\":[len(y_test)]}), \"regr_noise_metrics.csv\")" ] }, { "cell_type": "markdown", "id": "11a13962", "metadata": {}, "source": [ "## 3) Clustering: Streets by mobility signature\n", "We compute, for each street, a **mode share vector** based on trips that originate on that street. \n", "Then we cluster streets using **K-Means** and visualize cluster sizes." ] }, { "cell_type": "code", "execution_count": 8, "id": "8b77b4d6", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Saved figure: ./ml_figs\\cluster_elbow.png\n", "Saved table: ./ml_tables\\cluster_sizes.csv\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", "
clustern_streets
003
112
225
335
\n", "
" ], "text/plain": [ " cluster n_streets\n", "0 0 3\n", "1 1 2\n", "2 2 5\n", "3 3 5" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Saved figure: ./ml_figs\\cluster_sizes_bar.png\n", "Saved table: ./ml_tables\\cluster_assignments.csv\n" ] } ], "source": [ "# Build mode share per street (origin)\n", "mode_pvt = trips.pivot_table(index='origin_street_id', columns='mode', values='trip_id', aggfunc='count', fill_value=0)\n", "mode_share = mode_pvt.div(mode_pvt.sum(axis=1).replace(0,1), axis=0)\n", "\n", "# Try several k and pick a reasonable one (elbow heuristic)\n", "inertias = []\n", "ks = list(range(2, 9))\n", "for k in ks:\n", " km = KMeans(n_clusters=k, random_state=42, n_init=10)\n", " km.fit(mode_share)\n", " inertias.append(km.inertia_)\n", "\n", "plt.figure(figsize=(6,4))\n", "plt.plot(ks, inertias, marker=\"o\")\n", "plt.xlabel(\"k\"); plt.ylabel(\"Inertia\")\n", "plt.title(\"K-Means Elbow — Mobility Mode Share by Street\")\n", "savefig(\"cluster_elbow.png\")\n", "\n", "# Choose k=4 as a reasonable choice for demonstration\n", "k = 4\n", "km = KMeans(n_clusters=k, random_state=42, n_init=10)\n", "labels = km.fit_predict(mode_share)\n", "mode_share_clusters = mode_share.copy()\n", "mode_share_clusters['cluster'] = labels\n", "\n", "# Summarize clusters\n", "cluster_sizes = mode_share_clusters['cluster'].value_counts().sort_index().reset_index()\n", "cluster_sizes.columns = ['cluster','n_streets']\n", "savetab(cluster_sizes, \"cluster_sizes.csv\")\n", "display(cluster_sizes)\n", "\n", "plt.figure(figsize=(6,4))\n", "plt.bar(cluster_sizes['cluster'], cluster_sizes['n_streets'])\n", "plt.xlabel(\"Cluster\"); plt.ylabel(\"Number of streets\")\n", "plt.title(\"Street Clusters by Mobility Signature\")\n", "savefig(\"cluster_sizes_bar.png\")\n", "\n", "# Save cluster assignments\n", "assignments = mode_share_clusters.reset_index()[['origin_street_id','cluster']]\n", "savetab(assignments, \"cluster_assignments.csv\")" ] }, { "cell_type": "markdown", "id": "9dc1e2af", "metadata": {}, "source": [ "### Bonus: Interpreting models\n", "As a quick interpretability pass, we can look at feature importance from the random forests (MDI)." ] }, { "cell_type": "code", "execution_count": 9, "id": "73f43632", "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", "
featureimportance
0temp_c0.302091
1wind_kph0.219432
2month0.188248
3precip_mm0.082657
4snow_cm0.044882
5reported_by_sensor0.017501
6reported_by_officer0.016649
7reported_by_citizen0.016428
8category_animal0.012986
9category_theft0.012825
10category_disturbance0.012002
11category_noise0.011956
12category_other0.011694
13category_vandalism0.010977
14category_traffic0.010530
15has_bike_infra_True0.010157
16has_bike_infra_False0.009653
17heatwave_False0.004153
18heatwave_nan0.002672
19heatwave_True0.002508
\n", "
" ], "text/plain": [ " feature importance\n", "0 temp_c 0.302091\n", "1 wind_kph 0.219432\n", "2 month 0.188248\n", "3 precip_mm 0.082657\n", "4 snow_cm 0.044882\n", "5 reported_by_sensor 0.017501\n", "6 reported_by_officer 0.016649\n", "7 reported_by_citizen 0.016428\n", "8 category_animal 0.012986\n", "9 category_theft 0.012825\n", "10 category_disturbance 0.012002\n", "11 category_noise 0.011956\n", "12 category_other 0.011694\n", "13 category_vandalism 0.010977\n", "14 category_traffic 0.010530\n", "15 has_bike_infra_True 0.010157\n", "16 has_bike_infra_False 0.009653\n", "17 heatwave_False 0.004153\n", "18 heatwave_nan 0.002672\n", "19 heatwave_True 0.002508" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Saved table: ./ml_tables\\clf_top_feature_importances.csv\n", "Saved figure: ./ml_figs\\clf_feature_importances.png\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", "
featureimportance
0precip_mm0.231075
1temp_c0.223292
2snow_cm0.115154
3wind_kph0.098303
4month0.075172
5r__vehicles_per_day0.069528
6r__bikes_per_day0.068437
7r__peds_per_day0.066777
8heatwave_nan0.048992
9heatwave_False0.001682
10heatwave_True0.001590
\n", "
" ], "text/plain": [ " feature importance\n", "0 precip_mm 0.231075\n", "1 temp_c 0.223292\n", "2 snow_cm 0.115154\n", "3 wind_kph 0.098303\n", "4 month 0.075172\n", "5 r__vehicles_per_day 0.069528\n", "6 r__bikes_per_day 0.068437\n", "7 r__peds_per_day 0.066777\n", "8 heatwave_nan 0.048992\n", "9 heatwave_False 0.001682\n", "10 heatwave_True 0.001590" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Saved table: ./ml_tables\\regr_top_feature_importances.csv\n", "Saved figure: ./ml_figs\\regr_feature_importances.png\n" ] } ], "source": [ "# Extract MDI importances from pipelines (requires access to transformed feature names)\n", "def get_feature_names(preprocessor, num_cols, cat_cols, X):\n", " num_feats = num_cols\n", " # OneHot names\n", " oh = preprocessor.named_transformers_['cat']\n", " cat_names = list(oh.get_feature_names_out(cat_cols))\n", " return list(num_feats) + cat_names\n", "\n", "# Classification importances\n", "try:\n", " clf_model = clf.named_steps['rf']\n", " feat_names = get_feature_names(clf.named_steps['pre'], \n", " ['temp_c','precip_mm','snow_cm','wind_kph','month'],\n", " ['category','reported_by','heatwave','has_bike_infra'],\n", " X_train)\n", " imp = pd.Series(clf_model.feature_importances_, index=feat_names).sort_values(ascending=False)\n", " top_imp = imp.head(20).reset_index()\n", " top_imp.columns = ['feature','importance']\n", " display(top_imp)\n", " savetab(top_imp, \"clf_top_feature_importances.csv\")\n", "\n", " plt.figure(figsize=(8,5))\n", " plt.barh(top_imp['feature'][::-1], top_imp['importance'][::-1])\n", " plt.title(\"Top Feature Importances — Incident Severity RF\")\n", " plt.xlabel(\"Importance\")\n", " savefig(\"clf_feature_importances.png\")\n", "except Exception as e:\n", " print(\"Could not compute classification importances:\", e)\n", "\n", "# Regression importances\n", "try:\n", " regr_model = regr.named_steps['rf']\n", " feat_names = get_feature_names(regr.named_steps['pre'],\n", " ['temp_c','precip_mm','snow_cm','wind_kph','month','r__vehicles_per_day','r__bikes_per_day','r__peds_per_day'],\n", " ['heatwave'],\n", " X_train)\n", " imp = pd.Series(regr_model.feature_importances_, index=feat_names).sort_values(ascending=False)\n", " top_imp = imp.head(20).reset_index()\n", " top_imp.columns = ['feature','importance']\n", " display(top_imp)\n", " savetab(top_imp, \"regr_top_feature_importances.csv\")\n", "\n", " plt.figure(figsize=(8,5))\n", " plt.barh(top_imp['feature'][::-1], top_imp['importance'][::-1])\n", " plt.title(\"Top Feature Importances — Noise RF\")\n", " plt.xlabel(\"Importance\")\n", " savefig(\"regr_feature_importances.png\")\n", "except Exception as e:\n", " print(\"Could not compute regression importances:\", e)" ] }, { "cell_type": "markdown", "id": "0963228d", "metadata": {}, "source": [ "### Next ideas\n", "- Use **time-based splits** (train on early years, test on later) for more realistic forecasting.\n", "- Try **Gradient Boosting** or **XGBoost/LightGBM** for stronger baselines.\n", "- Add **spatial features** (distance to parks, density of POIs) for richer models.\n", "- Package results into an **HTML report** or a small **Streamlit** app." ] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.12.7" } }, "nbformat": 4, "nbformat_minor": 5 }