{ "cells": [ { "cell_type": "markdown", "id": "f43dd523", "metadata": {}, "source": [ "# Car recommendation notebook (initial version)\n", "\n", "## Task framing\n", "\n", "This dataset does not contain rich repeated user histories like MovieLens, so the car task is framed as a **context-aware top-N recommendation** problem:\n", "\n", "- **Input/query**: a desired car profile and customer context \n", " (`Year`, `Price`, `Mileage`, `Color`, `Condition`, `Country`)\n", "- **Target item**: car **Brand + Model**\n", "- **Output**: top-N recommended car models for that query\n", "\n", "This is the closest honest recommender formulation for the current schema.\n", "\n", "## Model set in this initial version\n", "\n", "### Regular models\n", "1. Popularity by country\n", "2. Context KNN\n", "3. Linear softmax classifier\n", "\n", "### Neural models\n", "4. MLP classifier\n", "5. Two-tower retrieval model\n", "\n", "## Metrics\n", "The models are evaluated with:\n", "- `HitRate@5`, `HitRate@10`, `HitRate@20`\n", "- `MRR@10`\n", "- `NDCG@10`" ] }, { "cell_type": "code", "execution_count": 1, "id": "2188f21b", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Torch device: cuda\n" ] } ], "source": [ "from pathlib import Path\n", "import math\n", "import random\n", "import warnings\n", "warnings.filterwarnings(\"ignore\")\n", "\n", "import numpy as np\n", "import pandas as pd\n", "\n", "from sklearn.model_selection import train_test_split\n", "from sklearn.preprocessing import OneHotEncoder, StandardScaler, LabelEncoder\n", "from sklearn.compose import ColumnTransformer\n", "from sklearn.pipeline import Pipeline\n", "from sklearn.impute import SimpleImputer\n", "from sklearn.neighbors import NearestNeighbors\n", "from sklearn.linear_model import SGDClassifier\n", "\n", "from tqdm.auto import tqdm\n", "\n", "import torch\n", "import torch.nn as nn\n", "from torch.utils.data import Dataset, DataLoader\n", "\n", "RANDOM_STATE = 42\n", "np.random.seed(RANDOM_STATE)\n", "random.seed(RANDOM_STATE)\n", "torch.manual_seed(RANDOM_STATE)\n", "\n", "DEVICE = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n", "print(\"Torch device:\", DEVICE)" ] }, { "cell_type": "code", "execution_count": 2, "id": "3883bf1a", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Using dataset: /home/konnilol/Documents/uni/kursovaya-sem5/car_sales_dataset_with_person_details.csv\n" ] } ], "source": [ "# Paths and core settings\n", "\n", "BASE_DIR = Path.cwd()\n", "\n", "DATA_PATH = BASE_DIR / \"car_sales_dataset_with_person_details.csv\"\n", "\n", "# Optional absolute fallback for your current local setup\n", "if not DATA_PATH.exists():\n", " fallback = Path(\"/home/konnilol/Documents/uni/kursovaya-sem5/car_sales_dataset_with_person_details.csv\")\n", " if fallback.exists():\n", " DATA_PATH = fallback\n", "\n", "if not DATA_PATH.exists():\n", " raise FileNotFoundError(f\"CSV file not found: {DATA_PATH}\")\n", "\n", "TARGET_TOP_N = 20\n", "\n", "# Keep only items with enough observations so top-N evaluation is meaningful\n", "MIN_ITEM_SUPPORT = 20\n", "\n", "# Limit evaluation set for faster iteration\n", "MAX_EVAL_ROWS = 5000\n", "\n", "# KNN settings\n", "KNN_NEIGHBORS = 100\n", "\n", "# Neural settings\n", "BATCH_SIZE = 512\n", "MLP_EPOCHS = 8\n", "TWOTOWER_EPOCHS = 8\n", "LEARNING_RATE = 1e-3\n", "EMBED_DIM = 32\n", "\n", "print(\"Using dataset:\", DATA_PATH.resolve())" ] }, { "cell_type": "code", "execution_count": 3, "id": "c41c1b7e", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Shape: (1000000, 11)\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", "
BrandModelYearPriceMileageColorConditionFirst NameLast NameAddressCountry
0HondaCivic202325627.2058513GreenCertified Pre-OwnedEmilyHarris456 Oak AveBrazil
1MazdaMazda3200012027.1460990BrownCertified Pre-OwnedJohnHarris101 Maple DrItaly
2MazdaCX-5201449194.931703GreenCertified Pre-OwnedKarenWilson202 Birch BlvdUK
3HyundaiTucson200311955.9425353SilverUsedSusanMartinez123 Main StMexico
4Land RoverRange Rover201210910.0176854OrangeUsedCharlesMiller456 Oak AveUSA
\n", "
" ], "text/plain": [ " Brand Model Year Price Mileage Color \\\n", "0 Honda Civic 2023 25627.20 58513 Green \n", "1 Mazda Mazda3 2000 12027.14 60990 Brown \n", "2 Mazda CX-5 2014 49194.93 1703 Green \n", "3 Hyundai Tucson 2003 11955.94 25353 Silver \n", "4 Land Rover Range Rover 2012 10910.01 76854 Orange \n", "\n", " Condition First Name Last Name Address Country \n", "0 Certified Pre-Owned Emily Harris 456 Oak Ave Brazil \n", "1 Certified Pre-Owned John Harris 101 Maple Dr Italy \n", "2 Certified Pre-Owned Karen Wilson 202 Birch Blvd UK \n", "3 Used Susan Martinez 123 Main St Mexico \n", "4 Used Charles Miller 456 Oak Ave USA " ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Brand object\n", "Model object\n", "Year int64\n", "Price float64\n", "Mileage int64\n", "Color object\n", "Condition object\n", "First Name object\n", "Last Name object\n", "Address object\n", "Country object\n", "dtype: object\n" ] } ], "source": [ "# Load data\n", "\n", "df = pd.read_csv(DATA_PATH)\n", "\n", "print(\"Shape:\", df.shape)\n", "display(df.head())\n", "print(df.dtypes)" ] }, { "cell_type": "code", "execution_count": 4, "id": "f8209587", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Rows after cleanup: 1000000\n", "Unique items before filtering: 88\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", "
BrandModelYearPriceMileageColorConditionCountryitem_id
0HondaCivic202325627.2058513GreenCertified Pre-OwnedBrazilHonda :: Civic
1MazdaMazda3200012027.1460990BrownCertified Pre-OwnedItalyMazda :: Mazda3
2MazdaCX-5201449194.931703GreenCertified Pre-OwnedUKMazda :: CX-5
3HyundaiTucson200311955.9425353SilverUsedMexicoHyundai :: Tucson
4Land RoverRange Rover201210910.0176854OrangeUsedUSALand Rover :: Range Rover
\n", "
" ], "text/plain": [ " Brand Model Year Price Mileage Color \\\n", "0 Honda Civic 2023 25627.20 58513 Green \n", "1 Mazda Mazda3 2000 12027.14 60990 Brown \n", "2 Mazda CX-5 2014 49194.93 1703 Green \n", "3 Hyundai Tucson 2003 11955.94 25353 Silver \n", "4 Land Rover Range Rover 2012 10910.01 76854 Orange \n", "\n", " Condition Country item_id \n", "0 Certified Pre-Owned Brazil Honda :: Civic \n", "1 Certified Pre-Owned Italy Mazda :: Mazda3 \n", "2 Certified Pre-Owned UK Mazda :: CX-5 \n", "3 Used Mexico Hyundai :: Tucson \n", "4 Used USA Land Rover :: Range Rover " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Define recommendation target and query features\n", "\n", "required_cols = [\n", " \"Brand\", \"Model\", \"Year\", \"Price\", \"Mileage\",\n", " \"Color\", \"Condition\", \"Country\"\n", "]\n", "\n", "missing = [c for c in required_cols if c not in df.columns]\n", "if missing:\n", " raise ValueError(f\"Missing required columns: {missing}\")\n", "\n", "df = df[required_cols].copy()\n", "df = df.dropna().reset_index(drop=True)\n", "\n", "# Item definition: Brand + Model\n", "df[\"item_id\"] = df[\"Brand\"].astype(str).str.strip() + \" :: \" + df[\"Model\"].astype(str).str.strip()\n", "\n", "# Query features used by recommenders\n", "num_cols = [\"Year\", \"Price\", \"Mileage\"]\n", "cat_cols = [\"Color\", \"Condition\", \"Country\"]\n", "feature_cols = num_cols + cat_cols\n", "target_col = \"item_id\"\n", "\n", "print(\"Rows after cleanup:\", len(df))\n", "print(\"Unique items before filtering:\", df[target_col].nunique())\n", "display(df.head())" ] }, { "cell_type": "code", "execution_count": 5, "id": "0a48f3ea", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Rows after item filtering: 1000000\n", "Unique items after filtering: 88\n", "Top items:\n" ] }, { "data": { "text/plain": [ "item_id\n", "Chrysler :: Voyager 18572\n", "Chrysler :: 300 18497\n", "Chrysler :: Pacifica 18297\n", "Chevrolet :: Silverado 11391\n", "Mercedes-Benz :: E-Class 11301\n", "Hyundai :: Sonata 11298\n", "Kia :: Optima 11255\n", "Mercedes-Benz :: GLC 11249\n", "Jeep :: Cherokee 11245\n", "Toyota :: Tacoma 11241\n", "Name: count, dtype: int64" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Train shape: (800000, 9)\n", "Test shape : (5000, 9)\n" ] } ], "source": [ "# Filter rare items and create train/test split\n", "\n", "item_counts = df[target_col].value_counts()\n", "kept_items = item_counts[item_counts >= MIN_ITEM_SUPPORT].index\n", "\n", "df_model = df[df[target_col].isin(kept_items)].copy().reset_index(drop=True)\n", "\n", "print(\"Rows after item filtering:\", len(df_model))\n", "print(\"Unique items after filtering:\", df_model[target_col].nunique())\n", "print(\"Top items:\")\n", "display(df_model[target_col].value_counts().head(10))\n", "\n", "if df_model[target_col].nunique() < 5:\n", " raise ValueError(\n", " \"Too few supported items after filtering. Lower MIN_ITEM_SUPPORT and rerun.\"\n", " )\n", "\n", "train_df, test_df = train_test_split(\n", " df_model,\n", " test_size=0.2,\n", " random_state=RANDOM_STATE,\n", " stratify=df_model[target_col]\n", ")\n", "\n", "train_df = train_df.reset_index(drop=True)\n", "test_df = test_df.reset_index(drop=True)\n", "\n", "if MAX_EVAL_ROWS is not None and len(test_df) > MAX_EVAL_ROWS:\n", " test_df = test_df.sample(MAX_EVAL_ROWS, random_state=RANDOM_STATE).reset_index(drop=True)\n", "\n", "print(\"Train shape:\", train_df.shape)\n", "print(\"Test shape :\", test_df.shape)" ] }, { "cell_type": "code", "execution_count": 6, "id": "63c8b73f", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Number of items: 88\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", "
BrandModelavg_yearavg_priceavg_mileagemain_colormain_condition
item_id
Audi :: A3AudiA32011.97083142423.44053698860.997852GrayUsed
Audi :: A4AudiA42011.91612442401.07580399596.451157BlackUsed
Audi :: A6AudiA62012.03386842415.86548399910.641023GrayUsed
Audi :: Q5AudiQ52011.93685942487.54679899363.572943OrangeNew
Audi :: Q7AudiQ72012.03488642450.890595100324.285568RedNew
\n", "
" ], "text/plain": [ " Brand Model avg_year avg_price avg_mileage main_color \\\n", "item_id \n", "Audi :: A3 Audi A3 2011.970831 42423.440536 98860.997852 Gray \n", "Audi :: A4 Audi A4 2011.916124 42401.075803 99596.451157 Black \n", "Audi :: A6 Audi A6 2012.033868 42415.865483 99910.641023 Gray \n", "Audi :: Q5 Audi Q5 2011.936859 42487.546798 99363.572943 Orange \n", "Audi :: Q7 Audi Q7 2012.034886 42450.890595 100324.285568 Red \n", "\n", " main_condition \n", "item_id \n", "Audi :: A3 Used \n", "Audi :: A4 Used \n", "Audi :: A6 Used \n", "Audi :: Q5 New \n", "Audi :: Q7 New " ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Build item vocabulary and helper tables\n", "\n", "label_encoder = LabelEncoder()\n", "label_encoder.fit(train_df[target_col])\n", "\n", "train_df = train_df[train_df[target_col].isin(label_encoder.classes_)].copy()\n", "test_df = test_df[test_df[target_col].isin(label_encoder.classes_)].copy()\n", "\n", "train_df[\"item_idx\"] = label_encoder.transform(train_df[target_col])\n", "test_df[\"item_idx\"] = label_encoder.transform(test_df[target_col])\n", "\n", "all_items = label_encoder.classes_.tolist()\n", "num_items = len(all_items)\n", "\n", "item_popularity = train_df[target_col].value_counts()\n", "global_popularity = item_popularity.index.tolist()\n", "\n", "country_item_popularity = (\n", " train_df.groupby([\"Country\", target_col])\n", " .size()\n", " .reset_index(name=\"cnt\")\n", " .sort_values([\"Country\", \"cnt\"], ascending=[True, False])\n", ")\n", "\n", "country_top_items = {\n", " country: grp[target_col].tolist()\n", " for country, grp in country_item_popularity.groupby(\"Country\")\n", "}\n", "\n", "item_meta = (\n", " train_df.groupby(target_col)\n", " .agg(\n", " Brand=(\"Brand\", \"first\"),\n", " Model=(\"Model\", \"first\"),\n", " avg_year=(\"Year\", \"mean\"),\n", " avg_price=(\"Price\", \"mean\"),\n", " avg_mileage=(\"Mileage\", \"mean\"),\n", " main_color=(\"Color\", lambda s: s.mode().iloc[0] if not s.mode().empty else s.iloc[0]),\n", " main_condition=(\"Condition\", lambda s: s.mode().iloc[0] if not s.mode().empty else s.iloc[0]),\n", " )\n", ")\n", "\n", "print(\"Number of items:\", num_items)\n", "display(item_meta.head())" ] }, { "cell_type": "code", "execution_count": 7, "id": "08042e2f", "metadata": {}, "outputs": [], "source": [ "# Metrics\n", "\n", "def hit_rate_at_k(true_item, ranked_items, k):\n", " return 1.0 if true_item in ranked_items[:k] else 0.0\n", "\n", "def mrr_at_k(true_item, ranked_items, k):\n", " topk = ranked_items[:k]\n", " if true_item in topk:\n", " rank = topk.index(true_item) + 1\n", " return 1.0 / rank\n", " return 0.0\n", "\n", "def ndcg_at_k(true_item, ranked_items, k):\n", " topk = ranked_items[:k]\n", " if true_item in topk:\n", " rank = topk.index(true_item) + 1\n", " return 1.0 / math.log2(rank + 1)\n", " return 0.0\n", "\n", "def evaluate_model(test_frame, recommend_fn, ks=(5, 10, 20), show_progress=True):\n", " rows = []\n", "\n", " iterator = tqdm(test_frame.itertuples(index=False), total=len(test_frame)) if show_progress else test_frame.itertuples(index=False)\n", "\n", " for row in iterator:\n", " ranked = recommend_fn(row, max(ks))\n", " true_item = getattr(row, target_col)\n", "\n", " result = {\"true_item\": true_item}\n", " for k in ks:\n", " result[f\"HR@{k}\"] = hit_rate_at_k(true_item, ranked, k)\n", " result[\"MRR@10\"] = mrr_at_k(true_item, ranked, 10)\n", " result[\"NDCG@10\"] = ndcg_at_k(true_item, ranked, 10)\n", " rows.append(result)\n", "\n", " detail_df = pd.DataFrame(rows)\n", " summary = detail_df.drop(columns=[\"true_item\"]).mean().to_dict()\n", " return summary, detail_df" ] }, { "cell_type": "code", "execution_count": 8, "id": "3b7574aa", "metadata": {}, "outputs": [], "source": [ "# Regular model 1: popularity by country\n", "\n", "def dedupe_fill(primary, fallback, n):\n", " out = []\n", " seen = set()\n", " for item in primary + fallback:\n", " if item not in seen:\n", " out.append(item)\n", " seen.add(item)\n", " if len(out) >= n:\n", " break\n", " return out\n", "\n", "def recommend_popularity(row, n=10):\n", " country = getattr(row, \"Country\")\n", " country_items = country_top_items.get(country, [])\n", " return dedupe_fill(country_items, global_popularity, n)" ] }, { "cell_type": "code", "execution_count": 9, "id": "69d53903", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "KNN train matrix shape: (800000, 26)\n" ] } ], "source": [ "# Regular model 2: context KNN on query features\n", "\n", "preprocessor = ColumnTransformer(\n", " transformers=[\n", " (\"num\", Pipeline([\n", " (\"imputer\", SimpleImputer(strategy=\"median\")),\n", " (\"scaler\", StandardScaler())\n", " ]), num_cols),\n", " (\"cat\", Pipeline([\n", " (\"imputer\", SimpleImputer(strategy=\"most_frequent\")),\n", " (\"onehot\", OneHotEncoder(handle_unknown=\"ignore\"))\n", " ]), cat_cols),\n", " ]\n", ")\n", "\n", "X_train_ctx = preprocessor.fit_transform(train_df[feature_cols])\n", "\n", "knn = NearestNeighbors(\n", " n_neighbors=min(KNN_NEIGHBORS, len(train_df)),\n", " metric=\"cosine\",\n", " algorithm=\"brute\",\n", " n_jobs=-1\n", ")\n", "knn.fit(X_train_ctx)\n", "\n", "train_items_array = train_df[target_col].to_numpy()\n", "train_countries_array = train_df[\"Country\"].to_numpy()\n", "\n", "print(\"KNN train matrix shape:\", X_train_ctx.shape)\n", "\n", "def recommend_knn(row, n=10):\n", " query_df = pd.DataFrame([{c: getattr(row, c) for c in feature_cols}])\n", " q = preprocessor.transform(query_df)\n", "\n", " distances, indices = knn.kneighbors(q, n_neighbors=min(KNN_NEIGHBORS, len(train_df)))\n", " distances = distances[0]\n", " indices = indices[0]\n", "\n", " scores = {}\n", " for dist, idx in zip(distances, indices):\n", " item = train_items_array[idx]\n", " country_bonus = 0.15 if train_countries_array[idx] == getattr(row, \"Country\") else 0.0\n", " weight = (1.0 / (dist + 1e-6)) + country_bonus\n", " scores[item] = scores.get(item, 0.0) + weight\n", "\n", " ranked = [item for item, _ in sorted(scores.items(), key=lambda x: x[1], reverse=True)]\n", " fallback = country_top_items.get(getattr(row, \"Country\"), [])\n", " return dedupe_fill(ranked, fallback + global_popularity, n)" ] }, { "cell_type": "code", "execution_count": 10, "id": "026f2236", "metadata": {}, "outputs": [], "source": [ "# Regular model 3: linear softmax classifier\n", "\n", "linear_model = SGDClassifier(\n", " loss=\"log_loss\",\n", " penalty=\"l2\",\n", " alpha=1e-5,\n", " max_iter=50,\n", " tol=1e-3,\n", " random_state=RANDOM_STATE\n", ")\n", "\n", "linear_model.fit(X_train_ctx, train_df[\"item_idx\"])\n", "\n", "def recommend_linear(row, n=10):\n", " query_df = pd.DataFrame([{c: getattr(row, c) for c in feature_cols}])\n", " q = preprocessor.transform(query_df)\n", "\n", " probs = linear_model.predict_proba(q)[0]\n", " top_idx = np.argsort(probs)[::-1][:n]\n", " return label_encoder.inverse_transform(top_idx).tolist()" ] }, { "cell_type": "code", "execution_count": 11, "id": "e25ed13c", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Neural train size: 800000\n", "Neural test size : 5000\n", "Items: 88\n" ] } ], "source": [ "# Tabular encoders for neural models\n", "\n", "def build_vocab(series):\n", " values = [\"[UNK]\"] + sorted(series.astype(str).unique().tolist())\n", " return {v: i for i, v in enumerate(values)}\n", "\n", "color_vocab = build_vocab(train_df[\"Color\"])\n", "condition_vocab = build_vocab(train_df[\"Condition\"])\n", "country_vocab = build_vocab(train_df[\"Country\"])\n", "\n", "num_means = train_df[num_cols].mean()\n", "num_stds = train_df[num_cols].std().replace(0, 1.0)\n", "\n", "def encode_cat(value, vocab):\n", " return vocab.get(str(value), 0)\n", "\n", "def encode_numeric_frame(frame):\n", " arr = frame[num_cols].copy()\n", " arr = (arr - num_means) / num_stds\n", " return arr.astype(np.float32).to_numpy()\n", "\n", "def encode_categorical_frame(frame):\n", " color = frame[\"Color\"].astype(str).map(lambda x: encode_cat(x, color_vocab)).astype(np.int64).to_numpy()\n", " condition = frame[\"Condition\"].astype(str).map(lambda x: encode_cat(x, condition_vocab)).astype(np.int64).to_numpy()\n", " country = frame[\"Country\"].astype(str).map(lambda x: encode_cat(x, country_vocab)).astype(np.int64).to_numpy()\n", " return color, condition, country\n", "\n", "train_num = encode_numeric_frame(train_df)\n", "test_num = encode_numeric_frame(test_df)\n", "\n", "train_color, train_condition, train_country = encode_categorical_frame(train_df)\n", "test_color, test_condition, test_country = encode_categorical_frame(test_df)\n", "\n", "y_train = train_df[\"item_idx\"].astype(np.int64).to_numpy()\n", "y_test = test_df[\"item_idx\"].astype(np.int64).to_numpy()\n", "\n", "print(\"Neural train size:\", len(train_df))\n", "print(\"Neural test size :\", len(test_df))\n", "print(\"Items:\", num_items)" ] }, { "cell_type": "code", "execution_count": 12, "id": "96f23867", "metadata": {}, "outputs": [], "source": [ "# Dataset classes\n", "\n", "class CarQueryDataset(Dataset):\n", " def __init__(self, num_x, color_x, condition_x, country_x, y):\n", " self.num_x = torch.tensor(num_x, dtype=torch.float32)\n", " self.color_x = torch.tensor(color_x, dtype=torch.long)\n", " self.condition_x = torch.tensor(condition_x, dtype=torch.long)\n", " self.country_x = torch.tensor(country_x, dtype=torch.long)\n", " self.y = torch.tensor(y, dtype=torch.long)\n", "\n", " def __len__(self):\n", " return len(self.y)\n", "\n", " def __getitem__(self, idx):\n", " return (\n", " self.num_x[idx],\n", " self.color_x[idx],\n", " self.condition_x[idx],\n", " self.country_x[idx],\n", " self.y[idx]\n", " )\n", "\n", "train_dataset = CarQueryDataset(train_num, train_color, train_condition, train_country, y_train)\n", "test_dataset = CarQueryDataset(test_num, test_color, test_condition, test_country, y_test)\n", "\n", "train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True)" ] }, { "cell_type": "code", "execution_count": 13, "id": "79a7b0bf", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "MLP epoch 1/8 - loss: 4.4729\n", "MLP epoch 2/8 - loss: 4.4719\n", "MLP epoch 3/8 - loss: 4.4717\n", "MLP epoch 4/8 - loss: 4.4716\n", "MLP epoch 5/8 - loss: 4.4714\n", "MLP epoch 6/8 - loss: 4.4712\n", "MLP epoch 7/8 - loss: 4.4710\n", "MLP epoch 8/8 - loss: 4.4708\n" ] } ], "source": [ "# Neural model 1: MLP classifier\n", "\n", "class QueryEncoder(nn.Module):\n", " def __init__(self, n_colors, n_conditions, n_countries, embed_dim=16, num_dim=3, hidden_dim=128, out_dim=128):\n", " super().__init__()\n", " self.color_emb = nn.Embedding(n_colors, embed_dim)\n", " self.condition_emb = nn.Embedding(n_conditions, embed_dim)\n", " self.country_emb = nn.Embedding(n_countries, embed_dim)\n", "\n", " input_dim = num_dim + embed_dim * 3\n", " self.net = nn.Sequential(\n", " nn.Linear(input_dim, hidden_dim),\n", " nn.ReLU(),\n", " nn.Dropout(0.2),\n", " nn.Linear(hidden_dim, out_dim),\n", " nn.ReLU(),\n", " )\n", "\n", " def forward(self, num_x, color_x, condition_x, country_x):\n", " x = torch.cat([\n", " num_x,\n", " self.color_emb(color_x),\n", " self.condition_emb(condition_x),\n", " self.country_emb(country_x)\n", " ], dim=1)\n", " return self.net(x)\n", "\n", "class MLPClassifierRecommender(nn.Module):\n", " def __init__(self, n_colors, n_conditions, n_countries, n_items, embed_dim=16, hidden_dim=128):\n", " super().__init__()\n", " self.encoder = QueryEncoder(\n", " n_colors=n_colors,\n", " n_conditions=n_conditions,\n", " n_countries=n_countries,\n", " embed_dim=embed_dim,\n", " hidden_dim=hidden_dim,\n", " out_dim=hidden_dim\n", " )\n", " self.head = nn.Linear(hidden_dim, n_items)\n", "\n", " def forward(self, num_x, color_x, condition_x, country_x):\n", " z = self.encoder(num_x, color_x, condition_x, country_x)\n", " return self.head(z)\n", "\n", "mlp_model = MLPClassifierRecommender(\n", " n_colors=len(color_vocab),\n", " n_conditions=len(condition_vocab),\n", " n_countries=len(country_vocab),\n", " n_items=num_items,\n", " embed_dim=EMBED_DIM,\n", " hidden_dim=128\n", ").to(DEVICE)\n", "\n", "criterion = nn.CrossEntropyLoss()\n", "optimizer = torch.optim.Adam(mlp_model.parameters(), lr=LEARNING_RATE)\n", "\n", "for epoch in range(MLP_EPOCHS):\n", " mlp_model.train()\n", " total_loss = 0.0\n", "\n", " for num_x, color_x, condition_x, country_x, y in train_loader:\n", " num_x = num_x.to(DEVICE)\n", " color_x = color_x.to(DEVICE)\n", " condition_x = condition_x.to(DEVICE)\n", " country_x = country_x.to(DEVICE)\n", " y = y.to(DEVICE)\n", "\n", " optimizer.zero_grad()\n", " logits = mlp_model(num_x, color_x, condition_x, country_x)\n", " loss = criterion(logits, y)\n", " loss.backward()\n", " optimizer.step()\n", "\n", " total_loss += loss.item() * len(y)\n", "\n", " epoch_loss = total_loss / len(train_loader.dataset)\n", " print(f\"MLP epoch {epoch+1}/{MLP_EPOCHS} - loss: {epoch_loss:.4f}\")\n", "\n", "@torch.no_grad()\n", "def recommend_mlp(row, n=10):\n", " mlp_model.eval()\n", "\n", " num_x = torch.tensor([[\n", " (float(getattr(row, \"Year\")) - num_means[\"Year\"]) / num_stds[\"Year\"],\n", " (float(getattr(row, \"Price\")) - num_means[\"Price\"]) / num_stds[\"Price\"],\n", " (float(getattr(row, \"Mileage\")) - num_means[\"Mileage\"]) / num_stds[\"Mileage\"],\n", " ]], dtype=torch.float32, device=DEVICE)\n", "\n", " color_x = torch.tensor([encode_cat(getattr(row, \"Color\"), color_vocab)], dtype=torch.long, device=DEVICE)\n", " condition_x = torch.tensor([encode_cat(getattr(row, \"Condition\"), condition_vocab)], dtype=torch.long, device=DEVICE)\n", " country_x = torch.tensor([encode_cat(getattr(row, \"Country\"), country_vocab)], dtype=torch.long, device=DEVICE)\n", "\n", " logits = mlp_model(num_x, color_x, condition_x, country_x)[0]\n", " top_idx = torch.topk(logits, k=min(n, num_items)).indices.cpu().numpy()\n", " return label_encoder.inverse_transform(top_idx).tolist()" ] }, { "cell_type": "code", "execution_count": 14, "id": "ea7c18a1", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "TwoTower epoch 1/8 - loss: 4.4788\n", "TwoTower epoch 2/8 - loss: 4.4774\n", "TwoTower epoch 3/8 - loss: 4.4773\n", "TwoTower epoch 4/8 - loss: 4.4773\n", "TwoTower epoch 5/8 - loss: 4.4773\n", "TwoTower epoch 6/8 - loss: 4.4773\n", "TwoTower epoch 7/8 - loss: 4.4773\n", "TwoTower epoch 8/8 - loss: 4.4773\n" ] } ], "source": [ "# Neural model 2: two-tower retrieval\n", "\n", "class TwoTowerRecommender(nn.Module):\n", " def __init__(self, n_colors, n_conditions, n_countries, n_items, embed_dim=32, hidden_dim=128):\n", " super().__init__()\n", " self.query_encoder = QueryEncoder(\n", " n_colors=n_colors,\n", " n_conditions=n_conditions,\n", " n_countries=n_countries,\n", " embed_dim=embed_dim,\n", " hidden_dim=hidden_dim,\n", " out_dim=embed_dim\n", " )\n", " self.item_embedding = nn.Embedding(n_items, embed_dim)\n", "\n", " def forward(self, num_x, color_x, condition_x, country_x):\n", " q = self.query_encoder(num_x, color_x, condition_x, country_x)\n", " item_emb = self.item_embedding.weight\n", " scores = q @ item_emb.T\n", " return scores\n", "\n", "twotower_model = TwoTowerRecommender(\n", " n_colors=len(color_vocab),\n", " n_conditions=len(condition_vocab),\n", " n_countries=len(country_vocab),\n", " n_items=num_items,\n", " embed_dim=EMBED_DIM,\n", " hidden_dim=128\n", ").to(DEVICE)\n", "\n", "criterion_tt = nn.CrossEntropyLoss()\n", "optimizer_tt = torch.optim.Adam(twotower_model.parameters(), lr=LEARNING_RATE)\n", "\n", "for epoch in range(TWOTOWER_EPOCHS):\n", " twotower_model.train()\n", " total_loss = 0.0\n", "\n", " for num_x, color_x, condition_x, country_x, y in train_loader:\n", " num_x = num_x.to(DEVICE)\n", " color_x = color_x.to(DEVICE)\n", " condition_x = condition_x.to(DEVICE)\n", " country_x = country_x.to(DEVICE)\n", " y = y.to(DEVICE)\n", "\n", " optimizer_tt.zero_grad()\n", " scores = twotower_model(num_x, color_x, condition_x, country_x)\n", " loss = criterion_tt(scores, y)\n", " loss.backward()\n", " optimizer_tt.step()\n", "\n", " total_loss += loss.item() * len(y)\n", "\n", " epoch_loss = total_loss / len(train_loader.dataset)\n", " print(f\"TwoTower epoch {epoch+1}/{TWOTOWER_EPOCHS} - loss: {epoch_loss:.4f}\")\n", "\n", "@torch.no_grad()\n", "def recommend_twotower(row, n=10):\n", " twotower_model.eval()\n", "\n", " num_x = torch.tensor([[\n", " (float(getattr(row, \"Year\")) - num_means[\"Year\"]) / num_stds[\"Year\"],\n", " (float(getattr(row, \"Price\")) - num_means[\"Price\"]) / num_stds[\"Price\"],\n", " (float(getattr(row, \"Mileage\")) - num_means[\"Mileage\"]) / num_stds[\"Mileage\"],\n", " ]], dtype=torch.float32, device=DEVICE)\n", "\n", " color_x = torch.tensor([encode_cat(getattr(row, \"Color\"), color_vocab)], dtype=torch.long, device=DEVICE)\n", " condition_x = torch.tensor([encode_cat(getattr(row, \"Condition\"), condition_vocab)], dtype=torch.long, device=DEVICE)\n", " country_x = torch.tensor([encode_cat(getattr(row, \"Country\"), country_vocab)], dtype=torch.long, device=DEVICE)\n", "\n", " scores = twotower_model(num_x, color_x, condition_x, country_x)[0]\n", " top_idx = torch.topk(scores, k=min(n, num_items)).indices.cpu().numpy()\n", " return label_encoder.inverse_transform(top_idx).tolist()" ] }, { "cell_type": "code", "execution_count": 15, "id": "fece0740", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Example query:\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", "
YearPriceMileageColorConditionCountryitem_id
150120075180.4984626BlackCertified Pre-OwnedFranceBMW :: Z4
\n", "
" ], "text/plain": [ " Year Price Mileage Color Condition Country item_id\n", "1501 2007 5180.49 84626 Black Certified Pre-Owned France BMW :: Z4" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "Popularity:\n", "['Chrysler :: Pacifica', 'Chrysler :: 300', 'Chrysler :: Voyager', 'Lexus :: NX', 'Kia :: Sportage', 'Volkswagen :: Atlas', 'Chevrolet :: Trax', 'Ford :: F-150', 'Hyundai :: Santa Fe', 'Honda :: Civic']\n", "KNN:\n", "['Kia :: Seltos', 'BMW :: 3 Series', 'Chevrolet :: Trax', 'Audi :: Q7', 'Subaru :: Legacy', 'Audi :: Q5', 'Ford :: Mustang', 'Toyota :: Tacoma', 'Ford :: Fusion', 'Hyundai :: Kona']\n", "Linear:\n", "['Chrysler :: Pacifica', 'Chrysler :: 300', 'Dodge :: Journey', 'Dodge :: Ram 1500', 'Subaru :: Forester', 'Land Rover :: Velar', 'Lexus :: IS', 'Chevrolet :: Equinox', 'Hyundai :: Sonata', 'Mercedes-Benz :: E-Class']\n", "MLP:\n", "['Chrysler :: 300', 'Chrysler :: Pacifica', 'Chrysler :: Voyager', 'Toyota :: Tacoma', 'Hyundai :: Santa Fe', 'Volkswagen :: Tiguan', 'Subaru :: Crosstrek', 'Land Rover :: Range Rover', 'Nissan :: Murano', 'Jeep :: Wrangler']\n", "TwoTower:\n", "['BMW :: X3', 'BMW :: 5 Series', 'Audi :: Q7', 'BMW :: 3 Series', 'Audi :: A4', 'Audi :: A3', 'Audi :: A6', 'Audi :: Q5', 'BMW :: X5', 'BMW :: Z4']\n" ] } ], "source": [ "# Example recommendations for one test query\n", "\n", "example_row = test_df.sample(1, random_state=RANDOM_STATE).iloc[0]\n", "\n", "print(\"Example query:\")\n", "display(example_row[feature_cols + [target_col]].to_frame().T)\n", "\n", "print(\"Popularity:\")\n", "print(recommend_popularity(example_row, 10))\n", "\n", "print(\"KNN:\")\n", "print(recommend_knn(example_row, 10))\n", "\n", "print(\"Linear:\")\n", "print(recommend_linear(example_row, 10))\n", "\n", "print(\"MLP:\")\n", "print(recommend_mlp(example_row, 10))\n", "\n", "print(\"TwoTower:\")\n", "print(recommend_twotower(example_row, 10))" ] }, { "cell_type": "code", "execution_count": 16, "id": "261628ca", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "================================================================================\n", "Evaluating: PopularityByCountry\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 5000/5000 [00:00<00:00, 211334.02it/s]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "================================================================================\n", "Evaluating: ContextKNN\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 5000/5000 [05:25<00:00, 15.37it/s]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "================================================================================\n", "Evaluating: LinearSoftmax\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 5000/5000 [00:14<00:00, 352.71it/s]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "================================================================================\n", "Evaluating: MLPClassifier\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 5000/5000 [00:02<00:00, 1680.72it/s]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "================================================================================\n", "Evaluating: TwoTower\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 5000/5000 [00:02<00:00, 1806.94it/s]\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", "
ModelHR@5HR@10HR@20MRR@10NDCG@10
0LinearSoftmax0.07200.13000.23660.0401160.060743
1MLPClassifier0.07080.12780.23940.0405760.060644
2PopularityByCountry0.07020.12600.24540.0425000.061681
3TwoTower0.06860.12480.24580.0383800.058247
4ContextKNN0.06060.12080.23620.0347300.054437
\n", "
" ], "text/plain": [ " Model HR@5 HR@10 HR@20 MRR@10 NDCG@10\n", "0 LinearSoftmax 0.0720 0.1300 0.2366 0.040116 0.060743\n", "1 MLPClassifier 0.0708 0.1278 0.2394 0.040576 0.060644\n", "2 PopularityByCountry 0.0702 0.1260 0.2454 0.042500 0.061681\n", "3 TwoTower 0.0686 0.1248 0.2458 0.038380 0.058247\n", "4 ContextKNN 0.0606 0.1208 0.2362 0.034730 0.054437" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# Evaluate all models\n", "\n", "results = []\n", "\n", "model_registry = {\n", " \"PopularityByCountry\": recommend_popularity,\n", " \"ContextKNN\": recommend_knn,\n", " \"LinearSoftmax\": recommend_linear,\n", " \"MLPClassifier\": recommend_mlp,\n", " \"TwoTower\": recommend_twotower,\n", "}\n", "\n", "for model_name, fn in model_registry.items():\n", " print(\"=\" * 80)\n", " print(\"Evaluating:\", model_name)\n", " summary, detail = evaluate_model(test_df, fn, ks=(5, 10, 20), show_progress=True)\n", " summary[\"Model\"] = model_name\n", " results.append(summary)\n", "\n", "results_df = pd.DataFrame(results)[[\"Model\", \"HR@5\", \"HR@10\", \"HR@20\", \"MRR@10\", \"NDCG@10\"]]\n", "results_df = results_df.sort_values([\"HR@10\", \"NDCG@10\"], ascending=False).reset_index(drop=True)\n", "\n", "display(results_df)" ] }, { "cell_type": "markdown", "id": "78e97eac", "metadata": {}, "source": [ "## Notes for the next iteration\n", "\n", "Likely next improvements:\n", "1. Better item definition: compare `Brand + Model` vs `Brand + Model + Condition`\n", "2. More query features: add price buckets, age, mileage buckets\n", "3. Hard-negative training for the neural retrieval model\n", "4. Better regular model: CatBoost or LightGBM ranking/classification\n", "5. Better neural model: deeper MLP, residual blocks, feature crosses, or a stronger retrieval loss" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "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": "3.14.3" } }, "nbformat": 4, "nbformat_minor": 5 }