{ "cells": [ { "cell_type": "markdown", "id": "72a8c1e8", "metadata": {}, "source": [ "# Car recommendation notebook (segment-based)\n", "\n", "This version reformulates the car task as a **regular recommender system**.\n", "\n", "Because the dataset has only one purchase per person and no real user history, we create **pseudo-users as market segments** built from:\n", "- `Country`\n", "- `Condition`\n", "- `Price` bin\n", "- `Mileage` bin\n", "- `Year` bin\n", "\n", "Each segment has a history of purchased car models. \n", "The task becomes:\n", "\n", "**Given a segment's past purchases, recommend the next car model that is likely to be relevant for that segment.**\n", "\n", "Models in this notebook:\n", "- Regular: Popularity, ItemKNN, EASE\n", "- Neural: NeuMF, MultVAE\n" ] }, { "cell_type": "code", "execution_count": 1, "id": "c6399ce0", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "/usr/lib/python3.14/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n", " from .autonotebook import tqdm as notebook_tqdm\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "Torch device: cuda\n" ] } ], "source": [ "\n", "from pathlib import Path\n", "import math\n", "import random\n", "import warnings\n", "\n", "import numpy as np\n", "import pandas as pd\n", "from scipy.sparse import csr_matrix\n", "from sklearn.preprocessing import LabelEncoder\n", "from sklearn.neighbors import NearestNeighbors\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", "warnings.filterwarnings(\"ignore\")\n", "\n", "RANDOM_STATE = 42\n", "np.random.seed(RANDOM_STATE)\n", "random.seed(RANDOM_STATE)\n", "torch.manual_seed(RANDOM_STATE)\n", "if torch.cuda.is_available():\n", " torch.cuda.manual_seed_all(RANDOM_STATE)\n", "\n", "DEVICE = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n", "print(\"Torch device:\", DEVICE)\n" ] }, { "cell_type": "code", "execution_count": 2, "id": "c980574b", "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": [ "\n", "# Paths and core settings\n", "\n", "BASE_DIR = Path(\"/home/konnilol/Documents/uni/kursovaya-sem5\")\n", "CSV_PATH = BASE_DIR / \"car_sales_dataset_with_person_details.csv\"\n", "\n", "if not CSV_PATH.exists():\n", " raise FileNotFoundError(f\"Dataset not found: {CSV_PATH}\")\n", "\n", "# Segment construction\n", "N_PRICE_BINS = 8\n", "N_MILEAGE_BINS = 8\n", "N_YEAR_BINS = 6\n", "\n", "# Filtering\n", "MIN_ITEMS_PER_SEGMENT = 8\n", "MIN_SEGMENTS_PER_ITEM = 8\n", "\n", "# Ranking\n", "TOP_KS = [5, 10, 20]\n", "\n", "# ItemKNN\n", "ITEM_KNN_NEIGHBORS = 30\n", "\n", "# EASE\n", "EASE_LAMBDA = 200.0\n", "\n", "# NeuMF\n", "NEUMF_NEG_PER_POS = 4\n", "NEUMF_EPOCHS = 10\n", "NEUMF_BATCH_SIZE = 4096\n", "NEUMF_LR = 1e-3\n", "NEUMF_MF_DIM = 32\n", "NEUMF_MLP_DIM = 64\n", "\n", "# MultVAE\n", "MULTVAE_EPOCHS = 60\n", "MULTVAE_BATCH_SIZE = 256\n", "MULTVAE_LR = 1e-3\n", "MULTVAE_HIDDEN_DIM = 256\n", "MULTVAE_LATENT_DIM = 64\n", "MULTVAE_KL_ANNEAL_EPOCHS = 20\n", "\n", "print(\"Using dataset:\", CSV_PATH)\n" ] }, { "cell_type": "code", "execution_count": 3, "id": "ddf27c4b", "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" }, { "data": { "text/plain": [ "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" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "# Load data\n", "\n", "df = pd.read_csv(CSV_PATH)\n", "\n", "print(\"Shape:\", df.shape)\n", "display(df.head())\n", "display(df.dtypes)\n" ] }, { "cell_type": "code", "execution_count": 4, "id": "98e34d8e", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Raw aggregated interactions: (632818, 3)\n", "Unique segments: 11520\n", "Unique 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", "
segment_iditem_idcount
0Australia | Certified Pre-Owned | p0 | m0 | y0Audi :: A33
1Australia | Certified Pre-Owned | p0 | m0 | y0Audi :: A44
2Australia | Certified Pre-Owned | p0 | m0 | y0Audi :: A63
3Australia | Certified Pre-Owned | p0 | m0 | y0Audi :: Q51
4Australia | Certified Pre-Owned | p0 | m0 | y0Audi :: Q72
\n", "
" ], "text/plain": [ " segment_id item_id count\n", "0 Australia | Certified Pre-Owned | p0 | m0 | y0 Audi :: A3 3\n", "1 Australia | Certified Pre-Owned | p0 | m0 | y0 Audi :: A4 4\n", "2 Australia | Certified Pre-Owned | p0 | m0 | y0 Audi :: A6 3\n", "3 Australia | Certified Pre-Owned | p0 | m0 | y0 Audi :: Q5 1\n", "4 Australia | Certified Pre-Owned | p0 | m0 | y0 Audi :: Q7 2" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "# Create item ids and market segments\n", "\n", "work = df.copy()\n", "\n", "work[\"Brand\"] = work[\"Brand\"].astype(str).str.strip()\n", "work[\"Model\"] = work[\"Model\"].astype(str).str.strip()\n", "work[\"Color\"] = work[\"Color\"].astype(str).str.strip()\n", "work[\"Condition\"] = work[\"Condition\"].astype(str).str.strip()\n", "work[\"Country\"] = work[\"Country\"].astype(str).str.strip()\n", "\n", "work[\"item_id\"] = work[\"Brand\"] + \" :: \" + work[\"Model\"]\n", "\n", "# Quantile bins keep segment sizes more balanced\n", "work[\"PriceBin\"] = pd.qcut(\n", " work[\"Price\"],\n", " q=N_PRICE_BINS,\n", " labels=[f\"p{i}\" for i in range(N_PRICE_BINS)],\n", " duplicates=\"drop\"\n", ").astype(str)\n", "\n", "work[\"MileageBin\"] = pd.qcut(\n", " work[\"Mileage\"],\n", " q=N_MILEAGE_BINS,\n", " labels=[f\"m{i}\" for i in range(N_MILEAGE_BINS)],\n", " duplicates=\"drop\"\n", ").astype(str)\n", "\n", "work[\"YearBin\"] = pd.qcut(\n", " work[\"Year\"],\n", " q=N_YEAR_BINS,\n", " labels=[f\"y{i}\" for i in range(N_YEAR_BINS)],\n", " duplicates=\"drop\"\n", ").astype(str)\n", "\n", "segment_cols = [\"Country\", \"Condition\", \"PriceBin\", \"MileageBin\", \"YearBin\"]\n", "work[\"segment_id\"] = work[segment_cols].astype(str).agg(\" | \".join, axis=1)\n", "\n", "# Aggregate to implicit interactions: segment -> item purchase count\n", "interactions = (\n", " work.groupby([\"segment_id\", \"item_id\"], as_index=False)\n", " .size()\n", " .rename(columns={\"size\": \"count\"})\n", ")\n", "\n", "print(\"Raw aggregated interactions:\", interactions.shape)\n", "print(\"Unique segments:\", interactions[\"segment_id\"].nunique())\n", "print(\"Unique items :\", interactions[\"item_id\"].nunique())\n", "display(interactions.head())\n" ] }, { "cell_type": "code", "execution_count": 5, "id": "f7a2f8a4", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "After filtering:\n", "Interactions : (632818, 3)\n", "Segments : 11520\n", "Items : 88\n", "Min items/seg : 38\n", "Min segs/item : 6988\n" ] }, { "data": { "text/plain": [ "item_id\n", "Chrysler :: Voyager 9182\n", "Chrysler :: 300 9137\n", "Chrysler :: Pacifica 9109\n", "Chevrolet :: Silverado 7271\n", "Hyundai :: Santa Fe 7218\n", "Jeep :: Cherokee 7217\n", "Toyota :: RAV4 7210\n", "Toyota :: Tacoma 7209\n", "Dodge :: Durango 7197\n", "Kia :: Forte 7194\n", "Volkswagen :: Jetta 7193\n", "Lexus :: ES 7189\n", "Hyundai :: Sonata 7187\n", "Chevrolet :: Trax 7185\n", "Ford :: Mustang 7185\n", "dtype: int64" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "# Iterative filtering so every segment and every item has enough support\n", "\n", "for _ in range(5):\n", " segment_sizes = interactions.groupby(\"segment_id\").size()\n", " keep_segments = segment_sizes[segment_sizes >= MIN_ITEMS_PER_SEGMENT].index\n", " interactions = interactions[interactions[\"segment_id\"].isin(keep_segments)].copy()\n", "\n", " item_sizes = interactions.groupby(\"item_id\").size()\n", " keep_items = item_sizes[item_sizes >= MIN_SEGMENTS_PER_ITEM].index\n", " interactions = interactions[interactions[\"item_id\"].isin(keep_items)].copy()\n", "\n", "segment_sizes = interactions.groupby(\"segment_id\").size()\n", "item_sizes = interactions.groupby(\"item_id\").size()\n", "\n", "print(\"After filtering:\")\n", "print(\"Interactions :\", interactions.shape)\n", "print(\"Segments :\", interactions['segment_id'].nunique())\n", "print(\"Items :\", interactions['item_id'].nunique())\n", "print(\"Min items/seg :\", int(segment_sizes.min()))\n", "print(\"Min segs/item :\", int(item_sizes.min()))\n", "display(item_sizes.sort_values(ascending=False).head(15))\n" ] }, { "cell_type": "code", "execution_count": 6, "id": "82319390", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Train interactions: (621298, 5)\n", "Test interactions : (11520, 5)\n", "Users: 11520\n", "Items: 88\n", "Matrix shape: (11520, 88)\n" ] } ], "source": [ "\n", "# Encode users/items and create leave-one-out split\n", "\n", "segment_encoder = LabelEncoder()\n", "item_encoder = LabelEncoder()\n", "\n", "interactions[\"user_idx\"] = segment_encoder.fit_transform(interactions[\"segment_id\"])\n", "interactions[\"item_idx\"] = item_encoder.fit_transform(interactions[\"item_id\"])\n", "\n", "num_users = interactions[\"user_idx\"].nunique()\n", "num_items = interactions[\"item_idx\"].nunique()\n", "\n", "rng = np.random.default_rng(RANDOM_STATE)\n", "\n", "test_idx = []\n", "for user_idx, g in interactions.groupby(\"user_idx\"):\n", " picked = g.sample(n=1, random_state=int(rng.integers(1_000_000))).index[0]\n", " test_idx.append(picked)\n", "\n", "test_interactions = interactions.loc[test_idx].copy()\n", "train_interactions = interactions.drop(index=test_idx).copy()\n", "\n", "print(\"Train interactions:\", train_interactions.shape)\n", "print(\"Test interactions :\", test_interactions.shape)\n", "print(\"Users:\", num_users)\n", "print(\"Items:\", num_items)\n", "\n", "# Sparse matrices\n", "rows = train_interactions[\"user_idx\"].to_numpy()\n", "cols = train_interactions[\"item_idx\"].to_numpy()\n", "vals = train_interactions[\"count\"].astype(np.float32).to_numpy()\n", "\n", "X_counts = csr_matrix((vals, (rows, cols)), shape=(num_users, num_items), dtype=np.float32)\n", "X_binary = csr_matrix((np.ones_like(vals), (rows, cols)), shape=(num_users, num_items), dtype=np.float32)\n", "\n", "# Dense versions are safe here because the item catalog is small\n", "X_counts_dense = X_counts.toarray().astype(np.float32)\n", "X_binary_dense = X_binary.toarray().astype(np.float32)\n", "\n", "user_seen = {\n", " int(uid): set(g[\"item_idx\"].tolist())\n", " for uid, g in train_interactions.groupby(\"user_idx\")\n", "}\n", "\n", "test_item_by_user = {\n", " int(uid): int(item)\n", " for uid, item in zip(test_interactions[\"user_idx\"], test_interactions[\"item_idx\"])\n", "}\n", "\n", "item_ids = item_encoder.classes_\n", "global_popularity = np.asarray(X_counts.sum(axis=0)).ravel()\n", "global_pop_rank = np.argsort(global_popularity)[::-1]\n", "\n", "print(\"Matrix shape:\", X_binary.shape)\n" ] }, { "cell_type": "code", "execution_count": 7, "id": "3bd6dd5e", "metadata": {}, "outputs": [], "source": [ "\n", "# Metrics\n", "\n", "def hit_rate_at_k(recs, true_item, k):\n", " return 1.0 if true_item in recs[:k] else 0.0\n", "\n", "def mrr_at_k(recs, true_item, k):\n", " recs_k = recs[:k]\n", " if true_item in recs_k:\n", " rank = recs_k.index(true_item) + 1\n", " return 1.0 / rank\n", " return 0.0\n", "\n", "def ndcg_at_k(recs, true_item, k):\n", " recs_k = recs[:k]\n", " if true_item in recs_k:\n", " rank = recs_k.index(true_item) + 1\n", " return 1.0 / math.log2(rank + 1.0)\n", " return 0.0\n", "\n", "def evaluate_model(model_func, model_name, user_ids, ks=(5, 10, 20)):\n", " hits = {k: 0.0 for k in ks}\n", " mrr = 0.0\n", " ndcg = 0.0\n", " valid = 0\n", "\n", " max_k = max(ks)\n", "\n", " print(\"=\" * 80)\n", " print(\"Evaluating:\", model_name)\n", "\n", " for uid in tqdm(user_ids):\n", " true_item = test_item_by_user.get(int(uid))\n", " if true_item is None:\n", " continue\n", "\n", " recs = model_func(int(uid), n=max_k)\n", " if not isinstance(recs, list):\n", " recs = list(recs)\n", "\n", " valid += 1\n", " for k in ks:\n", " hits[k] += hit_rate_at_k(recs, true_item, k)\n", " mrr += mrr_at_k(recs, true_item, 10)\n", " ndcg += ndcg_at_k(recs, true_item, 10)\n", "\n", " row = {\n", " \"Model\": model_name,\n", " \"HR@5\": hits[5] / valid if valid else 0.0,\n", " \"HR@10\": hits[10] / valid if valid else 0.0,\n", " \"HR@20\": hits[20] / valid if valid else 0.0,\n", " \"MRR@10\": mrr / valid if valid else 0.0,\n", " \"NDCG@10\": ndcg / valid if valid else 0.0,\n", " }\n", " return row\n" ] }, { "cell_type": "markdown", "id": "d69f57f9", "metadata": {}, "source": [ "## Regular model 1: Popularity baseline" ] }, { "cell_type": "code", "execution_count": 8, "id": "cb2afb07", "metadata": {}, "outputs": [], "source": [ "\n", "def recommend_popularity(user_idx, n=10):\n", " seen = user_seen.get(user_idx, set())\n", " return [int(i) for i in global_pop_rank if int(i) not in seen][:n]\n" ] }, { "cell_type": "markdown", "id": "33bc78c5", "metadata": {}, "source": [ "## Regular model 2: ItemKNN" ] }, { "cell_type": "code", "execution_count": 9, "id": "d1c2e1c1", "metadata": {}, "outputs": [], "source": [ "\n", "# Item-item KNN on binary segment-item interactions\n", "\n", "item_user = X_binary.T.tocsr()\n", "\n", "item_knn = NearestNeighbors(\n", " metric=\"cosine\",\n", " algorithm=\"brute\",\n", " n_neighbors=min(ITEM_KNN_NEIGHBORS, num_items),\n", " n_jobs=-1,\n", ")\n", "item_knn.fit(item_user)\n", "\n", "knn_distances, knn_indices = item_knn.kneighbors(\n", " item_user,\n", " n_neighbors=min(ITEM_KNN_NEIGHBORS, num_items)\n", ")\n", "knn_similarities = (1.0 - knn_distances).astype(np.float32)\n", "\n", "user_item_strength = {\n", " int(uid): {\n", " int(item): float(cnt)\n", " for item, cnt in zip(g[\"item_idx\"], g[\"count\"])\n", " }\n", " for uid, g in train_interactions.groupby(\"user_idx\")\n", "}\n", "\n", "def recommend_itemknn(user_idx, n=10):\n", " seen = user_seen.get(user_idx, set())\n", " strengths = user_item_strength.get(user_idx, {})\n", "\n", " scores = np.zeros(num_items, dtype=np.float32)\n", "\n", " for item_idx, cnt in strengths.items():\n", " weight = np.log1p(cnt)\n", " nbrs = knn_indices[item_idx]\n", " sims = knn_similarities[item_idx]\n", "\n", " for nbr_idx, sim in zip(nbrs, sims):\n", " if nbr_idx == item_idx:\n", " continue\n", " scores[nbr_idx] += weight * sim\n", "\n", " if seen:\n", " scores[list(seen)] = -np.inf\n", "\n", " n = min(n, num_items - len(seen))\n", " if n <= 0:\n", " return []\n", "\n", " top = np.argpartition(scores, -n)[-n:]\n", " top = top[np.argsort(scores[top])[::-1]]\n", " return top.tolist()\n" ] }, { "cell_type": "markdown", "id": "7d9b5413", "metadata": {}, "source": [ "## Regular model 3: EASE" ] }, { "cell_type": "code", "execution_count": 10, "id": "758ab370", "metadata": {}, "outputs": [], "source": [ "\n", "# EASE on binary interactions\n", "\n", "G = X_binary_dense.T @ X_binary_dense\n", "G[np.diag_indices(num_items)] += EASE_LAMBDA\n", "\n", "P = np.linalg.inv(G)\n", "B = -P / np.diag(P)\n", "B[np.diag_indices(num_items)] = 0.0\n", "\n", "def recommend_ease(user_idx, n=10):\n", " scores = X_binary_dense[user_idx] @ B\n", " seen = user_seen.get(user_idx, set())\n", "\n", " if seen:\n", " scores[list(seen)] = -np.inf\n", "\n", " n = min(n, num_items - len(seen))\n", " if n <= 0:\n", " return []\n", "\n", " top = np.argpartition(scores, -n)[-n:]\n", " top = top[np.argsort(scores[top])[::-1]]\n", " return top.tolist()\n" ] }, { "cell_type": "markdown", "id": "887cfa54", "metadata": {}, "source": [ "## Neural model 1: NeuMF" ] }, { "cell_type": "code", "execution_count": 11, "id": "452e132d", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "NeuMF epoch 1/10 - loss: 0.5025\n", "NeuMF epoch 2/10 - loss: 0.4987\n", "NeuMF epoch 3/10 - loss: 0.4974\n", "NeuMF epoch 4/10 - loss: 0.4952\n", "NeuMF epoch 5/10 - loss: 0.4914\n", "NeuMF epoch 6/10 - loss: 0.4857\n", "NeuMF epoch 7/10 - loss: 0.4782\n", "NeuMF epoch 8/10 - loss: 0.4694\n", "NeuMF epoch 9/10 - loss: 0.4596\n", "NeuMF epoch 10/10 - loss: 0.4495\n" ] } ], "source": [ "\n", "class ImplicitPairDataset(Dataset):\n", " def __init__(self, user_seen_dict, num_items, neg_per_pos=4, seed=42):\n", " self.user_seen_dict = user_seen_dict\n", " self.num_items = num_items\n", " self.neg_per_pos = neg_per_pos\n", " self.seed = seed\n", " self.rng = np.random.default_rng(seed)\n", " self.positive_pairs = [\n", " (int(u), int(i))\n", " for u, items in user_seen_dict.items()\n", " for i in items\n", " ]\n", " self.samples = []\n", " self.refresh()\n", "\n", " def refresh(self):\n", " samples = []\n", " for u, i in self.positive_pairs:\n", " samples.append((u, i, 1.0))\n", "\n", " negs = 0\n", " pos_set = self.user_seen_dict[u]\n", " while negs < self.neg_per_pos:\n", " j = int(self.rng.integers(self.num_items))\n", " if j not in pos_set:\n", " samples.append((u, j, 0.0))\n", " negs += 1\n", "\n", " self.samples = samples\n", "\n", " def __len__(self):\n", " return len(self.samples)\n", "\n", " def __getitem__(self, idx):\n", " u, i, y = self.samples[idx]\n", " return (\n", " torch.tensor(u, dtype=torch.long),\n", " torch.tensor(i, dtype=torch.long),\n", " torch.tensor(y, dtype=torch.float32),\n", " )\n", "\n", "class NeuMF(nn.Module):\n", " def __init__(self, num_users, num_items, mf_dim=32, mlp_dim=64):\n", " super().__init__()\n", " self.user_mf = nn.Embedding(num_users, mf_dim)\n", " self.item_mf = nn.Embedding(num_items, mf_dim)\n", "\n", " self.user_mlp = nn.Embedding(num_users, mlp_dim)\n", " self.item_mlp = nn.Embedding(num_items, mlp_dim)\n", "\n", " self.mlp = nn.Sequential(\n", " nn.Linear(mlp_dim * 2, 128),\n", " nn.ReLU(),\n", " nn.Dropout(0.15),\n", " nn.Linear(128, 64),\n", " nn.ReLU(),\n", " )\n", "\n", " self.out = nn.Linear(mf_dim + 64, 1)\n", "\n", " def forward(self, user_idx, item_idx):\n", " mf_part = self.user_mf(user_idx) * self.item_mf(item_idx)\n", " mlp_part = self.mlp(\n", " torch.cat([self.user_mlp(user_idx), self.item_mlp(item_idx)], dim=1)\n", " )\n", " x = torch.cat([mf_part, mlp_part], dim=1)\n", " return self.out(x).squeeze(1)\n", "\n", " @torch.no_grad()\n", " def score_all_items(self, user_idx):\n", " self.eval()\n", " device = next(self.parameters()).device\n", " item_idx = torch.arange(self.item_mf.num_embeddings, device=device)\n", " user_vec = torch.full_like(item_idx, fill_value=int(user_idx))\n", " scores = self.forward(user_vec, item_idx)\n", " return scores.detach().cpu().numpy()\n", "\n", "neumf_model = NeuMF(\n", " num_users=num_users,\n", " num_items=num_items,\n", " mf_dim=NEUMF_MF_DIM,\n", " mlp_dim=NEUMF_MLP_DIM,\n", ").to(DEVICE)\n", "\n", "neumf_dataset = ImplicitPairDataset(\n", " user_seen_dict=user_seen,\n", " num_items=num_items,\n", " neg_per_pos=NEUMF_NEG_PER_POS,\n", " seed=RANDOM_STATE,\n", ")\n", "\n", "neumf_loader = DataLoader(\n", " neumf_dataset,\n", " batch_size=NEUMF_BATCH_SIZE,\n", " shuffle=True,\n", " num_workers=0,\n", " pin_memory=(DEVICE == \"cuda\"),\n", ")\n", "\n", "neumf_optimizer = torch.optim.Adam(neumf_model.parameters(), lr=NEUMF_LR)\n", "neumf_criterion = nn.BCEWithLogitsLoss()\n", "\n", "for epoch in range(NEUMF_EPOCHS):\n", " neumf_dataset.refresh()\n", " neumf_model.train()\n", " total_loss = 0.0\n", "\n", " for user_idx, item_idx, target in neumf_loader:\n", " user_idx = user_idx.to(DEVICE)\n", " item_idx = item_idx.to(DEVICE)\n", " target = target.to(DEVICE)\n", "\n", " neumf_optimizer.zero_grad()\n", " logits = neumf_model(user_idx, item_idx)\n", " loss = neumf_criterion(logits, target)\n", " loss.backward()\n", " neumf_optimizer.step()\n", "\n", " total_loss += loss.item() * len(target)\n", "\n", " epoch_loss = total_loss / len(neumf_dataset)\n", " print(f\"NeuMF epoch {epoch + 1}/{NEUMF_EPOCHS} - loss: {epoch_loss:.4f}\")\n", "\n", "@torch.no_grad()\n", "def recommend_neumf(user_idx, n=10):\n", " scores = neumf_model.score_all_items(user_idx)\n", " seen = user_seen.get(user_idx, set())\n", "\n", " if seen:\n", " scores[list(seen)] = -np.inf\n", "\n", " n = min(n, num_items - len(seen))\n", " if n <= 0:\n", " return []\n", "\n", " top = np.argpartition(scores, -n)[-n:]\n", " top = top[np.argsort(scores[top])[::-1]]\n", " return top.tolist()\n" ] }, { "cell_type": "markdown", "id": "3acaa6c4", "metadata": {}, "source": [ "## Neural model 2: MultVAE" ] }, { "cell_type": "code", "execution_count": 12, "id": "af3565cf", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "MultVAE epoch 1/60 - loss: 241.3210 - anneal: 0.050\n", "MultVAE epoch 10/60 - loss: 241.6675 - anneal: 0.500\n", "MultVAE epoch 20/60 - loss: 241.4924 - anneal: 1.000\n", "MultVAE epoch 30/60 - loss: 241.4797 - anneal: 1.000\n", "MultVAE epoch 40/60 - loss: 241.4659 - anneal: 1.000\n", "MultVAE epoch 50/60 - loss: 241.4500 - anneal: 1.000\n", "MultVAE epoch 60/60 - loss: 241.4394 - anneal: 1.000\n" ] } ], "source": [ "\n", "class MultVAE(nn.Module):\n", " def __init__(self, num_items, hidden_dim=256, latent_dim=64):\n", " super().__init__()\n", " self.encoder = nn.Sequential(\n", " nn.Linear(num_items, hidden_dim),\n", " nn.Tanh(),\n", " nn.Linear(hidden_dim, hidden_dim),\n", " nn.Tanh(),\n", " )\n", " self.mu = nn.Linear(hidden_dim, latent_dim)\n", " self.logvar = nn.Linear(hidden_dim, latent_dim)\n", "\n", " self.decoder = nn.Sequential(\n", " nn.Linear(latent_dim, hidden_dim),\n", " nn.Tanh(),\n", " nn.Linear(hidden_dim, num_items),\n", " )\n", "\n", " def forward(self, x):\n", " h = self.encoder(x)\n", " mu = self.mu(h)\n", " logvar = self.logvar(h)\n", "\n", " if self.training:\n", " std = torch.exp(0.5 * logvar)\n", " eps = torch.randn_like(std)\n", " z = mu + eps * std\n", " else:\n", " z = mu\n", "\n", " logits = self.decoder(z)\n", " return logits, mu, logvar\n", "\n", "def multvae_loss_fn(logits, x, mu, logvar, anneal=1.0):\n", " recon = -(torch.log_softmax(logits, dim=1) * x).sum(dim=1).mean()\n", " kl = -0.5 * torch.mean(torch.sum(1 + logvar - mu.pow(2) - logvar.exp(), dim=1))\n", " return recon + anneal * kl, recon.detach(), kl.detach()\n", "\n", "multvae_model = MultVAE(\n", " num_items=num_items,\n", " hidden_dim=MULTVAE_HIDDEN_DIM,\n", " latent_dim=MULTVAE_LATENT_DIM,\n", ").to(DEVICE)\n", "\n", "multvae_optimizer = torch.optim.Adam(multvae_model.parameters(), lr=MULTVAE_LR)\n", "\n", "X_multvae = torch.tensor(X_binary_dense, dtype=torch.float32)\n", "\n", "for epoch in range(MULTVAE_EPOCHS):\n", " multvae_model.train()\n", " perm = torch.randperm(X_multvae.size(0))\n", " total_loss = 0.0\n", "\n", " anneal = min(1.0, (epoch + 1) / MULTVAE_KL_ANNEAL_EPOCHS)\n", "\n", " for start in range(0, X_multvae.size(0), MULTVAE_BATCH_SIZE):\n", " idx = perm[start:start + MULTVAE_BATCH_SIZE]\n", " batch = X_multvae[idx].to(DEVICE)\n", "\n", " multvae_optimizer.zero_grad()\n", " logits, mu, logvar = multvae_model(batch)\n", " loss, recon, kl = multvae_loss_fn(logits, batch, mu, logvar, anneal=anneal)\n", " loss.backward()\n", " multvae_optimizer.step()\n", "\n", " total_loss += loss.item() * batch.size(0)\n", "\n", " epoch_loss = total_loss / X_multvae.size(0)\n", " if (epoch + 1) % 10 == 0 or epoch == 0:\n", " print(f\"MultVAE epoch {epoch + 1}/{MULTVAE_EPOCHS} - loss: {epoch_loss:.4f} - anneal: {anneal:.3f}\")\n", "\n", "@torch.no_grad()\n", "def recommend_multvae(user_idx, n=10):\n", " multvae_model.eval()\n", "\n", " x = torch.tensor(X_binary_dense[user_idx:user_idx + 1], dtype=torch.float32, device=DEVICE)\n", " logits, _, _ = multvae_model(x)\n", " scores = logits[0].detach().cpu().numpy()\n", "\n", " seen = user_seen.get(user_idx, set())\n", " if seen:\n", " scores[list(seen)] = -np.inf\n", "\n", " n = min(n, num_items - len(seen))\n", " if n <= 0:\n", " return []\n", "\n", " top = np.argpartition(scores, -n)[-n:]\n", " top = top[np.argsort(scores[top])[::-1]]\n", " return top.tolist()\n" ] }, { "cell_type": "code", "execution_count": 13, "id": "1e88dc7d", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Example segment:\n", "France | Certified Pre-Owned | p2 | m4 | y0\n", "\n", "Seen items (first 15):\n", "['Audi :: A6', 'Audi :: Q7', 'BMW :: 5 Series', 'BMW :: X5', 'BMW :: Z4', 'Chevrolet :: Equinox', 'Chevrolet :: Malibu', 'Chevrolet :: Silverado', 'Chevrolet :: Trax', 'Chrysler :: 300', 'Chrysler :: Pacifica', 'Chrysler :: Voyager', 'Dodge :: Challenger', 'Dodge :: Journey', 'Ford :: Escape']\n", "\n", "Popularity:\n", "['Kia :: Optima', 'Kia :: Forte', 'Hyundai :: Tucson', 'Toyota :: Tacoma', 'Volkswagen :: Tiguan', 'Dodge :: Charger', 'Volkswagen :: Jetta', 'Honda :: CR-V', 'Toyota :: RAV4', 'Audi :: A4']\n", "\n", "ItemKNN:\n", "['Kia :: Forte', 'Toyota :: RAV4', 'Toyota :: Tacoma', 'Dodge :: Durango', 'Volkswagen :: Jetta', 'Hyundai :: Tucson', 'Jeep :: Renegade', 'Kia :: Optima', 'Land Rover :: Velar', 'Audi :: A3']\n", "\n", "EASE:\n", "['Dodge :: Ram 1500', 'Honda :: CR-V', 'Volkswagen :: Passat', 'Audi :: A4', 'Mazda :: Mazda6', 'Kia :: Optima', 'Volkswagen :: Atlas', 'Toyota :: RAV4', 'Nissan :: Titan', 'Hyundai :: Kona']\n", "\n", "NeuMF:\n", "['Kia :: Forte', 'Audi :: Q5', 'Nissan :: Murano', 'Dodge :: Ram 1500', 'Mazda :: CX-5', 'Mazda :: Mazda6', 'BMW :: 3 Series', 'Volkswagen :: Passat', 'Ford :: Explorer', 'Toyota :: Camry']\n", "\n", "MultVAE:\n", "['Volkswagen :: Jetta', 'Dodge :: Durango', 'Toyota :: RAV4', 'Kia :: Optima', 'Hyundai :: Tucson', 'Nissan :: Murano', 'Kia :: Forte', 'Honda :: CR-V', 'Ford :: Explorer', 'Toyota :: Tacoma']\n" ] } ], "source": [ "\n", "# Example recommendations for one segment\n", "\n", "example_user = int(train_interactions[\"user_idx\"].sample(1, random_state=RANDOM_STATE).iloc[0])\n", "example_segment = segment_encoder.inverse_transform([example_user])[0]\n", "seen_items = [item_ids[i] for i in sorted(user_seen[example_user])[:15]]\n", "\n", "print(\"Example segment:\")\n", "print(example_segment)\n", "print()\n", "print(\"Seen items (first 15):\")\n", "print(seen_items)\n", "print()\n", "print(\"Popularity:\")\n", "print([item_ids[i] for i in recommend_popularity(example_user, n=10)])\n", "print()\n", "print(\"ItemKNN:\")\n", "print([item_ids[i] for i in recommend_itemknn(example_user, n=10)])\n", "print()\n", "print(\"EASE:\")\n", "print([item_ids[i] for i in recommend_ease(example_user, n=10)])\n", "print()\n", "print(\"NeuMF:\")\n", "print([item_ids[i] for i in recommend_neumf(example_user, n=10)])\n", "print()\n", "print(\"MultVAE:\")\n", "print([item_ids[i] for i in recommend_multvae(example_user, n=10)])\n" ] }, { "cell_type": "code", "execution_count": 14, "id": "0c6507f1", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "================================================================================\n", "Evaluating: Popularity\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 11520/11520 [00:00<00:00, 107535.01it/s]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "================================================================================\n", "Evaluating: ItemKNN\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 11520/11520 [00:08<00:00, 1348.76it/s]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "================================================================================\n", "Evaluating: EASE\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 11520/11520 [00:00<00:00, 74883.31it/s]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "================================================================================\n", "Evaluating: NeuMF\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 11520/11520 [00:03<00:00, 2949.40it/s]\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "================================================================================\n", "Evaluating: MultVAE\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 11520/11520 [00:02<00:00, 3972.26it/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
0ItemKNN0.1736110.3237850.6118060.1085830.157806
1MultVAE0.1763020.3210070.6102430.1102430.158543
2Popularity0.1706600.3210070.6118060.1090120.157498
3EASE0.1692710.3175350.6058160.1076570.155620
4NeuMF0.1598090.3070310.6030380.0913760.140766
\n", "
" ], "text/plain": [ " Model HR@5 HR@10 HR@20 MRR@10 NDCG@10\n", "0 ItemKNN 0.173611 0.323785 0.611806 0.108583 0.157806\n", "1 MultVAE 0.176302 0.321007 0.610243 0.110243 0.158543\n", "2 Popularity 0.170660 0.321007 0.611806 0.109012 0.157498\n", "3 EASE 0.169271 0.317535 0.605816 0.107657 0.155620\n", "4 NeuMF 0.159809 0.307031 0.603038 0.091376 0.140766" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "# Evaluate all models\n", "\n", "eval_users = np.array(sorted(test_item_by_user.keys()))\n", "\n", "results = []\n", "results.append(evaluate_model(recommend_popularity, \"Popularity\", eval_users, ks=TOP_KS))\n", "results.append(evaluate_model(recommend_itemknn, \"ItemKNN\", eval_users, ks=TOP_KS))\n", "results.append(evaluate_model(recommend_ease, \"EASE\", eval_users, ks=TOP_KS))\n", "results.append(evaluate_model(recommend_neumf, \"NeuMF\", eval_users, ks=TOP_KS))\n", "results.append(evaluate_model(recommend_multvae, \"MultVAE\", eval_users, ks=TOP_KS))\n", "\n", "results_df = pd.DataFrame(results).sort_values([\"HR@10\", \"NDCG@10\"], ascending=False).reset_index(drop=True)\n", "display(results_df)\n" ] }, { "cell_type": "markdown", "id": "bdc8372a", "metadata": {}, "source": [ "## Notes\n", "\n", "If the results are still weak, the next iteration should focus on:\n", "1. better segment design\n", "2. stronger weighting of interaction counts\n", "3. tuning EASE regularization\n", "4. tuning NeuMF negative sampling and epochs\n", "5. tuning MultVAE latent size and KL annealing\n" ] } ], "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 }