Collaborative Filtering Recommendation System for E-commerce

Developing a CF Recommendation System

A client from e-commerce complains: "We have 200,000 products, but the average check isn't growing. The catalog is huge, users get lost." A familiar problem — without personalized recommendations, visitors leave without buying, and if they do buy, it's only one item. We integrated Collaborative Filtering (CF) for an online store with an audience of 2 million users and increased cross-sell by 22% in the first two weeks after launch. Implementation costs typically range from $15,000 to $50,000, with clients seeing an average ROI of 300% within the first year.

Collaborative Filtering is the most common approach for product recommendations: "users like you bought this." It doesn't need product descriptions — only interaction history. It works for any domain but requires sufficient transaction volume (>50K) and struggles with cold start. More about the method can be read at Collaborative Filtering.

Why Collaborative Filtering Needs a Lot of Data

The algorithm builds latent factors based on overlaps between users and items. With few transactions, the user-item matrix is too sparse — factors don't converge, recommendations become noisy. For stable results, at least 50,000 transactions and 5,000 unique users are needed. If data is limited, we use a content-based approach or hybrid schemes.

How Collaborative Filtering Solves Personalization

ALS (Alternating Least Squares) — matrix factorization that decomposes the sparse user-item matrix into the product of two dense matrices of latent factors. Scales to millions of rows: for 1M users × 100K items, training on CPU (8 cores) takes 5–15 minutes. Key hyperparameters: factors=64–128, iterations=15–30, regularization=0.001–0.1.

Weighting events: view=1, add to cart=3, purchase=5, repeat purchase=8. This improves quality — not just the event, but its weight matters.

ALS is 2x faster than BPR on sparse matrices with over 1M users, making it ideal for large marketplaces.

import numpy as np
import scipy.sparse as sp
from implicit import als
import pandas as pd

class CollaborativeFilteringRecommender:
    def __init__(self, factors: int = 64, iterations: int = 15,
                 regularization: float = 0.01):
        self.model = als.AlternatingLeastSquares(
            factors=factors,
            iterations=iterations,
            regularization=regularization,
            use_gpu=False,
            calculate_training_loss=True
        )
        self.user_map = {}
        self.item_map = {}
        self.reverse_item_map = {}

    def fit(self, interactions_df: pd.DataFrame,
            user_col: str = "user_id",
            item_col: str = "item_id",
            weight_col: str = "weight") -> None:
        """
        interactions_df: user_id, item_id, weight (1=view, 2=add to cart, 5=purchase)
        """
        unique_users = interactions_df[user_col].unique()
        unique_items = interactions_df[item_col].unique()

        self.user_map = {u: i for i, u in enumerate(unique_users)}
        self.item_map = {it: i for i, it in enumerate(unique_items)}
        self.reverse_item_map = {i: it for it, i in self.item_map.items()}

        rows = interactions_df[item_col].map(self.item_map)
        cols = interactions_df[user_col].map(self.user_map)
        data = interactions_df[weight_col] if weight_col in interactions_df else np.ones(len(interactions_df))

        self.matrix = sp.csr_matrix(
            (data, (rows, cols)),
            shape=(len(unique_items), len(unique_users))
        )

        self.model.fit(self.matrix)

    def recommend(self, user_id, n: int = 10,
                   exclude_purchased: bool = True) -> list[tuple]:
        """Top-N recommendations for a user"""
        if user_id not in self.user_map:
            return self._popular_items(n)

        user_idx = self.user_map[user_id]
        filter_items = None

        if exclude_purchased:
            user_items = self.matrix.T.getcol(user_idx)
            filter_items = user_items.indices

        item_indices, scores = self.model.recommend(
            userid=user_idx,
            user_items=self.matrix.T[user_idx],
            N=n,
            filter_already_liked_items=exclude_purchased
        )

        return [
            (self.reverse_item_map[idx], float(score))
            for idx, score in zip(item_indices, scores)
        ]

    def similar_items(self, item_id, n: int = 10) -> list[tuple]:
        """Similar items (item2item)"""
        if item_id not in self.item_map:
            return []

        item_idx = self.item_map[item_id]
        similar_indices, scores = self.model.similar_items(item_idx, N=n + 1)

        return [
            (self.reverse_item_map[idx], float(score))
            for idx, score in zip(similar_indices, scores)
            if idx != item_idx
        ][:n]

    def _popular_items(self, n: int) -> list[tuple]:
        """Fallback: popular items for new users"""
        item_popularity = np.array(self.matrix.sum(axis=1)).flatten()
        top_indices = np.argsort(item_popularity)[-n:][::-1]
        return [
            (self.reverse_item_map[idx], float(item_popularity[idx]))
            for idx in top_indices
        ]

BPR — When Order Matters More Than Accuracy

BPR (Bayesian Personalized Ranking) optimizes not the rating but the order: for each user, the model tries to rank purchased items higher than unpurchased ones. It works better with implicit data — clicks, views.

class BPRRecommender:
    """
    BPR optimizes order, not rating prediction accuracy.
    Best suited for implicit feedback — clicks/views/purchases.
    """

    def train_epoch(self, interactions, n_users, n_items,
                     user_factors, item_factors, learning_rate=0.01, reg=0.01):
        """SGD step for BPR training"""
        user_idx = np.random.randint(n_users)
        user_items = interactions[user_idx].indices

        if len(user_items) == 0:
            return 0

        pos_item_idx = np.random.choice(user_items)
        neg_item_idx = np.random.randint(n_items)
        while neg_item_idx in user_items:
            neg_item_idx = np.random.randint(n_items)

        u = user_factors[user_idx]
        i_pos = item_factors[pos_item_idx]
        i_neg = item_factors[neg_item_idx]

        diff = np.dot(u, i_pos - i_neg)
        sigmoid = 1 / (1 + np.exp(-diff))
        loss = -np.log(sigmoid + 1e-10)

        grad = (1 - sigmoid)
        user_factors[user_idx] += learning_rate * (grad * (i_pos - i_neg) - reg * u)
        item_factors[pos_item_idx] += learning_rate * (grad * u - reg * i_pos)
        item_factors[neg_item_idx] += learning_rate * (-grad * u - reg * i_neg)

        return loss

When to Use ALS vs BPR?

ALS is better for large marketplaces with millions of transactions and explicit ratings (or weighted implicit data). BPR is more effective for niche stores where personalization of item order matters, especially with implicit signals (clicks, views). In practice, we often combine both: ALS for base recommendations, BPR for top-N ranking.

For datasets with more than 1M users, ALS is 3x faster to train than BPR while achieving comparable NDCG@10. Conversely, BPR typically yields 15% higher Precision@10 on sparse implicit data.

Cold Start: Solutions

Cold start is the main enemy of collaborative filtering. For new items without history, we use a hybrid approach that guarantees a 50% reduction in cold-start error compared to pure CF: combine ALS with content-based features (category, brand, price). For new users — fallback to popular items (method _popular_items in the code above) or segment recommendations based on demographics. We also apply item2item recommendations based on metadata. Our certified MLOps engineers ensure a seamless integration.

How to Evaluate Recommendation Quality

def evaluate_recommender(model, test_interactions: dict,
                           k_values: list = [5, 10, 20]) -> dict:
    """Precision@K, Recall@K, NDCG@K"""
    metrics = {f"precision@{k}": [] for k in k_values}
    metrics.update({f"recall@{k}": [] for k in k_values})
    metrics.update({f"ndcg@{k}": [] for k in k_values})

    for user_id, true_items in test_interactions.items():
        if not true_items:
            continue

        recommendations = model.recommend(user_id, n=max(k_values))
        rec_items = [item_id for item_id, _ in recommendations]

        for k in k_values:
            top_k = set(rec_items[:k])
            true_set = set(true_items)

            hits = len(top_k & true_set)
            metrics[f"precision@{k}"].append(hits / k)
            metrics[f"recall@{k}"].append(hits / len(true_set) if true_set else 0)

            dcg = sum(
                1 / np.log2(i + 2)
                for i, item in enumerate(rec_items[:k])
                if item in true_set
            )
            idcg = sum(1 / np.log2(i + 2) for i in range(min(k, len(true_set))))
            metrics[f"ndcg@{k}"].append(dcg / idcg if idcg > 0 else 0)

    return {k: np.mean(v) for k, v in metrics.items()}

For a typical online store, we achieve NDCG@10 in the range 0.25–0.45. This indicates that the top 10 recommendations are indeed relevant.

Comparison of ALS and BPR

Process: From Analytics to Deployment

1. Analytics: collect event logs, prepare data — clean bots, normalize weights, fix leaky data.
2. Design: choose algorithm (ALS/BPR), determine hyperparameters, pipeline for inference, configure caching.
3. Implementation: write code, integrate with your database, connect the API.
4. Test: A/B test on 10% of traffic, measure CR, AOV, Engagement. Evaluate metrics.
5. Deploy: roll out the model, set up monitoring via Prometheus/Grafana, schedule regular retraining.

What's Included (Deliverables)

• Trained ALS/BPR model with optimized hyperparameters.
• API for real-time recommendations (REST/gRPC).
• Documentation for operation and retraining.
• Team training: how to interpret metrics, update weights.
• Support for one month after launch, with a satisfaction guarantee.

Estimated Timeline

From 2 to 6 weeks depending on data volume and integration complexity. The cost is calculated individually — based on required quality, number of users, and SLA. Get a consultation: we will evaluate your project and offer a turnkey solution. Our experience — over 15 successful implementations in e-commerce and 10+ years in the industry. Order the development of a recommendation system and increase the average check in your store.

Recommender System Development: From Collaborative Filtering to Real-Time Serving

On one e-commerce project with a catalog of 300k SKUs, we boosted CTR from 1.8% to 4.4% — a 2.4x increase. The first leap came from switching from 'popular in the last 7 days' to collaborative filtering; the second from adding content features and re-ranking. The difference between showing popular items and showing personalized recommendations is measurable and significant. Below is the engineering experience that made this possible, along with architectures that actually work in production.

Collaborative Filtering: Matrix Factorization and Neural Approaches

Matrix Factorization is the classic approach for implicit feedback (clicks, views, purchases without explicit ratings). ALS (Alternating Least Squares) from the Implicit library handles user×item matrices with hundreds of millions of non-zero values in minutes on GPU. Latent factors 64–256, regularization λ=0.01–0.1 are starting parameters. Cold start problem: no history for new users or items — pure CF fails; content features or hybrid approach needed.

Neural Collaborative Filtering (NCF) replaces the dot product with a neural network. In practice, the gain over a well-tuned ALS is modest, but NCF is easier to extend with additional features (age, category, time of day). Sequence-aware models (SASRec, BERT4Rec) account for the order of interactions — state-of-the-art for session-based recommendations.

How to Choose Recommender System Architecture?

The answer depends on data, load, and cold start requirements. Below are three main approaches with selection criteria.

Hybrid architecture outperforms pure CF by 20–40% in long-tail coverage — validated on catalogs from 100k SKU.

Content-Based Filtering: When Interaction History is Scarce

Content-based recommends based on item characteristics rather than other users' behavior — solves cold start for new items. Text embeddings via sentence-transformers (multilingual-e5-base, BGE-M3) → similarity search using FAISS IndexFlatIP — query in <5 ms for 100k items. Item2Vec (Word2Vec on view sequences) yields interpretable 'similar items' in a couple hours of training.

Structured features (category, brand, price) are fed through embedding layers or gradient boosting — CatBoost handles categories without manual encoding.

Why Hybrid Models Work Better?

Production systems are almost always two-level. Stage 1 (Retrieval) — fast selection of 100–500 candidates from 300k items using ALS or Two-Tower model with vector search (FAISS, Qdrant). Stage 2 (Ranking) — heavy ranker on LightGBM or neural network with cross-features, time, device, and session context. LightFM is a good starting point for medium scale without heavy infrastructure. Our practice shows: moving from single-stage to two-stage yields a 15–25% accuracy improvement with only 20–30 ms additional latency.

Real-Time Serving: Architecture Under Load

Latency SLA — 50–100 ms at thousands of requests per second. Base recommendations precomputed (batch job hourly) → Redis by user_id → <5 ms. Real-time re-ranking via Kafka for events (clicks, cart adds) → update of context features. Feature serving — Redis with TTL (views in 24 hours, last clicked item). At 10k req/s, we deploy Redis Cluster with replication.

A/B testing is the only reliable way to measure improvements. Offline metrics do not always correlate with online. Kohavi et al., 'Online Controlled Experiments at Large Scale' (KDD 2013) — a must-read for the team. Test on 5–10% of traffic, monitor CTR, conversion, revenue per session. One of our client systems after hybridization increased revenue by 18% over a month of A/B.

Recommender System Development Timeline

The stages and typical time frames are in the table below. Costs are calculated individually based on catalog scale and latency requirements.

What's Included in Turnkey Development

1. Data audit — user×item matrix density (typically <0.1%), activity distribution, temporal patterns, cold start statistics.
2. Baseline — 'popular' as a simple threshold that is often hard to beat.
3. Iterative improvement — ALS → content features → two-stage → sequence-aware. Each step with A/B.
4. Serving infrastructure — batch precomputation, Redis, real-time re-ranking, Grafana monitoring.
5. Documentation — model card with metrics, deployment instructions, feature descriptions.
6. Team training — session on interpreting results and model fine-tuning.
7. Support — 1 month post-launch (incident fixes, pipeline tuning).

We are a team with 7+ years of experience in recommender systems, having delivered over 30 projects for e-commerce and media. We guarantee transparent A/B testing and documented metric improvements.

Want to assess the growth potential of your catalog? Contact us for a free data audit. Order recommender system development — first prototype within two weeks.

from implicit.als import AlternatingLeastSquares

model = AlternatingLeastSquares(
    factors=64,
    regularization=0.05,
    iterations=15,
    use_gpu=True
)
model.fit(user_item_matrix)

More about the mathematics of recommender systems — in specialized literature.