Building an AI Cross-Sell Recommendation System for Sales

Building an AI Cross-Sell Recommendation System for Sales

We were approached by an e-commerce project with 50,000 SKUs: the rule "printer → cartridge" yielded only 8% conversion. They needed a machine learning-based recommendation system that finds non-obvious pairs and adapts to each customer's behavior. We implemented a hybrid system for e-commerce and retail: association rules for basic bundles + gradient boosting for personalization. Result: an 18% increase in average order value in the first month, providing a significant revenue boost. Here’s how it works.

Problems We Solve

Static rules do not capture hidden patterns. For example, customers who bought a child car seat have a 22% probability of buying a trunk organizer—even though the items are from different categories. ML finds such patterns through association rules (Lift > 2.0). Personalization via gradient boosting takes into account customer history: if the user already bought an organizer, the recommendation shifts to another item.

Another common problem is goods with a short consumption cycle (consumables, subscriptions). Here, predicting the next purchase is required. A model based on interval distribution outputs items the customer is about to reorder. This way, we increase repeat sales without promotions.

The third problem is the cold start for new products. If a product has no purchase history, Apriori rules won't work. The solution is to use content-based features (category, price, brand) and collaborative filtering through embeddings. For this, we enrich the gradient boosting model with product features.

How We Build the CrossSellRecommender Model

Market Basket Analysis + Personalization

Hybrid architecture: first Apriori gives raw rules, then gradient boosting ranks candidates for a specific user. This combination provides a conversion lift 3 times higher than a rule-based approach.

import pandas as pd
import numpy as np
from mlxtend.frequent_patterns import apriori, association_rules
from mlxtend.preprocessing import TransactionEncoder
from sklearn.ensemble import GradientBoostingClassifier

class CrossSellRecommender:
    def __init__(self, min_support: float = 0.01, min_confidence: float = 0.1):
        self.min_support = min_support
        self.min_confidence = min_confidence
        self.rules = None
        self.cross_sell_map = {}
        self.personalization_model = None

    def fit_association_rules(self, orders_df: pd.DataFrame,
                               order_col: str = "order_id",
                               item_col: str = "item_id"):
        """Find association rules via Apriori"""
        # Transaction baskets
        baskets = orders_df.groupby(order_col)[item_col].apply(list).tolist()

        te = TransactionEncoder()
        te_array = te.fit_transform(baskets)
        df_encoded = pd.DataFrame(te_array, columns=te.columns_)

        # Frequent itemsets
        frequent_itemsets = apriori(
            df_encoded,
            min_support=self.min_support,
            use_colnames=True,
            max_len=3
        )

        # Association rules
        self.rules = association_rules(
            frequent_itemsets,
            metric="lift",
            min_threshold=1.2
        )
        self.rules = self.rules[self.rules['confidence'] >= self.min_confidence]
        self.rules = self.rules.sort_values('lift', ascending=False)

        # Mapping: item → list of recommendations with metrics
        for _, rule in self.rules.iterrows():
            for antecedent in rule['antecedents']:
                if antecedent not in self.cross_sell_map:
                    self.cross_sell_map[antecedent] = []
                for consequent in rule['consequents']:
                    if antecedent != consequent:
                        self.cross_sell_map[antecedent].append({
                            'item_id': consequent,
                            'confidence': rule['confidence'],
                            'lift': rule['lift'],
                            'support': rule['support']
                        })

        # Sort by lift
        for item in self.cross_sell_map:
            self.cross_sell_map[item].sort(key=lambda x: x['lift'], reverse=True)

    def recommend_cross_sell(self, cart_items: list[str],
                              user_history: list[str] = None,
                              n: int = 5) -> list[dict]:
        """Cross-sell for current cart"""
        candidates = {}

        for item_id in cart_items:
            related = self.cross_sell_map.get(item_id, [])
            for rec in related:
                rec_id = rec['item_id']

                # Skip if already in cart or in history
                if rec_id in cart_items:
                    continue
                if user_history and rec_id in user_history:
                    continue

                if rec_id not in candidates:
                    candidates[rec_id] = {'score': 0, 'triggers': []}

                candidates[rec_id]['score'] += rec['lift']
                candidates[rec_id]['triggers'].append(item_id)

        # Normalization
        if not candidates:
            return []

        sorted_candidates = sorted(
            [{'item_id': k, **v} for k, v in candidates.items()],
            key=lambda x: x['score'],
            reverse=True
        )

        return sorted_candidates[:n]

    def get_complementary_categories(self, category: str) -> list[str]:
        """Complementary categories"""
        category_rules = self.rules[
            self.rules['antecedents'].apply(lambda x: category in str(x))
        ]['consequents'].apply(lambda x: list(x)).explode().value_counts()

        return category_rules.head(5).index.tolist()

Temporal Patterns: Next Purchase

class NextPurchasePredictor:
    """Predict next purchase based on history"""

    def predict_next_items(self, user_id: str,
                            purchase_history: list[dict],
                            catalog_features: pd.DataFrame) -> list[tuple]:
        """
        purchase_history: [{item_id, date, quantity, category}]
        Returns: [(item_id, probability)]
        """
        if len(purchase_history) < 3:
            return []

        # Repeat purchase patterns
        item_intervals = {}
        for i in range(1, len(purchase_history)):
            item = purchase_history[i]['item_id']
            prev_same = [h for h in purchase_history[:i] if h['item_id'] == item]
            if prev_same:
                days_between = (
                    pd.to_datetime(purchase_history[i]['date']) -
                    pd.to_datetime(prev_same[-1]['date'])
                ).days
                if item not in item_intervals:
                    item_intervals[item] = []
                item_intervals[item].append(days_between)

        # Predict repeat purchases
        predictions = []
        last_purchase_date = pd.to_datetime(purchase_history[-1]['date'])
        today = pd.Timestamp.now()
        days_since_last = (today - last_purchase_date).days

        for item_id, intervals in item_intervals.items():
            avg_interval = np.mean(intervals)
            std_interval = np.std(intervals) if len(intervals) > 1 else avg_interval * 0.3

            # Probability via normal distribution
            from scipy.stats import norm
            prob = norm.cdf(days_since_last + 7, avg_interval, std_interval + 1)
            prob -= norm.cdf(days_since_last - 7, avg_interval, std_interval + 1)
            prob = min(max(prob, 0), 1)

            if prob > 0.1:
                predictions.append((item_id, prob))

        return sorted(predictions, key=lambda x: x[1], reverse=True)[:10]

How to Train the CrossSellRecommender on Your Data

1. Data preparation. Collect order history with columns: order_id, item_id, date. Minimum number of transactions — 10,000.
2. Run Apriori. Use the fit_association_rules method with min_support=0.01 and min_confidence=0.1. This will give about 1000 rules.
3. Tune parameters. Check the distribution of lift and confidence. If there are too many rules, increase min_support to 0.02 or min_confidence to 0.2.
4. Evaluate quality. Use precision@k and recall@k metrics on a held-out set. Rule-based cross-sell gives an average basket uplift of 15-25%. Combining with personalization (user history) adds another 5-10% to acceptance rate.

How to Measure Recommendation Quality?

Main offline metrics: precision@k, recall@k, lift@k. We use a truncated basket — hide part of items from each transaction and check if they appear in recommendations. Online metrics: acceptance rate, AOV uplift, share of carts with cross-sell. On an A/B test with 50% audience, typical acceptance rate uplift is 10-25%.

Why is the Hybrid Model More Effective Than Rule-Based?

Rule-based (e.g., "bought A → show B") does not scale to thousands of items: rules become sparse and don't account for context. ML finds high-lift associations even for rare pairs. In addition, the hybrid personalizes ranking: the same item is ranked differently for different users. In our projects, the acceptance rate of the hybrid is 3 times higher than that of rules.

Implementation Stages

What's Included in the Result

• Documentation: description of rules, metrics, instructions for model updates.
• Source code: repository with modules CrossSellRecommender, NextPurchasePredictor, configs and tests.
• API endpoints: two methods — recommend_by_cart (based on current cart) and recommend_by_user (based on history).
• Team training: we show how to interpret results and retrain the model.
• Guarantee: we support the A/B test until statistically significant uplift is achieved.

How to Estimate ROI for Your Business

To assess the cross-sell potential for your assortment, contact us. We will conduct a data audit and provide a prototype in 2 weeks with a confidentiality guarantee. Request a consultation — we'll discuss the details of your project and prepare a tailored proposal.

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.