Product Affinity

Problem & Business Impact

The multiplication of telecom offers and services (plans, options, devices, accessories, digital services) creates a choice paradox for customers and merchandising complexity for the operator. Generic recommendation approaches ("best-sellers", "new arrivals") ignore individual preferences and generate low conversion rates (<3%). Customers miss products relevant to them, and the operator misses revenue opportunities.

Product affinity uses machine learning to analyze past purchase behaviors, navigation patterns, and customer similarities to recommend products with highest conversion probability. This approach transforms digital touchpoints (app, website, emails) into personalized discovery engines, increasing both customer satisfaction (better experience) and revenue (superior conversion).

Measured impact:

• 68% increase in recommendation conversion rate (2.8% → 4.7%)
• 22% increase in average basket (devices + accessories)
• Generation of €2.4M in additional annual revenue
• 15% improvement in NPS on digital shopping experience

Data & Key Features

Main data sources

• Transaction history: Product purchases, services, timing, channel
• Web/app navigation: Pages visited, products viewed, time spent, abandons
• Customer profile: Segment, ARPU, active services, demographics
• Product catalog: Categories, features, pricing, availability
• Market trends: Best-sellers by segment, new releases, seasonality

Engineered features

• User-product matrix (interaction history)
• Product embeddings (vector representation of characteristics)
• User embeddings (vector representation of preferences)
• Product-product similarity score (collaborative)
• User-product compatibility score (content-based)
• Temporal context (seasonality, events, promotions)
• Popularity by segment (peer group trends)
• Cart abandonment probability

Models & Methods

Hybrid recommendation system

1. Collaborative Filtering (Matrix Factorization - ALS): Similarity behavior-based recommendations

   • "Customers who bought X also bought Y"
   • Top 10 recommendation precision: 38%
   • Handles popular products and established patterns well
   • Limitation: Cold start (new products/customers)

2. Content-Based Filtering (Neural Networks): Attribute-based recommendations

   • Analysis of product characteristics (price, category, specifications)
   • Matching with customer profile and preferences
   • Solves cold start problem
   • Precision: 32% (lower but complementary)

3. Hybrid Recommender (Ensemble): Optimal approach combination

   • Adaptive weighting by context (new customer → content-based)
   • Integration of business rules (margins, inventory, commercial strategies)
   • Final top 10 precision: 44%
   • Recommendation diversity (avoid bubble effect)

Recommendation strategies by context

• Homepage app/site: Top 5 high-affinity products
• Product page: "Frequently bought with", "Customers also viewed"
• Cart: Cross-sell accessories, upgrades, bundles
• Post-purchase email: Complementary products, consumables
• Store: Advisor suggestions during interaction

Real-time Integration

Personalized recommendation pipeline

Event (page visit, cart add) → Context enrichment (session)
    ↓
Real-time affinity scoring (``<50ms``) → Top N product ranking
    ↓
Filtering (stock, eligibility) → Display personalization
    ↓
Tracking (impression, click, purchase) → Model feedback

Real-time optimizations

• Pre-computing: Offline calculation of embeddings and static affinities
• Feature store: Fast access to customer profile and history
• Caching: Pre-calculated recommendations for common segments
• A/B testing: Continuous algorithm experimentation

Experience personalization

• Presentation order: High-affinity products at top
• Messages: "Recommended for you", "Based on your purchases"
• Visuals: Highlighting of profile-relevant features
• Pricing: Display of personalized promotions if eligible

KPIs & Performance Targets

| Metric | Target | Current | |--------|--------|---------| | Recommendation conversion rate | >4.5% | 2.8% → 4.7% | | Recommendation click rate | >12% | 8.3% → 13.2% | | Average basket | >€180 | €147 → €179 | | Revenue via recommendations | €2M/year | €2.4M | | Top 10 precision (hit rate) | >40% | 44% | | API latency | <50ms p95 | 38ms |

Deployment & Monitoring

Recommendation architecture

1. Batch processing (daily): Model training, embedding calculation
2. Feature store: Customer profile, product catalog synchronization
3. Scoring API: Real-time endpoint for recommendations (<50ms)
4. Frontend integration: App/web SDK for recommendation display
5. Event tracking: Impressions, clicks, cart adds, purchases

Performance monitoring

• Dashboard: Click rate, conversion, revenue by placement
• Diversity analysis (avoid repetitive recommendations)
• Freshness tracking (new product integration)
• Continuous A/B testing of algorithms and presentations
• Weekly retraining with new interactions

Business optimization

• Margin calibration (favor high-margin products if comparable affinity)
• Inventory management (promote overstock products)
• Commercial strategies (product launch, end-of-life)
• Cannibalization analysis (recommendations vs organic sales)

FAQ & Prerequisites

Q: How to handle cold start (new customers/products)? A: For new customers: content-based filtering based on demographic profile and segment. For new products: feature analysis + recommendations to identified early adopters.

Q: How to avoid bubble effect (always similar recommendations)? A: Diversity injection in ranking: 70% pure affinity, 30% exploration (new releases, underrepresented categories). Precision/diversity trade-off calibration.

Q: Are recommendations explainable to the customer? A: Yes, each recommendation includes justification: "Based on your previous purchases", "Popular among customers like you", "Complements your X well". Transparency improves trust.

Q: How to measure incremental impact? A: A/B testing with control group (no personalized recommendations). Measurement of conversion and revenue uplift. Typical impact: +30-50% additional revenue vs baseline.

Q: What are the technical prerequisites? A: Data warehouse with transaction and navigation history (6+ months), centralized product catalog, performant feature store, scalable scoring API (<50ms), frontend SDK for integration, event tracking for feedback.