AI Engine
Understanding the machine learning behind Smart Recommendations
Last updated: 2025-01-16
The Smart Recommendations AI Engine uses advanced machine learning to analyze your store's data and generate highly relevant product suggestions. This guide explains how the algorithms work and how to optimize them for your store.
Data Inputs
The AI engine processes multiple data sources to generate recommendations:
Data Sources
- Order history — What customers purchase together
- Browse behavior — Products viewed, time on page, scroll depth
- Search queries — What customers are looking for
- Cart additions/removals — Purchase intent signals
- Product attributes — Categories, tags, descriptions
- Inventory levels — Stock availability
- Pricing data — Price points and margins
- Customer segments — VIP, new, returning, etc.
Recommendation Algorithms
Frequently Bought Together
Identifies products that are commonly purchased in the same order:
{
"algorithm": "frequently_bought_together",
"method": "association_rules",
"parameters": {
"min_support": 0.01,
"min_confidence": 0.3,
"lookback_days": 90,
"min_co_occurrences": 5
},
"output": {
"product_pairs": [
{
"product_a": "moisturizer",
"product_b": "cleanser",
"confidence": 0.65,
"lift": 2.3
}
]
}
}Collaborative Filtering
Recommends products based on similar customers' behavior:
{
"algorithm": "collaborative_filtering",
"method": "matrix_factorization",
"parameters": {
"latent_factors": 50,
"regularization": 0.1,
"learning_rate": 0.01,
"iterations": 20
},
"similarity_metric": "cosine",
"min_similarity_threshold": 0.3
}Content-Based Filtering
Finds similar products based on attributes and descriptions:
{
"algorithm": "content_based",
"features": {
"title_embedding": { "weight": 0.3, "model": "tfidf" },
"description_embedding": { "weight": 0.2, "model": "tfidf" },
"category_match": { "weight": 0.25 },
"tag_overlap": { "weight": 0.15 },
"price_range": { "weight": 0.1 }
},
"similarity_threshold": 0.4
}Real-Time Scoring
Combines pre-computed models with real-time signals:
{
"realtime_signals": {
"current_cart": {
"weight": 0.4,
"boost_related_products": true
},
"session_views": {
"weight": 0.25,
"decay_factor": 0.8
},
"search_query": {
"weight": 0.2,
"match_type": "semantic"
},
"time_context": {
"weight": 0.15,
"factors": ["day_of_week", "time_of_day", "season"]
}
}
}Model Training
Models are trained and updated automatically:
Training Schedule
- Full model retrain — Weekly on Sundays
- Incremental updates — Daily at midnight
- Real-time learning — Immediate feedback incorporation
- A/B model testing — Automatic model comparison
{
"training": {
"schedule": {
"full_retrain": "0 3 * * 0",
"incremental": "0 1 * * *"
},
"data_requirements": {
"min_orders": 100,
"min_products": 50,
"min_customers": 200
},
"validation": {
"holdout_percentage": 0.2,
"metrics": ["precision@k", "recall@k", "ndcg"]
}
}
}Personalization
The engine personalizes recommendations for individual customers:
{
"personalization": {
"enabled": true,
"user_profile": {
"build_from": [
"purchase_history",
"browse_history",
"wishlist",
"reviews_written"
],
"profile_decay": 0.95,
"max_profile_items": 100
},
"cold_start": {
"strategy": "popularity_based",
"use_session_data": true,
"segment_defaults": true
}
}
}Cold Start Problem
For new visitors with no history, the engine uses popularity-based recommendations and session behavior. After 3-5 interactions, personalization kicks in.
Performance Optimization
The AI engine is optimized for speed:
Performance Features
- Pre-computed recommendations cached at edge
- Real-time scoring < 50ms latency
- Batch pre-computation during low-traffic hours
- Fallback recommendations for cache misses
- Adaptive computation based on traffic
Monitoring & Quality
Monitor recommendation quality:
{
"monitoring": {
"metrics": {
"precision_at_5": "Relevant items in top 5 / 5",
"recall_at_10": "Relevant items found / Total relevant",
"coverage": "Products recommended / Total products",
"diversity": "Uniqueness of recommendations",
"novelty": "Long-tail vs popular items"
},
"alerts": {
"ctr_drop": { "threshold": -20, "window": "7d" },
"coverage_drop": { "threshold": -10, "window": "1d" }
}
}
}Tuning Parameters
Algorithm Tuning
- Increase diversity — Lower similarity threshold, add random exploration
- Favor new products — Add recency boost to scoring
- Improve relevance — Increase model training frequency
- Handle sparse data — Lower min_support thresholds
Related Topics
Need more help? Contact support