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:

Frequently Bought Together
{
  "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:

Collaborative Filtering
{
  "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:

Content-Based Filtering
{
  "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:

Real-Time Scoring
{
  "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 Configuration
{
  "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 Config
{
  "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:

Quality Metrics
{
  "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

Was this page helpful?

Need more help? Contact support