Next Best Action & Product Recommendations

Keystone's Next Best Action (NBA) engine calculates individualized product recommendations and marketing intervention strategies for every customer profile in real time.

1. The Next Best Action Interface

Navigate to Intelligence -> Next Best Action (/intelligence/next-best-action) to inspect algorithmic recommendations across your audience.

Click to enlarge

What the Engine Determines:

• Recommended SKU / Category: The exact product with the highest statistical likelihood of being purchased next.
• Recommended Channel: Whether the user is most responsive to Email, SMS, Web Push, or Paid Social.
• Optimal Send Timing: The predicted hour and day of week when the user is most likely to open and purchase.
• Expected Conversion Lift: Projected revenue increase compared to a generic promotional broadcast.

2. Recommendation Algorithms

Keystone utilizes a multi-model ensemble:

graph TD
    Catalog[(Product Catalog & SKUs)] --> Matrix[Interaction Co-Occurrence Matrix]
    Orders[(Historical Orders)] --> Matrix
    Vectors[(Customer 1536D Vector Embeddings)] --> Similarity[Cosine Similarity Search]

    Matrix --> Model1[Collaborative Filtering (ALS)]
    Matrix --> Model2[Basket Association Rules (Market Basket)]
    Similarity --> Model3[Vector Trajectory Embedding Distance]

    Model1 & Model2 & Model3 --> Ranker[Gradient Boosted Ranking Layer]
    Ranker --> Exclusions[Business Exclusions Filter\nStock Out / Recent Purchase]
    Exclusions --> Output[Top 3 Ranked Products per Customer]

1. Collaborative Filtering (Alternating Least Squares): "Customers who bought items similar to your purchase history also loved these products."
2. Market Basket Association (Frequent Itemset Mining): Computes confidence and lift between co-purchased items (e.g. Customers who purchase Dental Handpieces purchase Replacement Ceramic Bearings within 30 days 42% of the time).
3. Vector Trajectory Embedding Distance: Compares the user's recent browsing trajectory against product vector embeddings in latent space.

3. Real-Time Personalization API

Frontends and headless e-commerce platforms can query recommendations on the fly using Keystone's low-latency personalization endpoint:

GET /api/v1/recommendations/{customer_id}?limit=4
X-API-Key: YOUR_API_KEY

Sample Response:

{
  "customer_id": "cust_88291",
  "recommendations": [
    {
      "product_id": "SKU-9921",
      "title": "High-Speed Ceramic Handpiece",
      "price": 249.99,
      "affinity_score": 0.94,
      "algorithm": "collaborative_vector"
    },
    {
      "product_id": "SKU-1044",
      "title": "Sterilization Cassette Pack",
      "price": 45.00,
      "affinity_score": 0.88,
      "algorithm": "basket_affinity"
    }
  ]
}