Product Recommendations
“Which products should this customer see?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which products should this customer see?
E-commerce retailers show the same trending products to every visitor, missing $50-100 per session in personalization uplift. Amazon attributes 35% of revenue to its recommendation engine (McKinsey). Most mid-market retailers lack the engineering resources to build similar systems. Collaborative filtering misses new customers (cold start), content-based filtering ignores purchase context (time of day, season, browsing path), and both fail to capture the rich relational structure between products, categories, brands, and customer segments.
Quick answer
The most effective product recommendation systems go beyond collaborative filtering by connecting customers, products, orders, browsing sessions, and reviews into a relational graph. This captures signals that matrix factorization misses: a customer who bought a DSLR camera and then browsed lens reviews for 2 minutes is a very different buyer than one who glanced at the same lens for 5 seconds. On the RelBench user-item recommendation task, KumoRFM scores 76.71 vs 62.44 for the next-best approach because it encodes these behavioral sequences alongside purchase history.
Approaches compared
4 ways to solve this problem
1. Popularity-based and trending lists
Show the same best-sellers or trending items to every visitor. Zero personalization, zero infrastructure cost.
Best for
Brand-new sites with fewer than 1,000 users and no behavioral data to train on.
Watch out for
Conversion rates plateau fast. Every visitor sees the same items, so you miss the 35% of revenue that Amazon attributes to personalization (McKinsey).
2. Collaborative filtering (matrix factorization, ALS)
Decomposes the user-item interaction matrix to find latent factors. The backbone of early Netflix and Spotify recommenders.
Best for
Retailers with dense purchase histories and a stable product catalog where most items have been purchased hundreds of times.
Watch out for
Cold-start problem is brutal: new users and new products get no signal. Also ignores rich contextual data like browsing duration, time of day, and category browsing paths.
3. Deep learning recommenders (two-tower, transformers)
Neural models that encode users and items into embedding spaces. Can incorporate some side features like category and brand.
Best for
Large e-commerce platforms with dedicated ML teams who can manage embedding pipelines and serving infrastructure.
Watch out for
Still operates on flattened feature vectors. Misses the multi-hop relationships: 'customers who bought this camera also returned that lens within 30 days' requires joining orders, returns, and product tables.
4. KumoRFM (relational foundation model)
Builds a graph connecting customers, products, orders, browsing sessions, reviews, and inventory. Learns purchase probability from multi-hop relational patterns.
Best for
Retailers who want to capture cross-entity signals (browsing-to-purchase paths, co-purchase timing, brand affinity evolution) without building feature pipelines.
Watch out for
Most valuable when you have 3+ connected data tables. A single user-item matrix will not reveal the relational advantage.
Key metric: RelBench user-item recommendation: KumoRFM 76.71 vs next-best 62.44. Amazon attributes 35% of revenue to recommendations (McKinsey).
Why relational data changes the answer
Traditional recommendation engines see a flat matrix of users and items. They know that User A bought Products 1, 3, and 7, and User B bought Products 1, 3, and 9, so they recommend Product 9 to User A. But they cannot see that User A browsed Product 9 for 2 minutes, added it to a wishlist, then abandoned cart because it was out of stock in their preferred color. That signal lives in the browsing_sessions and inventory tables, not in the purchase matrix.
Relational models join these tables into a graph and learn the full behavioral sequence. They capture that customers who buy a DSLR camera purchase a compatible lens 68% of the time within 30 days, but only if they have not returned a lens from that brand before. These multi-hop patterns (camera purchase -> brand affinity -> return history -> lens compatibility) produce recommendations that are both more accurate and more explainable than black-box embedding models.
Recommending products from a flat purchase matrix is like a bookstore clerk who only sees your receipt history. They know you bought three mystery novels, so they suggest a fourth. A relational model is like a clerk who also saw you linger at the true-crime shelf for 10 minutes, noticed you picked up a cookbook and put it back, and remembers that the mystery author you like just released a thriller. Same store, dramatically better recommendation.
How KumoRFM solves this
Relational intelligence built for retail and e-commerce data
Kumo builds a relational graph connecting customers, products, orders, browsing sessions, reviews, and inventory data. The model learns that Customer CU-3012 who recently bought a DSLR camera, then browsed memory cards and camera bags, and has a purchase history skewing toward premium brands, should see a recommended set of compatible accessories ranked by purchase probability. The graph captures that customers who buy this camera model also purchase a specific lens 68% of the time within 30 days.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
CUSTOMERS
| customer_id | segment | lifetime_orders | avg_order_value | joined_date |
|---|---|---|---|---|
| CU-3012 | Premium | 24 | $142 | 2022-06-15 |
| CU-3045 | Standard | 8 | $67 | 2023-11-20 |
| CU-3078 | New | 1 | $38 | 2025-09-10 |
ORDERS
| order_id | customer_id | total | items_count | timestamp |
|---|---|---|---|---|
| ORD-8801 | CU-3012 | $849.99 | 1 | 2025-09-05 |
| ORD-8802 | CU-3012 | $34.99 | 1 | 2025-09-12 |
| ORD-8803 | CU-3045 | $89.50 | 3 | 2025-09-14 |
PRODUCTS
| product_id | name | category | price | brand |
|---|---|---|---|---|
| P-5001 | Canon EOS R6 Mark II | Cameras | $2,499 | Canon |
| P-5002 | SanDisk 128GB SD Card | Accessories | $24.99 | SanDisk |
| P-5003 | Canon RF 50mm f/1.8 Lens | Lenses | $199.99 | Canon |
BROWSING_SESSIONS
| customer_id | product_id | action | duration_sec | timestamp |
|---|---|---|---|---|
| CU-3012 | P-5002 | view | 45 | 2025-09-14 10:22 |
| CU-3012 | P-5003 | view | 120 | 2025-09-14 10:25 |
| CU-3012 | P-5003 | add_to_wishlist | 5 | 2025-09-14 10:27 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(ORDERS.PRODUCT_ID, 0, 14, days) FOR EACH CUSTOMERS.CUSTOMER_ID, PRODUCTS.PRODUCT_ID RANK TOP 5
Prediction output
Every entity gets a score, updated continuously
| CUSTOMER_ID | PRODUCT_ID | PRODUCT_NAME | PURCHASE_PROB | RANK |
|---|---|---|---|---|
| CU-3012 | P-5003 | Canon RF 50mm f/1.8 Lens | 0.82 | 1 |
| CU-3012 | P-5002 | SanDisk 128GB SD Card | 0.74 | 2 |
| CU-3012 | P-5008 | Peak Design Camera Bag | 0.61 | 3 |
Understand why
Every prediction includes feature attributions — no black boxes
Customer CU-3012 (Premium segment)
Predicted: Canon RF 50mm Lens: 82% purchase probability
Top contributing features
Compatible camera purchased recently
Canon EOS R6
32% attribution
Browsing session with wishlist add
120s + saved
26% attribution
Co-purchase pattern (68% of similar buyers)
High affinity
20% attribution
Premium brand preference
85% premium
13% attribution
Purchase recency pattern
Buys within 14d
9% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about product recommendations
How do product recommendation engines handle new customers?
The cold-start problem is the Achilles heel of collaborative filtering. With zero purchase history, matrix factorization has nothing to work with. Relational models solve this by using the signals that do exist: browsing behavior, referral source, device type, geographic location, and time of visit. A new visitor browsing premium camera gear for 3 minutes gets very different recommendations than one clicking through budget phone cases, even if neither has purchased anything yet.
What is the ROI of a product recommendation engine?
Amazon attributes 35% of its revenue to recommendations (McKinsey). For a mid-market e-commerce retailer doing $200M in annual revenue, even a 10% lift in recommendation-driven sales adds $7M. The key metric is incremental AOV (average order value) per session. Well-tuned relational models typically increase AOV by 15-25% and conversion rates by 2-4x compared to popularity-based approaches.
Can recommendation models account for inventory levels?
They should, but most do not. Standard collaborative filtering recommends products regardless of stock status, leading to frustrating out-of-stock experiences. Relational models can join the inventory table directly, so they only surface products that are available in the customer's preferred size, color, or configuration. This eliminates the dead-end recommendation problem and improves conversion on recommended items by 15-20%.
Bottom line: Increase average order value by 15-25% and conversion rate by 2-4x through personalized recommendations, generating $20-50M in incremental annual revenue for a mid-size e-commerce retailer.
Related use cases
Explore more retail & e-commerce use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
