New Collection Launch Recs
“For each customer, which items from the new collection will they purchase in the next 30 days?”
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
For each customer, which items from the new collection will they purchase in the next 30 days?
New products have zero interaction history — collaborative filtering fails completely. Fashion retailers launch 200-500 new items per season and have no data on who will buy them. Traditional approaches default to showing new items to everyone or using simple attribute matching. First-week sell-through on new collections averages 8-12% when it should be 25-35%. For a fashion retailer doing $2B annually, improving new-collection sell-through by 10 points is worth $40-60M per season.
Quick answer
New collection launch recommendations solve the cold-start problem by using the relational graph to recommend products with zero interaction history. Graph-based models transfer knowledge from similar products (same brand, category, style, price point) and customer purchase patterns to target the right buyers from day one, improving first-week sell-through by 10-15 points.
Approaches compared
4 ways to solve this problem
1. Show to Everyone (Mass Merchandising)
Feature new products on the homepage and in marketing emails for all customers. The default launch strategy for most retailers.
Best for
High-profile launches with broad appeal (flagship products, celebrity collaborations) where awareness is the goal.
Watch out for
First-week sell-through averages 8-12% because most customers seeing the new product are not the right audience. Leads to heavy markdowns later in the season to clear unsold inventory.
2. Attribute-Based Matching
Match new products to customers based on shared attributes: same category, similar price range, same brand. If a customer bought a linen blazer last fall, show them linen shirts this spring.
Best for
Catalogs with rich, well-structured product metadata and clear attribute-to-preference mapping.
Watch out for
Attribute matching is brittle. A 'linen shirt' at $129 and a 'linen shirt' at $29 have the same attributes but completely different customer profiles. The model cannot distinguish premium from value positioning without price sensitivity data.
3. Transfer Learning from Similar Products
Find the most similar existing products (based on embeddings or attributes) and transfer their customer affinity scores to the new product. A data science approach for cold-start items.
Best for
Teams with existing product embeddings and the infrastructure to compute similarity at launch time.
Watch out for
Similarity is defined in embedding space, which often over-weights visual or textual similarity. Two products that look similar may have very different customer profiles. Also requires maintaining and updating product embeddings continuously.
4. KumoRFM (Graph Neural Networks on Relational Data)
Predicts which customers will buy new collection items by traversing the relational graph: same brand, similar category, matching price point, cross-customer purchase patterns. The WHERE clause filters predictions to the new collection. No interaction history required because the graph transfers signals through product relationships.
Best for
Fashion and lifestyle retailers launching 200+ new items per season who need personalized targeting from day one.
Watch out for
Requires product metadata (brand, category, price, attributes) for new items. If new products are launched without structured metadata, the graph has fewer relationship edges to transfer signals through.
Key metric: Graph-based cold-start targeting improves first-week sell-through by 10-15 points, translating to $40-60M in incremental seasonal revenue for fashion retailers and 20-30% reduction in markdowns.
Why relational data changes the answer
Product P601 (Linen Shirt, Spring 2025) has zero purchase history. Collaborative filtering returns nothing. But the relational graph is rich with transferable signals. P601 shares a brand with P301 (Linen Blazer, Fall 2024), which was purchased by Sarah Chen and similar premium-segment customers. The Tops category has 35% share of Sarah's purchase history. The $129 price point falls within her typical spend range ($95-$210). And 68% of customers in her graph neighborhood who bought last season's linen collection have been flagged as targets for the new one.
The critical insight is that cold-start recommendations require multi-hop graph traversal: from the new product to similar products (via brand and category), from those products to customers who purchased them, and from those customers to the new product's predicted audience. This three-hop path is invisible to any model that starts from the product's own interaction data (which does not exist). Graph neural networks traverse these paths automatically, transferring purchase affinity through product relationships. For fashion retailers doing $2B annually, a 10-15 point improvement in first-week sell-through on new collections translates to $40-60M in incremental seasonal revenue and drastically reduced end-of-season markdowns.
Recommending new products without graph signals is like a movie theater marketing a sequel to random moviegoers. Graph-based cold-start targeting is like marketing the sequel specifically to people who watched the original, bought the director's other films, and attend the same film festivals as other fans. The new movie has no audience data yet, but the relational graph knows exactly who will want to see it.
How KumoRFM solves this
Relational intelligence for true personalization
Kumo solves the cold-start problem by using the relational graph — similar products, same brand, category affinity, style attributes, and cross-customer purchase patterns — to recommend items with zero interaction data. The model learns that customers who bought last season's linen collection from the same brand, in similar colorways, at similar price points, are the best targets for the new Spring 2025 line. No interaction history required.
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 | name | segment | signup_date |
|---|---|---|---|
| C001 | Sarah Chen | premium | 2023-06-15 |
| C002 | James Wilson | standard | 2024-02-10 |
| C003 | Aisha Patel | premium | 2022-08-30 |
PURCHASES
| purchase_id | customer_id | product_id | amount | timestamp |
|---|---|---|---|---|
| PUR201 | C001 | P301 | 189.00 | 2024-09-15 |
| PUR202 | C001 | P305 | 145.00 | 2024-10-02 |
| PUR203 | C003 | P302 | 210.00 | 2024-09-28 |
PRODUCTS
| product_id | name | category | collection | price |
|---|---|---|---|---|
| P301 | Linen Blazer (Fall 2024) | Outerwear | Fall 2024 | 189.00 |
| P601 | Linen Shirt (Spring 2025) | Tops | Spring 2025 | 129.00 |
| P602 | Wide Leg Trouser (Spring 2025) | Bottoms | Spring 2025 | 165.00 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT LIST_DISTINCT( PURCHASES.PRODUCT_ID WHERE PRODUCTS.COLLECTION = "Spring 2025", 0, 30, days ) FOR EACH CUSTOMERS.CUSTOMER_ID
Prediction output
Every entity gets a score, updated continuously
| CUSTOMER_ID | CLASS | SCORE | TIMESTAMP |
|---|---|---|---|
| C001 | P601 | 0.87 | 2025-03-12 |
| C001 | P602 | 0.79 | 2025-03-12 |
| C003 | P601 | 0.82 | 2025-03-12 |
Understand why
Every prediction includes feature attributions — no black boxes
Customer C001 (Sarah Chen, premium segment)
Predicted: Will purchase P601 (Linen Shirt, Spring 2025) — score 0.87
Top contributing features
Same-brand linen purchases
2 linen items from same brand in Fall 2024
34% attribution
Category affinity (Tops)
35% of purchases are Tops
24% attribution
Price range match
$129 within typical spend range ($95-$210)
19% attribution
Graph neighbors (linen buyers)
68% of similar customers targeted
15% attribution
Seasonal purchase pattern
Buys new collections within 2 weeks of launch
8% 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 new collection launch recs
What is the cold-start problem in product recommendations?
The cold-start problem occurs when a new product has zero interaction data (no views, clicks, or purchases). Collaborative filtering fails completely because it relies on interaction history. Graph-based models solve this by transferring purchase signals through product relationships: same brand, similar category, comparable price, shared attributes with popular products.
How quickly can graph recommendations start working for new products?
From the moment the product is added to the catalog with basic metadata (brand, category, price, attributes). No interaction history is needed. As purchases accumulate, the model blends graph-transferred signals with observed behavior, continuously improving precision.
What sell-through improvement can you expect for new collections?
Graph-based targeting improves first-week sell-through by 10-15 percentage points (from 8-12% to 20-25%). This reduces end-of-season markdown rates by 20-30% because less inventory needs clearance pricing. For fashion retailers, the combined revenue lift and margin preservation is worth $40-60M per season.
Does cold-start recommendation work for completely new categories?
It works best when the new product has at least some relational connections (brand, attributes, price range) to existing products. A completely new category with a new brand and no comparable products has fewer graph connections to transfer signals through. In that case, attribute-based matching is a reasonable starting point until interaction data accumulates.
Bottom line: 10-15 point improvement in first-week sell-through for new collections. For fashion retailers, this translates to $40-60M in incremental seasonal revenue and drastically reduced markdowns.
Related use cases
Explore more personalization 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.
