Skip to main content

The Cold Start Item Problem

You've just added exciting new products to your e-commerce store, published groundbreaking articles, or uploaded fresh video content. Naturally, you want users to discover them. But there's a catch-22: most recommendation algorithms heavily rely on past user interactions (clicks, views, purchases) to determine what's relevant. Brand new items, by definition, lack this history. They are "cold starts."

This "cold start item" problem is a major headache. It means valuable new inventory or content can remain virtually invisible, buried beneath established popular items. Users miss out on potentially relevant discoveries, leading to stale experiences and filter bubbles. For businesses, creators, or sellers, it means new investments struggle to gain traction, get feedback, or generate revenue. Manually promoting every new item is unsustainable, and traditional algorithms often fail to give them a fair shot.

The Shaped Approach: Intelligent Exposure for New Items​

Shaped tackles the cold start item problem not as an afterthought, but as an integral part of its relevance platform, using a combination of intelligent strategies designed to give new items a fair chance without sacrificing overall recommendation quality.

How Shaped Surfaces New Items Intelligently:

  1. Attribute-Aware Ranking Models: Shaped's core ranking models (often based on Transformers) go beyond simple ID matching. They deeply analyze the item attributes you provide (text descriptions, categories, tags, brands, etc., defined in your fetch.items query). This means a new item can be effectively scored and ranked based on its inherent characteristics and how they relate to a user's preferences, even before it accumulates significant interaction data. A relevant new item can naturally rank higher than an irrelevant old one.
  2. Intelligent Exploration via Contextual Bandits: Shaped employs sophisticated algorithms like contextual multi-armed bandits to manage the crucial exploration-exploitation trade-off. This isn't just random injection; the bandits intelligently decide when to explore by showing potentially relevant new or less-popular items, learning from the user's feedback (or lack thereof). This allows new items to gather initial interactions and prove their value. You can influence this balance using the exploration_factor parameter.
  3. Optimized Candidate Retrieval: Before ranking even happens, Shaped's internal systems work to retrieve a diverse set of candidate items. This includes mechanisms to ensure a reasonable sample of relevant new items are considered in the first place, preventing them from being immediately filtered out simply due to lack of popularity.

These strategies work together seamlessly within the Shaped platform, providing automated and intelligent handling of new items.

Configuring New Item Exposure with Shaped​

While much of the cold start item handling is automated and learned by the models, you can influence how aggressively Shaped explores new items using configuration and API parameters.

Goal: Ensure new, relevant items get a fair chance to be seen by users, balancing discovery with exploitation of known preferences.

1. Define Your Shaped Model (YAML - Emphasize Features): Crucially, provide rich, descriptive item metadata in your fetch.items query. This empowers the attribute-aware ranking (Strategy 1).

item_cold_start_model.yaml
model:
  name: product_discovery_v2
  inference_config:
    exploration_factor: 0.1 # Example: dedicate ~10% effort to exploration
connectors:
  - type: Dataset
    name: item_metadata
    id: items
  - type: Dataset
    name: user_interactions
    id: interactions
fetch:
  items: |
    SELECT
      item_id,
      title,
      description,  # <-- Crucial for attribute-aware ranking
      category,     # <-- Important feature
      tags,         # <-- Important feature
      brand,
      image_url,
      product_url,
      publish_date  # <-- Useful metadata, potentially for rules/analysis
    FROM items
  events: |
    SELECT user_id, item_id, timestamp AS created_at, event_type FROM interactions
  • Key Point: Rich fields like description, tags, category allow Shaped to understand what a new item is, enabling it to be ranked appropriately even without interaction data.

2. Create the Model & Monitor Training: (Standard process)

shaped create-model --file item_cold_start_model.yaml

shaped view-model --model-name product_discovery_v2 # Wait for ACTIVE

3. Influence Exploration at Query Time (Application Logic): You can adjust the exploration level for specific requests or user segments using the exploration_factor in the rank API call.

  • Step A (Your Backend): Identify the user and context for the recommendation request.

  • Step B (Your Backend): Call Shaped's rank API, optionally overriding the exploration_factor.

Scenario 1: Use default exploration set in the model (or Shaped's internal default)​

const { ShapedClient } = require('@shapedai/shaped'); // Replace with actual SDK import

const apiKey = process.env.SHAPED_API_KEY; // Assumes this is set in environment
const modelName = 'product_discovery_v2';
const userId = 'USER_789';
const itemsLimit = 20;

const shapedClient = new ShapedClient({ apiKey: apiKey });
const responseDefault = await shapedClient.rank({
    modelName: modelName,
    userId: userId,
    limit: itemsLimit,
    returnMetadata: true
});
console.log(`Resulting ${responseDefault.metadata} ranked items for model: ${modelName}, user: ${userId}`);

Scenario 2: Increase exploration for this specific request​

const { ShapedClient } = require('@shapedai/shaped'); // Replace with actual SDK import

const apiKey = process.env.SHAPED_API_KEY; // Assumes this is set in environment
const modelName = 'product_discovery_v2';
const userId = 'USER_789';
const itemsLimit = 20;

const shapedClient = new ShapedClient({ apiKey: apiKey });
const responseDefault = await shapedClient.rank({
    modelName: modelName,
    userId: userId,
    limit: itemsLimit,
    returnMetadata: true
});
console.log(`Resulting ${responseDefault.metadata} ranked items for model: ${modelName}, user: ${userId}`);
  • Step C (Observe & Monitor): Analyze the recommendation lists. With higher exploration, you should expect to see a greater proportion of less common items, including relevant new ones that the bandit algorithms choose to test. Monitor metrics like CTR on new items, overall recommendation diversity, and user engagement to find the optimal exploration_factor for your audience.

4. How Attribute-Awareness Helps: Imagine two new items, A and B.

  • Item A: A new sci-fi novel by a known author whose previous books User X loved.
  • Item B: A new romance novel, a genre User X has never shown interest in.
  • Even with zero interactions for A and B, Shaped's attribute-aware ranking can analyze their features (genre, author, description). It will likely rank Item A much higher for User X than Item B, because its attributes match the user's learned preferences, giving the relevant new item a strong chance from the start. Exploration then helps refine this and discover unexpected hits.

Conclusion: Intelligent Discovery for Every Item​

Ignoring new items until they accumulate data is a recipe for missed opportunities and stale user experiences. The cold start item problem requires intelligent solutions that go beyond manual curation or random guesswork.

Shaped integrates sophisticated strategies directly into its platform: attribute-aware ranking models understand new items based on their content, intelligent bandit algorithms dynamically balance exploration and exploitation, and optimized retrieval ensures new items are considered. By providing rich item metadata and optionally tuning the exploration_factor, you empower Shaped to automatically surface relevant new products or content, fostering discovery, keeping experiences fresh, and giving every item a fair chance to succeed.

Ready to ensure your new items get the visibility they deserve?

Request a demo of Shaped today to discuss how Shaped handles item cold starts. Or, start exploring immediately with our free trial sandbox.