Cold‑Start Recommendation: Algorithmic Approaches and Strategies

1. Efficient Use of Side Information

Side information refers to any data beyond IDs, such as attribute features (e.g., user gender, age, item category, price) and auxiliary sources like knowledge graphs or domain‑specific data. Existing recommender models already incorporate user/item attributes, but cold‑start scenarios require more effective exploitation of these signals.

1.1 DropoutNet

DropoutNet modifies the training process by adding dropout, preventing the model from over‑relying on ID embeddings and forcing it to learn from content features, thereby improving cold‑start recommendations.

1.2 MetaEmbedding

MetaEmbedding generates a high‑quality initialization for ID embeddings by training a generator on existing items; the generator predicts embeddings for new items, enabling better cold‑start performance.

1.3 Meta Warm‑Up Framework (MWUF)

MWUF learns a meta‑stretch network that transforms cold‑start ID embeddings into a space better suited for deep models, and a meta‑offset network that refines embeddings using representations of all interacted users.

2. Leveraging Knowledge Graphs

Knowledge graphs (e.g., movie or book graphs) provide rich structural knowledge that can supplement scarce interaction data, allowing inference of user interests from related entities such as actors or directors.

3. Cross‑Domain Recommendation

When users or items lack interactions in the target domain but have data in auxiliary domains, cross‑domain recommendation transfers embeddings from source to target domains, mitigating cold‑start effects. Methods like TMCDR improve generalization for such transfers.

4. Multi‑Behavior Recommendation

Even cold‑start users may exhibit other behaviors (clicks, add‑to‑cart, etc.). Exploiting these auxiliary actions can enhance recommendations for the primary target behavior.

5. Efficient Use of Limited Interaction Data

5.1 Twitter Recommendation

This approach trains a classifier on items a user has interacted with and uses it to predict interest in cold‑start items, resembling metric‑based meta‑learning.

5.2 MeLU

MeLU applies MAML to learn a shared initialization for a deep recommender; each cold‑start user fine‑tunes this model with their limited interactions, yielding a personalized model.

5.3 MetaHIN

MetaHIN extends MeLU by incorporating heterogeneous information networks, further boosting cold‑start performance.

6. Explore & Exploit

Methods in this category decide when to explore additional user interests versus exploiting known preferences, and similarly when to promote items; the article references contextual bandit and policy‑gradient approaches for deeper study.

7. Other Cold‑Start Scenarios

7.1 New Recommendation Scenarios

When a platform adds a new domain (e.g., a maternity category on an e‑commerce site) with few interactions, meta‑learning methods like S²meta treat it as a few‑shot learning problem.

7.2 Cold‑Start Marketing Tasks

New marketing campaigns often target a small seed set of users. MetaHeac uses meta‑learning to model relationships among multiple campaigns, enabling broader propagation from limited seeds.

8. Summary

The article emphasizes that solving cold‑start recommendation requires efficient use of side information, limited interaction data, and explore‑exploit strategies, while also acknowledging the importance of product‑level interventions such as onboarding forms, appropriate tagging, and real‑time model updates.