Cold Start Problem

Cold start occurs when a recommender lacks enough interaction history for a new user, item, market, or context. Pure collaborative filtering cannot infer much from an empty row or column, so systems use content, priors, onboarding, or exploration until behavior arrives.

Serving-time fallbacks

A common serving-time fallback blends content and prior evidence:

$$s(u,i)=\alpha s_{\text{content}}(u,i)+(1-\alpha )s_{\text{prior}}(i).$$

For a new item, $s_{\text{content}}$ might come from metadata or image embeddings; for a new user, it might come from onboarding choices. Hybrid recommenders make this blend explicit.

Worked example

A new user profile $[0.8,0.2,0.0]$ can still score new items from content features and a popularity prior:

Item	Content vector	Content score	Popularity prior	Blend $0.75s_{\text{content}}+0.25s_{\text{prior}}$
0	$[0.9,0.1,0.0]$	0.74	0.30	0.630
1	$[0.1,0.2,0.9]$	0.12	0.70	0.265
2	$[0.5,0.5,0.0]$	0.50	0.40	0.475

The rank is item 0, item 2, then item 1. The first new item wins because content aligns with the user profile, even before collaborative data exists. Exploration versus exploitation controls how aggressively cold items are exposed for learning.

Caveats

Cold-start fixes can become popularity defaults that bury new inventory. Onboarding adds friction and can collect noisy stated preferences. Content features may be missing or weak, so track cold-start segments separately in evaluation.

References

• Adomavicius and Tuzhilin, 2005, Toward the Next Generation of Recommender Systems
• Koren, Bell, and Volinsky, 2009, Matrix Factorization Techniques for Recommender Systems

Nav

Section — Recommendation Systems and Personalization

← Bayesian Personalized Ranking Hybrid Recommenders →