The Cold Start Problem in Recommender SystemsOpen in Colab

The Cold Start Problem in Recommender Systems

A recommender system that depends on interaction history fails for any entity with no history. The cold start problem occurs for:

  1. New users: no purchase history, no clicks, no ratings.
  2. New items: a just-listed product, a newly published article, a newly uploaded video.
  3. New contexts: a platform expanding to a new country/language/vertical with few users.

Cold start is not a niche edge case — new users join and new items are listed continuously; the platform must serve them well from day 0 or lose them permanently.

Strategy 1: Content-based fallback

Core idea: when there's no interaction data, use the item's or user's own features instead of learned interaction embeddings.

Item cold start with content embedding

Train a content encoder jointly with the interaction-based item tower:

vcontent(i)=Wproj⋅ϕ(i)\mathbf{v}_{\text{content}}(i) = W_{\text{proj}} \cdot \phi(i)

where ϕ(i)\phi(i) is a feature vector (text embedding of description, image CLIP embedding, category one-hot, price bucket). WprojW_{\text{proj}} is learned to map the content feature space into the shared embedding space where user embeddings live.

At Airbnb, a new listing gets:

  • Text embedding from the listing description (fine-tuned BERT)
  • Image embedding from listing photos (CLIP)
  • Structured features: city, price, amenities, room type

These are concatenated and projected to the user-item embedding space. The listing is immediately retrievable by similarity to historical user embeddings, before any booking occurs.

User cold start with demographic + context

For new users, use available context: device type, referrer URL, country, time of day, initial search query. These form the user's "content features" before any interactions exist.

At Netflix, new users without history are served a taste cluster assignment based on their signup responses (genre preferences) and initial browsing. The cluster's interaction history acts as a proxy user embedding.

Strategy 2: Explicit onboarding

Actively elicit preferences during onboarding. The key is to show items that maximally discriminate the latent preference space — not the most popular items (which everyone likes and provide no signal).

Information gain maximization: select the onboarding slate S\mathcal{S} that maximizes expected entropy reduction:

S∗=arg⁡max⁡SE[H(P(user type))−H(P(user type∣responses to S))]\mathcal{S}^* = \arg\max_{\mathcal{S}} \mathbb{E}\left[H(P(\text{user type})) - H(P(\text{user type} \mid \text{responses to } \mathcal{S}))\right]

In practice, this is approximated by clustering the item catalog into KK "taste clusters" and selecting one representative item per cluster, covering the full diversity of the catalog.

Examples:

  • Spotify's artist onboarding: 5 genres × top 2 artists = 10 quick choices that classify the user into a coarse taste bucket.
  • Tinder's initial swipes: binary feedback on photos learns age/gender/style preferences within 20 interactions.
  • Duolingo's language assessment: 10 questions calibrate skill level and content difficulty.

Strategy 3: Meta-learning (MAML for few-shot users)

Model-Agnostic Meta-Learning (MAML) trains an initialization θ\theta such that a small number of gradient steps on user uu's first KK interactions produces a high-quality personalized model:

Training:

θ∗=arg⁡min⁡θ∑uL ⁣(fθ−α∇L(fθ,Dusupport),  Duquery)\theta^* = \arg\min_\theta \sum_u \mathcal{L}\!\left(f_{\theta - \alpha \nabla \mathcal{L}(f_\theta, D_u^{\text{support}})},\; D_u^{\text{query}}\right)

The inner loop (DusupportD_u^{\text{support}} = first 5 interactions) simulates the cold-start adaptation; the outer loss (DuqueryD_u^{\text{query}} = next 5 interactions) evaluates how well the inner adaptation generalized.

In production, this creates a "warm start" mechanism: the base model is fine-tuned for 1–5 gradient steps on any new user's first interactions, producing a personalized model far earlier than waiting for enough interactions to train from scratch.

Strategy 4: Popularity fallback with context

The simplest cold-start strategy: show the most popular items overall (or within context). Works surprisingly well when:

  • The user's context (country, device, referrer) strongly constrains preferences.
  • The catalog has clear "safe bets" with high broad appeal.

Contextual popularity fallback: rank by popularity within context:

score(i∣context)=P(click∣i,context)≈clicks on item i by users in same contextimpressions on item i in same context\text{score}(i \mid \text{context}) = P(\text{click} \mid i, \text{context}) \approx \frac{\text{clicks on item } i \text{ by users in same context}}{\text{impressions on item } i \text{ in same context}}

At Amazon and eBay, new users see "trending in your category" and "top sellers" — a valid cold-start strategy that's immediately available.

Strategy 5: Hybrid with exploration bonus

Combine the cold-start strategy (content-based or popularity) with an exploration bonus for items that have received few impressions:

score(i)=(1−α)⋅r^(u,i)+α⋅ExplorationBonus(i)\text{score}(i) = (1-\alpha) \cdot \hat{r}(u, i) + \alpha \cdot \text{ExplorationBonus}(i)

where ExplorationBonus(i)∝1ni\text{ExplorationBonus}(i) \propto \frac{1}{\sqrt{n_i}} decreases as item ii accumulates more impressions. New items get high exploration bonus; established items are ranked by predicted relevance.

This is exactly the UCB1 bandit formula applied at the item level, ensuring new items are tested before being ranked purely on content features.

Comparison table

StrategyRequiresLatencyQuality on day 0Quality after N interactions
Content embedding fallbackItem featuresLowGoodTransitions to interaction-based
Explicit onboardingUser answersLowGoodHigh (immediate signal)
MAML few-shotPre-trainingMediumGoodVery good (fast adapt)
Popularity fallbackHistorical logsVery lowAcceptablePoor (not personalized)
Exploration bonusAny base modelLowAcceptableGood (drives data collection)

Related concepts

Concept neighbourhood

How this concept connects — click any node to jump there, or explore the whole concept map.

RelatedReferenced by
The Cold Start Problem in Recomme…Session-Based Recommendat…The Recommendation ProblemDiversity, Cold Start & E…