Cold-Start

The core of the cold-start problem is the absence of interaction history. For a new user, there is no clickstream, purchase record, or rating to infer preferences. For a new item, there are no consumption patterns to identify its target audience. This violates the fundamental assumption of collaborative filtering that past behavior predicts future interests. The system must operate with zero-shot conditions, relying entirely on auxiliary data or exploration heuristics until a minimal signal threshold is reached.

The problem manifests in two distinct forms:

• User Cold-Start: A new user joins the platform with no profile or behavioral history. The system cannot personalize content, leading to a generic, one-size-fits-all experience that often results in poor engagement and early churn.
• Item Cold-Start: A new product, article, or piece of content is added to the catalog. Without interaction data, it remains invisible to collaborative filtering algorithms, creating a popularity bias where only established items are recommended, stifling catalog exploration and inventory turnover.

To bootstrap recommendations, cold-start systems depend on side information—auxiliary metadata that exists independently of interaction history. For users, this includes demographic attributes, sign-up source, device type, or responses to an onboarding questionnaire. For items, this encompasses product descriptions, categories, brand, price, and visual features. This metadata is encoded into content-based feature vectors, allowing the system to compute similarity between new entities and existing ones using purely attribute-based signals until behavioral data accumulates.

Cold-start forces the system into a pure exploration phase. Without prior knowledge, the agent must actively probe the action space to gather reward signals. This is where contextual multi-armed bandits become critical—algorithms like Thompson Sampling or LinUCB can use side information as context features to make informed exploration decisions. The system strategically selects actions to maximize information gain, rapidly building a preference model while minimizing the regret incurred during this learning period. An exploration budget is often defined to cap business risk.

A warm-start strategy mitigates cold-start by initializing models with prior knowledge before online learning begins. Techniques include:

• Transfer Learning: Pre-training embeddings on a related domain or older dataset and fine-tuning on the new entity's sparse data.
• Heuristic Priors: Using business rules or population averages as the initial reward estimate in a Bayesian bandit.
• Lookalike Modeling: Identifying existing users with similar side-information profiles and seeding the new user's representation with their aggregated preferences. This transforms a zero-data problem into a low-data problem, dramatically accelerating time-to-value.

Evaluating cold-start strategies is uniquely challenging because standard off-policy evaluation methods like Inverse Propensity Scoring (IPS) require sufficient data overlap. For new entities, there is no historical logging policy to correct against. Practitioners rely on:

• Simulation: Using hold-out data from mature users to artificially create cold-start scenarios.
• Shadow Deployment: Running the cold-start model in parallel on live traffic, logging predictions without affecting the user experience, and comparing against a baseline over time.
• Online A/B Tests: The gold standard, measuring real-world metrics like click-through rate and session length for new users assigned to different cold-start strategies.

The direct countermeasure to cold-start, involving the initialization of a model with prior knowledge or pre-trained parameters from historical data. Instead of starting from zero, a warm-started bandit uses aggregated population priors or embeddings from a pre-trained neural network. This dramatically reduces the time-to-first-value for new users or items, allowing the system to make educated guesses rather than random selections during the initial interaction phase.

The fundamental dilemma amplified by cold-start scenarios. With no prior data, the system must heavily explore to gather initial signal, sacrificing immediate reward. Key strategies include:

• Epsilon-Greedy: Forces random exploration at a fixed rate
• Upper Confidence Bound (UCB): Adds an uncertainty bonus to unproven actions
• Thompson Sampling: Samples from posterior probability distributions The optimal balance shifts from aggressive exploration during cold-start toward exploitation as confidence builds.

The numerical representation of a new user or item that serves as the input to a contextual bandit during cold-start. Effective vectors leverage side information—attributes available before any interaction history exists:

• User side information: Device type, referral source, geolocation, time of day
• Item side information: Category, brand, price, color, description embeddings
• Contextual features: Season, inventory level, session device These features allow the model to generalize from similar entities, bridging the cold-start gap.

A Bayesian probabilistic algorithm particularly effective for cold-start mitigation. It maintains a posterior probability distribution over each action's expected reward. During cold-start, these distributions are wide and uncertain, naturally encouraging exploration. As data accumulates, distributions narrow, shifting behavior toward exploitation. Key advantage: uncertainty is quantified explicitly, so new items with high-variance posteriors are automatically explored without manual tuning of exploration parameters.

Any attribute or metadata available before the first interaction that can inform initial predictions. Side information is the primary weapon against cold-start. Examples include:

• Demographic data: Age bracket, inferred income, loyalty tier
• Content metadata: Product descriptions, image embeddings, specifications
• Acquisition channel: Organic search, paid social, email campaign
• Third-party enrichment: Credit scores, household composition Without side information, the system collapses to a pure exploration problem with no generalization capability.

The constrained learning signal where only the reward for the chosen action is observed, leaving counterfactual outcomes unknown. This partial observability makes cold-start especially challenging—the system cannot learn from what it didn't try. Mitigation approaches include:

• Inverse Propensity Scoring (IPS): Re-weights observed rewards to debias logged data
• Doubly Robust Estimation: Combines IPS with a direct reward model
• Counterfactual Evaluation: Estimates policy performance offline before deployment These techniques enable learning from historical exploration data to warm-start new models.