User Cold Start

The defining characteristic is a null vector in the user-item interaction matrix. Standard collaborative filtering models, which rely on finding latent patterns in past clicks, purchases, or ratings, have no data to decompose. This forces the system to fall back on non-personalized global averages, resulting in a regression to the mean where every new user sees the same generic, popular items.

In a trained embedding model, a new user's representation is often initialized to zero or a global mean vector. This places them in a region of the latent space with maximum entropy. From the model's perspective, the user could belong to any preference cluster with equal probability. Algorithms like Thompson Sampling explicitly model this high posterior variance, triggering broad exploration to rapidly reduce uncertainty.

Without behavioral data, the system must pivot to exogenous features to make initial inferences. This includes:

• Demographics: Age, location, device type.
• Acquisition Context: UTM parameters, referring domain, or ad campaign keywords.
• Explicit Declarations: Selections made during an onboarding survey. The quality of cold-start performance is directly proportional to the predictive power of this auxiliary data.

A naive fallback strategy is to recommend the most globally popular items. While this maximizes the probability of a generic click, it creates a feedback loop that entrenches existing hits and prevents the discovery of niche content. This is particularly damaging for long-tail catalogs, as new users never provide the initial signal needed to surface diverse inventory, skewing the entire ecosystem toward a homogenized top 1%.

Before a persistent identity is established, the only usable data is the ephemeral clickstream of the current session. Session-based recommendation models using Recurrent Neural Networks (RNNs) or Transformers treat the sequence of in-session actions as the sole input. A single misclick or idle timeout can inject noise into a very sparse signal, making these early predictions highly brittle and sensitive to context.

The user cold start represents a critical activation gap. The system cannot demonstrate its value proposition until a minimum threshold of data is collected. This delay directly impacts key business metrics like Time-to-First-Transaction and Day-1 Retention. The longer the system takes to learn, the higher the probability of user churn, making the speed of preference elicitation a direct driver of customer lifetime value.

The active process of explicitly gathering a new user's tastes during onboarding. This transforms the cold start from a passive modeling problem into an active data acquisition strategy.

• Onboarding surveys ask users to rate sample items or select interests
• Interactive prompts adapt questions based on prior answers
• Progressive profiling spreads questions across sessions to reduce friction
• Directly constructs an initial user profile vector for immediate matching

Auxiliary data associated with a user beyond interaction history that serves as the primary signal during cold starts. This metadata establishes initial similarity links before behavioral patterns emerge.

• Demographics: age, location, device type
• Acquisition source: referring campaign, search keyword, UTM parameters
• Contextual signals: time of day, session entry point, browser language
• Enables lookalike modeling by matching new users to existing cohorts with similar attributes

A recommendation strategy that mitigates cold starts by matching item attributes to a user's explicitly stated preferences or a profile built from consumed items. No interaction history from other users is required.

• Analyzes intrinsic features: category, brand, price, description embeddings
• Computes cosine similarity between user preference vector and item vectors
• Pairs naturally with SBERT embeddings for semantic matching of textual descriptions
• The primary fallback when collaborative signals are absent

A reinforcement learning algorithm that selects actions based on contextual information about the user or situation. It intelligently explores new items for cold-start users by leveraging side information to guide exploration.

• Balances the exploration-exploitation trade-off using observed context
• Thompson Sampling provides a principled probabilistic approach to action selection
• Each arm pull provides an immediate reward signal to update the model
• Rapidly converges on user preferences without requiring a pre-built profile

A method that generates predictions based solely on the sequence of actions within a user's current anonymous session. It provides immediate personalization without requiring any long-term user profile.

• Uses sequential models like GRUs or Transformers on clickstream data
• Captures short-term intent from the first few interactions
• Ideal for first-visit scenarios where no login or identity exists
• Often combined with implicit feedback signals like dwell time and scroll depth

A machine learning paradigm where a model is trained to learn new tasks quickly from very few examples. In cold start contexts, it enables rapid adaptation to new user preferences without extensive historical data.

• Few-shot learning generalizes from as few as 1-5 initial interactions
• Model-Agnostic Meta-Learning (MAML) finds an initialization that fine-tunes rapidly
• Treats each user as a separate task during meta-training
• Dramatically reduces the number of interactions needed before personalization becomes effective