Active Learning

The model queries the user about items for which its prediction confidence is lowest. This focuses the onboarding survey on the most ambiguous data points.

• Least Confidence: Asks about the item with the lowest predicted probability for its most likely class.
• Margin Sampling: Queries the item where the difference between the top two predicted probabilities is smallest.
• Entropy Sampling: Selects the item with the highest entropy across all possible class predictions.

This avoids wasting questions on items the model already understands.

An ensemble of multiple diverse models is trained on the existing data. The active learner selects data points where the committee disagrees most about the prediction.

• Vote Entropy: Measures the disagreement in hard class votes across the committee.
• Kullback-Leibler Divergence: Calculates the average divergence between each committee member's probability distribution and the consensus.

This is highly effective for identifying items that expose fundamental model uncertainty.

Selects the data point that would cause the largest change to the current model's parameters if its label were known. The goal is to query instances that have the highest potential impact on learning.

• Uses gradient-based metrics to estimate the magnitude of parameter update.
• Particularly suited for gradient-descent-based models like neural networks.
• Directly optimizes for model improvement rather than just resolving uncertainty.

This strategy is computationally intensive but highly efficient for data selection.

Combines informativeness with representativeness to avoid querying outliers. An instance must not only be uncertain but also located in a dense region of the input space.

• Prevents the oracle from labeling anomalous or noisy data points.
• Uses clustering algorithms to measure the density around a candidate instance.
• Ensures the queried preference is relevant to a broad segment of the user population.

This is critical for building a robust initial user profile that generalizes well.

Selects a batch of diverse items to present in a single onboarding step, maximizing the breadth of information gained from one round of feedback.

• Avoids querying redundant items that share similar features.
• Uses determinantal point processes (DPPs) to model both item quality and pairwise dissimilarity.
• Ensures the onboarding survey covers a wide range of product categories and attributes quickly.

This accelerates the preference elicitation process for the new user.

Integrates active learning with content-based filtering to bootstrap the process. The system uses item metadata to generate an initial query set before any user feedback is collected.

• Step 1: Uses a knowledge graph to present a diverse seed set of popular items.
• Step 2: Applies uncertainty sampling on the user's sparse initial feedback.
• Step 3: Switches to density-weighted selection as the profile matures.

This staged approach provides immediate value while the model learns.

The fundamental dilemma that Active Learning helps navigate. The system must balance exploiting known high-reward items against exploring unknown items to gather new data.

• Contextual Bandit: An algorithm that uses side information to intelligently explore for cold-start users.
• Thompson Sampling: A probabilistic method that selects actions based on their probability of being optimal, naturally balancing uncertainty for new items.

A strategy that mitigates cold starts by analyzing the intrinsic attributes of items and matching them to a user's explicitly stated preferences.

• How it works: If a user selects 'sci-fi' during an Active Learning onboarding, the system recommends items tagged with 'sci-fi'.
• Key Enabler: Rich item metadata and a user profile built from the Active Learning survey.

The process of converting a user's preferences into a dense vector representation. Active Learning provides the initial data points to anchor this vector.

• Cosine Similarity: Used to find the nearest existing items to a new user's initial preference vector.
• Pre-Trained Embeddings: Provide a rich semantic initialization, allowing the system to understand that a preference for 'space operas' is related to 'sci-fi'.

A dynamic data collection strategy that builds a user profile over time. It contrasts with a lengthy upfront survey by asking non-intrusive questions at contextually relevant moments.

• Active Learning Synergy: The model can intelligently choose the next best question to ask during a user's journey, not just at sign-up.
• Goal: Continuously refine the user profile without causing survey fatigue.

An architecture that combines collaborative and content-based filtering. Active Learning bridges the gap between the two.

• How it works: Content-based filtering bootstraps recommendations using Active Learning responses, while collaborative filtering takes over once sufficient interaction data exists.
• Key Benefit: Leverages the strengths of each technique to provide seamless personalization from the very first session.