Content-Based Filtering

The process of transforming an item's intrinsic metadata into a structured feature vector. This vector serves as the mathematical representation of the item in a high-dimensional space, enabling direct comparison.

• TF-IDF Vectorization: A classical technique that weighs term importance based on frequency within a document and rarity across the corpus, commonly used for textual item descriptions.
• One-Hot Encoding: Converts categorical attributes like 'brand' or 'color' into binary columns, creating a sparse feature representation.
• Embedding Layers: Deep learning models learn dense, low-dimensional representations for categorical features, capturing latent semantic relationships between attribute values.

A user profile is built by aggregating the feature vectors of items the user has previously rated, purchased, or viewed. This creates a weighted centroid representing the user's taste in the item attribute space.

• Explicit Aggregation: The user profile is a direct average of the vectors of liked items, weighted by the explicit rating score.
• Rocchio's Algorithm: A relevance feedback method that modifies the user profile vector by moving it closer to relevant item vectors and away from non-relevant ones.
• Implicit Weighting: Behavioral signals like dwell time or purchase frequency are used to weight the contribution of each item's vector to the user profile without requiring explicit ratings.

Recommendations are generated by computing the similarity between the user profile vector and the vectors of all candidate items. The items with the highest similarity scores are surfaced to the user.

• Cosine Similarity: The most common metric, measuring the cosine of the angle between the user profile vector and an item vector. It is magnitude-invariant, focusing purely on directional alignment.
• Euclidean Distance: Measures the straight-line distance between two points in the vector space, suitable when absolute feature magnitudes are critical.
• Dot Product: A simple, computationally efficient similarity measure that also accounts for vector magnitudes, often used as the final scoring layer in neural networks.

The performance of a content-based system is critically dependent on the quality and discriminative power of the chosen item attributes. Irrelevant or noisy features degrade recommendation accuracy.

• Structured Metadata: Directly using database fields like price, weight, category, and release_date as features.
• Unstructured Text Processing: Applying NLP techniques like tokenization, stemming, and named entity recognition to convert product descriptions or reviews into feature vectors.
• Dimensionality Reduction: Techniques like PCA or autoencoders are used to compress sparse, high-dimensional feature vectors into a dense, lower-dimensional space, removing noise and reducing computational load.

Content-based filtering directly solves the item cold start because a new item can be recommended as soon as its metadata is ingested and vectorized, without waiting for any user interactions.

• Zero-Interaction Recommendation: A new product is immediately eligible for recommendation to users whose profiles match its attributes, bypassing the need for an interaction history.
• Side Information Dependency: The system's ability to handle new items is entirely dependent on the availability and quality of side information at the time of item ingestion.
• User Cold Start Limitation: This method does not inherently solve the user cold start unless combined with an explicit onboarding survey to build an initial user profile from stated preferences.

A fundamental limitation where the system only recommends items highly similar to the user's established profile, creating a 'filter bubble' and failing to suggest novel or diverse items.

• Lack of Serendipity: The engine cannot recommend an item from a completely different category that the user might unexpectedly like, as it lacks a mechanism for cross-domain discovery.
• Feature Saturation: The user receives diminishing returns from recommendations that are overly similar to their past consumption, leading to user fatigue.
• Mitigation via Diversity Injection: Post-processing steps can re-rank the top-N similar items to introduce controlled diversity, penalizing items that are too similar to each other even if they match the user profile.