Chapter 6: Personalized Recommendations: Algorithms that Adapt to User Behaviour

Collaborative Filtering Foundations  
Use user–item interaction matrices to identify patterns of co-consumption and suggest items that similar users purchased. 

Collaborative filtering leverages patterns in historical user–item interactions such as views, ratings, or purchases to recommend latest items based on the behaviour of similar users. The core assumption is that users who agreed in the past will continue to share preferences: if User A and User B both bought Items X and Y, then an item purchased by A but not yet by B (Item Z) is likely relevant to B. 

How It Works: 

1. Interaction Matrix Construction: Assemble a sparse matrix where rows represent users, columns represent items, and entries capture interaction strength (binary clicks, rating scores, purchase counts). 

1. Similarity Computation: Compute user–user or item–item similarities using cosine similarity or Pearson correlation on the interaction vectors. 

1. Neighbour Selection: For a target user, identify k most similar users (neighbours) and aggregate their high-rated or frequently purchased items. 

1. Recommendation Generation: Score candidate items by weighted neighbour preferences and present the top-ranked suggestions.  

By relying solely on collaborative signals, this method requires minimal domain knowledge and adapts naturally as interaction data grows. However, it can struggle with cold-start users or items lacking history. 

Key Characteristics 

• Data-Driven: Learns directly from interaction patterns without explicit item features. 

• Scalable: Matrix factorization techniques (SVD, ALS) reduce dimensionality for large-scale systems. 

• Adaptive: Improves as new user interactions accumulate. 

• Sparse Sensitivity: Performance degrades when interaction data is extremely sparse. 

• Neighbourhood-Based: Balances local (k-nearest neighbours) and global (factorization) approaches for accuracy.