Collaborative Filtering Course

Lesson 1 • Taxonomy of Filtering Approaches

Surveys content-based, collaborative, and hybrid filtering strategies. Positions collaborative filtering within the broader landscape.

Lesson 2 • What Recommender Systems Do

Defines recommendation systems and their core function of matching users to items. Establishes vocabulary used throughout the course.

Lesson 3 • Data Requirements and Collection

Identifies the types of data needed to power collaborative filtering models. Covers collection strategies and early data quality considerations.

Lesson 1 • Matrix Operations for Recommendations

Applies matrix multiplication, transposition, and inversion to rating matrices. Prepares students for matrix factorisation decomposition techniques.

Lesson 2 • Probability and Statistics Basics

Reviews conditional probability, expectation, and variance as they apply to rating prediction. Grounds probabilistic models introduced in later chapters.

Lesson 3 • Linear Algebra Essentials

Covers vectors, matrices, and dot products as representations of users and items. Directly supports similarity computation and matrix factorisation later.

Lesson 1 • Item-Based Nearest-Neighbour Filtering

Computes item-to-item similarity and recommends items related to those a user already rated. Contrasts scalability with user-based approaches.

Lesson 2 • Evaluating Memory-Based Models

Applies MAE, RMSE, and ranking metrics to assess prediction quality. Introduces train-test splitting and cross-validation for recommender evaluation.

Lesson 3 • User-Based Nearest-Neighbour Filtering

Finds similar users and aggregates their ratings to predict a target user's preferences. Introduces neighbourhood selection and weighted averaging.

Lesson 4 • Normalisation and Bias Correction

Addresses user rating bias through mean-centering and z-score normalisation. Improves prediction accuracy before similarity is computed.

Lesson 1 • Regularised Matrix Factorisation

Introduces the objective function with L2 regularisation to prevent overfitting latent factors. Derives the gradient update rules used in optimisation.

Lesson 2 • Incorporating Bias Terms

Extends the base factorisation model with global, user, and item bias parameters. Demonstrates measurable accuracy gains on standard benchmarks.

Lesson 3 • Singular Value Decomposition Fundamentals

Explains SVD as a low-rank approximation of the rating matrix. Connects latent dimensions to hidden user and item characteristics.

Lesson 4 • Stochastic Gradient Descent Training

Trains matrix factorisation models efficiently using SGD on observed ratings. Covers learning rate schedules and convergence monitoring.

Lesson 1 • Weighted Regularised Matrix Factorisation

Applies confidence-weighted loss to treat unobserved interactions as soft negatives. Derives the WRMF objective and its ALS solution.

Lesson 2 • Bayesian Personalised Ranking

Frames recommendation as pairwise ranking by maximising the likelihood that observed items rank above unobserved ones. Trains with SGD on sampled triplets.

Lesson 3 • Implicit vs. Explicit Feedback Revisited

Contrasts the statistical properties of implicit signals with explicit ratings. Establishes why standard factorisation assumptions must be revised.

Lesson 1 • Sequence-Aware Recommendation Models

Models temporal user behaviour with recurrent and attention-based architectures. Captures session context that static models miss.

Lesson 2 • Neural Collaborative Filtering Architecture

Replaces the dot product with a multilayer perceptron to model complex interactions. Introduces the GMF and MLP branches of the NCF framework.

Lesson 3 • Comparing Deep and Shallow Models

Benchmarks neural CF against matrix factorisation on accuracy, training cost, and interpretability. Guides model selection based on dataset size and latency constraints.

Lesson 4 • Autoencoders for Recommendation

Uses encoder-decoder networks to reconstruct rating vectors and impute missing values. Covers denoising autoencoders as a regularisation strategy.

Lesson 1 • Distributed Model Training

Scales factorisation and neural CF training across multiple machines using data and model parallelism. Addresses gradient synchronisation and fault tolerance.

Lesson 2 • Feature Stores and Embedding Serving

Manages user and item embeddings in a feature store for consistent training and serving. Covers versioning, retrieval latency, and embedding refresh schedules.

Lesson 3 • Monitoring and Maintaining Live Systems

Tracks recommendation quality in production using online metrics and data drift detection. Establishes retraining triggers and rollback procedures.

Lesson 4 • Approximate Nearest-Neighbour Search

Replaces exact similarity search with ANN indexes to retrieve candidates in milliseconds. Covers HNSW, LSH, and product quantisation methods.

Lesson 1 • Graph Neural Networks for CF

Models user-item interactions as a bipartite graph and propagates embeddings through graph convolutions. Covers LightGCN and its simplifications.

Lesson 2 • Federated Collaborative Filtering

Trains recommendation models across decentralised user devices without sharing raw interaction data. Addresses communication efficiency and privacy guarantees.

Lesson 3 • Conversational and Interactive Recommendations

Incorporates user feedback through multi-turn dialogue to refine recommendations dynamically. Connects reinforcement learning and bandit frameworks to CF.

Lesson 4 • Emerging Research Directions

Reviews large language model integration, causal inference, and self-supervised pretraining in recommendation research. Guides students in identifying tractable open problems.