Personalised recommendation systems

Definition: - Systems that suggest items (products, content, services) tailored to an individual user’s preferences, behavior, and context. Main approaches: - Collaborative Filtering: recommends based on behavior of similar users (user-based or item-based). Strength: captures complex tastes without item metadata. Weakness: cold-start, sparsity. (See: Resnick et al., 1994; Sarwar et al., 2001.) - Content-Based Filtering: recommends items similar to those the user liked, using item features. Strength: handles new items; interpretable. Weakness: limited serendipity, needs good features. (See: Pazzani & Billsus, 2007.) - Hybrid Methods: combine collaborative and content approaches to offset weaknesses (e.g., matrix factorization + side information). (See: Burke, 2002.) - Context-Aware and Session-Based Models: incorporate time, location, device, or short-session signals; useful for dynamic preferences. - Deep Learning & Representation Learning: use neural networks (embeddings, sequence models, transformers) to model complex patterns and sequential behavior. Key components: - Data sources: explicit feedback (ratings), implicit feedback (clicks, views, purchases), contextual signals, user/item metadata. - Algorithms: nearest neighbors, matrix factorization, factorization machines, neural recommenders, bandits for exploration. - Evaluation metrics: precision, recall, MAP, NDCG, AUC, click-through rate, and business KPIs; offline vs. online (A/B) testing. Challenges and trade-offs: - Cold-start for new users/items. - Scalability to many users and items. - Diversity vs. accuracy (filter bubbles). - Privacy and fairness concerns; need for transparency and user control. - Long-term engagement vs. short-term clicks; optimizing for long-term utility may require causal or reinforcement learning methods. Ethical and philosophical considerations: - Autonomy and manipulation: recommendations can shape preferences and behavior. - Bias amplification and fairness: training data can encode social biases. - Privacy: collecting and profiling users raises consent and surveillance issues. (See: O’Neil, 2016; Barocas & Selbst, 2016; Ekstrand et al., 2018.) Concise further reading: - Ricci, Rokach, Shapira (eds.), Recommender Systems Handbook (2015). - Burke, R., Hybrid Recommender Systems, 2002. - Ekstrand, T., et al., 2018, “All the cool kids” (fairness in recommender systems).