Recommender Systems – Wikipedia Notes

Core Concepts and Vocabulary

• Collective Intelligence

  • Wisdom that emerges from large groups of users; recommender systems harness this by aggregating many users’ signals to make single-user suggestions.

• Relevance

  • Central optimisation goal: surface items that are most pertinent to a user’s current intent or long-term interests.

• Star Ratings

  • Classic explicit-feedback signal (e.g., 1–5 stars) used to infer preferences; often normalised into a common rating scale.

• Long-Tail Effect

  • Ability of RecSys to expose niche items lying outside the popular “head”; improves catalog utilisation and user satisfaction.

Methods & Classic Challenges

• Cold-Start

  • Lack of interaction data for new users/items; mitigated via hybrid models, content features, or $\epsilon$-greedy / multi-armed-bandit exploration.

• Scalability

  • Industrial systems must handle $\text{O}(10^6!!\text{–}!10^9)$ users/items; addressed with dimensionality reduction, sharding, approximate nearest neighbours (ANN), etc.

• Sparsity

  • Ratings matrix is overwhelmingly empty; similarity measures and factor models must cope with missing data.

• Dimensionality Reduction

  • Techniques (e.g.

  • Singular Value Decomposition,

  • Probabilistic Matrix Factorisation,

  • Auto-encoders) compress high-dimensional user–item space into latent factors.

• Preference Elicitation

  • Methods for gathering initial signals (onboarding quizzes, forced-choice pairs, slider ratings, etc.).

• Similarity Search

  • k-Nearest Neighbour, cosine similarity, Pearson correlation; critical for memory-based CF and candidate-generation pipelines.

Canonical Implementations

• Collaborative Search Engine – recommends search results based on aggregated user click patterns.

• Content Discovery Platform – pushes personalised articles/TV/academic papers to web, mobile, set-top boxes.

• Decision Support System – assists users in selecting products or actions (e.g., financial advisors).

• Music Genome Project – 450-attribute ontology powering Pandora’s content-based radio.

• Product Finder – interactive filters plus RecSys ranking for e-commerce.

Landmark Research Groups & Events

• GroupLens Research – University of Minnesota; pioneers of MovieLens dataset.

• MovieLens Dataset – widely used benchmark; millions of explicit ratings.

• Netflix Prize (2006–2009)

  • $\$1{,}000{,}000$ competition; target: $10\%$ RMSE improvement.

  • Winning solution: ensemble of 107 algorithms; sparked surge in matrix-factorisation & ensemble methods.

• ACM RecSys Conference – premier academic venue (est. 2007).

Recommender-System Overview

• Operate at intersection of information filtering & decision support.

• Widely deployed: playlists (Spotify/YouTube), product carousels (Amazon), open-web feeds (news/social).

• Inputs may be unimodal (music) or multimodal (books + search queries + social graph).

• Domains also include restaurants, dating, expert finding, finance.

Two Classical Paradigms

Collaborative Filtering (CF)

• Assumption: users with similar histories will prefer similar future items.

• Memory-Based CF

  • User-based or item-based k-NN; Amazon popularised item-to-item version.

• Model-Based CF

  • Learns parametric models; e.g., Matrix Factorisation solves
    $\min<em>{P,Q} \sum</em>{(u,i)\in K} (r<em>{u,i} - P</em>u^\top Q<em>i)^2 + \lambda (|P</em>u|^2+|Q_i|^2)$

• Explicit vs Implicit Data

  • Explicit: star ratings, rankings, thumbs-up.

  • Implicit: clicks, dwell-time, purchases, social follows.

• Known Issues: cold-start, scalability, sparsity; bandit exploration addresses the first.

Content-Based Filtering (CBF)

• Principle: recommend items similar to those the user liked, based on item feature vectors.

• Item Representation

  • Often $tf\text{–}idf$ bag-of-words:
    $tfidf<em>{t,d}=tf</em>{t,d}\times idf_t$

  • Or dense embeddings from CNN/RNN/Transformer for images, audio, text.

• User Profile

  • Weighted centroid of liked-item vectors; updated via feedback.

• Machine-Learning Classifiers

  • Naïve Bayes, SVM, decision trees, neural nets estimate $P(\text{like}\,|\,\text{features})$.

• Strength: handles new items instantly; Weakness: cannot suggest serendipitously dissimilar content.

Hybrid Approaches

• Weighted – blend CF & CBF scores (e.g., Netflix).

• Switching – pick algorithm adaptively per context.

• Mixed – present interleaved lists from multiple engines.

• Cascade / Meta-Level – stage-wise refinement or one model consumes another’s output.

Advanced & Emerging Technologies

Session-Based RecSys

• Rely solely on short-term interaction sequence; key for anonymous traffic.

• Models: RNNs, GRU4Rec, Transformers, XLNet-style self-attention.

Reinforcement-Learning (RL) RecSys

• Treat user as environment; agent selects item $a<em>t$, receives reward $r</em>t$ (click, dwell).

• Contextual-bandit and policy-gradient methods optimise long-term engagement.

Multi-Criteria RecSys

• Predict vector $\vec r_{u,i}=(r^1,r^2,\dots,r^m)$ instead of scalar rating; use MCDM to aggregate.

Risk-Aware RecSys

• Balance relevance with risk of disturbing user (time-of-day, situation).

• DRARS: contextual bandit + content model.

Mobile & Location-Based RecSys

• Leverage GPS, temporal patterns; must handle noisy, heterogeneous, privacy-sensitive data.

• Example: Uber/Lyft driver pickup-point recommendations via trajectory mining.

Generative Recommenders

• Frame recommendation as sequence generation; treat every user action as a token.

• HSTU (Hierarchical Sequential Transduction Units) handles trillion-parameter streams; custom self-attention scalable to long histories.

Algorithmic Building Blocks

• k-Nearest Neighbour (k-NN) for similarity search; approximated via locality-sensitive hashing (LSH) or FAISS/ScaNN.

• Pearson Correlation similarity:
  $\text{sim}(u,v)=\frac{\sum<em>i (r</em>{u,i}-\bar r<em>u)(r</em>{v,i}-\bar r<em>v)}{\sqrt{\sum</em>i (r<em>{u,i}-\bar r</em>u)^2}\,\sqrt{\sum<em>i (r</em>{v,i}-\bar r_v)^2}}$

• Embeddings – learn low-dim vectors for users/items via Word2Vec, Doc2Vec, graph neural nets, etc.

Evaluation Methodologies

• Offline – replay on historical dataset; metrics:

  • $MSE$ / $RMSE$,

  • Precision@k, Recall@k, $nDCG$,

  • Diversity, Novelty, Coverage.

• User Studies – small panels judge recommendation lists.

• Online A/B Tests – live traffic; KPIs: click-through rate (CTR), conversion, dwell-time.

• Known caveats: offline-online mismatch, dataset biases, duplicate data.

Beyond Accuracy – Additional Quality Dimensions

• Diversity – variety within list; increases satisfaction.

• Serendipity – surprising yet relevant items.

• Trust & Explainability – transparency builds acceptance; e.g., “because you watched …”.

• Privacy – GDPR, anonymisation challenges (Netflix Prize deanonymisation case).

• Robustness / Shilling Attacks – detect fraudulent profiles.

• Labelling Effects – “Sponsored” vs “Organic” tags alter CTR.

• Recommender Persistence – re-showing items can boost engagement.

Reproducibility Concerns

• Studies show <40 % of neural RecSys papers replicate; calls for shared code, standard splits, robust baselines, framework support (LensKit, RecBole, RecPack, ReChorus, etc.).

AI-Driven Models & Architectures

KNN-Based Collaborative Filters

• Represent each user as an $n$-dimensional vector; compute statistical distance; choose $k$ nearest neighbours; infer ratings from neighbour aggregate.

Neural Networks

• ANN / MLP – ingest side features (time, session signals, social trends).

• Two-Tower Model

  • Separate user-tower $f<em>u(\cdot)$ and item-tower $g</em>i(\cdot)$ producing embeddings $\mathbf u, \mathbf v$; relevance $s=\mathbf u^\top \mathbf v$.

  • Item vectors pre-computed enables fast ANN retrieval, then rerank via deeper model.

Natural Language Processing in RecSys

• Extract aspects & sentiment from reviews; use LSA, SVD, LDA or Transformers (BERT) to enrich item metadata.

Domain-Specific Applications & Examples

• Academic Content Discovery – tackle ≈$6000$ new papers/day; services augment Google Scholar alarms; challenge for early-career researchers lacking publication history.

• Decision-Making / Bridging-Based Ranking – tools like Polis, Remesh surface consensus-building content; Twitter & YouTube piloting community-note style ranking.

• Connected Television – central portal merges linear broadcast + OTT streams; personalised guides differentiate service providers.

Historical Milestones

• 1979 – Grundy by Elaine Rich: stereotype-based book advisor.

• 1990 – Digital Bookshelf (Karlgren et al.).

• 1994–1995 – GroupLens, Bellcore, MIT agents: foundational collaborative filtering papers.

• 2010 – ACM Software Systems Award to GroupLens.

• Numerous patents (Amazon item-to-item, etc.) and >50 open-source libraries now available.

Ethical, Legal & Practical Implications

• Supreme-court debate (Gonzalez v. Google) on whether search vs recommendation are distinct technologies.

• Patent landscape influences innovation & licensing.

• Privacy lawsuits (Doe v. Netflix) halted second Netflix Prize.

• Trade-off between personalisation and data protection leading to federated-learning & on-device inference research.

Key Mathematical & Statistical References

• RMSE: $RMSE = \sqrt{\frac{1}{|K|}\sum<em>{(u,i)\in K} (\hat r</em>{u,i} - r_{u,i})^2}$

• Discounted Cumulative Gain:
  $DCG@k = \sum<em>{i=1}^k \frac{2^{rel</em>i}-1}{\log_2(i+1)}$

• Multi-Armed Bandit Exploration: Upper-Confidence-Bound (UCB1)
  $UCB<em>i = \bar x</em>i + \sqrt{\frac{2\ln n}{n_i}}$

• Transformer Self-Attention simplified:
  $\text{Attention}(Q,K,V)=\text{softmax}\left(\frac{QK^\top}{\sqrt d}\right)V$

These bullet-point notes capture every notable definition, example, algorithm, challenge, evaluation nuance, historical fact, ethical concern, and advanced technique discussed across all transcript pages, providing a comprehensive stand-alone study guide for recommender-system theory and practice.