Recommender Systems

The definitive Netflix Prize paper establishing matrix factorization as the dominant CF paradigm.

Foundational paper for implicit feedback (clicks, views); introduces confidence-weighted MF with ALS optimization.

Standard pairwise ranking loss for implicit feedback; BPR-OPT criterion ubiquitous in modern training.

Combines neighborhood and latent factor models; integrates user/item biases; essential for blending approaches.

Introduced probabilistic treatment of MF that scales linearly with observations; handles Netflix-scale sparsity with automatic regularization via adaptive priors.

General predictor combining SVM flexibility with factorization model power; models all pairwise feature interactions in linear time. Subsumes SVD++, PITF, and specialized models.

TimeSVD++ explicitly models drifting user preferences, item popularity evolution, and time-varying biases. Core component of Netflix Prize winning solution.

Seminal industry paper establishing two-stage deep learning (candidate generation + ranking) at billion-scale.

Comprehensive industry paper on two-tower embedding models; covers hard negative mining, ANN indexing, production deployment.

Applies self-attention to sequential recommendations; the Transformer foundation inspiring BERT4Rec and modern sequential models.

Adapts BERT's bidirectional self-attention with Cloze task for sequential predictions.

Replaces inner product with neural networks; introduces GMF+MLP framework (NeuMF); highly cited baseline.

Jointly trains wide linear models (memorization) with deep networks (generalization) using shared embeddings. Productionized on Google Play for 1B+ users; seminal industry paper establishing wide+deep paradigm.

Combines FM component (low-order interactions) with deep component (high-order) via shared embeddings. Unlike Wide&Deep, requires no manual feature engineering. End-to-end training.

Attention mechanism adaptively learns user interest representations from behavioral history relative to candidate items—solving fixed-length embedding bottleneck. Deployed on Alibaba main traffic.

Foundational paper applying causal inference to recommenders; introduces IPS-based unbiased estimators.

Addresses position bias in click data; propensity-weighted ranking SVM; essential for implicit feedback.

Introduces SNIPS reducing IPS variance without bias; widely used in production.

Comprehensive tutorial covering IPS, SNIPS, Doubly Robust estimators with implementation guidance.

Combines imputation models with propensity scoring for unbiased performance estimation; theoretically principled doubly robust framework specifically for recommender systems with MNAR data.

CausE optimizes recommendation policy for causal treatment effects (not just prediction) using domain adaptation to handle biased logged data. RecSys Best Paper Award.

Two-stage approach using substitute confounders from exposure modeling; addresses unobserved confounding in observational recommendation data from principled causal perspective.

Pioneering paper on balancing accuracy with diversity; introduces intra-list diversity metrics.

First rigorous longitudinal study measuring filter bubble effects at individual level using MovieLens.

Analyzes how recommenders affect aggregate sales diversity (long tail vs. blockbusters).

Spotify's production system combining contextual bandits for exploration with explainable recommendations.

Introduces LinUCB algorithm for contextual bandits; 12.5% CTR lift over context-free bandits on 33M+ Yahoo events. Foundational paper for bandit-based recommendations.

Comprehensive evaluation across display advertising and news recommendation demonstrating Thompson Sampling achieves state-of-the-art results; reignited practical interest in TS.

Shows accuracy-optimized recommenders crowd out users' lesser interests; proposes KL-divergence calibration metrics and efficient greedy re-ranking. Netflix research, RecSys Best Paper Nominee.

Uses dropout to force reliance on content features when collaborative signals unavailable; practical and widely adopted.

Applies meta-learning to item cold-start; learns to adapt quickly from user history.

Combines deep learning for content (stacked denoising autoencoders) with CF; influential hybrid approach.

Applies MAML framework to user cold-start; learns initialization that quickly adapts to new users.

First paper framing cold-start as zero-shot learning problem; proposes Low-rank Linear Auto-Encoder (LLAE) addressing domain shift and spurious correlations. Novel theoretical connection enabling attribute-based recommendations for entirely new users.

State-of-the-art combining transfer learning and meta-learning for cross-domain cold start; uses source domain to warm-start target domain predictions for zero-interaction users.

First billion-scale GCN deployment. Random-walk neighborhood sampling, curriculum training with hard negatives, MapReduce inference for Pinterest's 3B node graph. Seminal industrial GNN paper.

First to explicitly encode collaborative signal via high-order connectivity on user-item bipartite graphs through embedding propagation. Foundational theoretical framework for GNN-based CF.

Demonstrates feature transformation and nonlinear activation contribute little to CF; proposes simplified GCN using only neighborhood aggregation—16% improvement over NGCF with simpler architecture.

Unifies user-item interaction graphs with knowledge graphs via attention-based neighbor discrimination; enables more accurate, diverse, and explainable recommendations.

Task-specific gating networks over shared expert submodels; explicitly learns task relationships from data. Deployed at Google; widely adopted at Pinterest, LinkedIn, Meta.

Addresses 'seesaw phenomenon' where improving one task hurts others; explicitly separates shared/task-specific experts with progressive routing. Achieved 2.23% lift at Tencent.

Models CVR over entire impression space (not just clicks) to eliminate sample selection bias; jointly trains CTR and CTCVR with shared embeddings. Alibaba/Taobao deployment.

PE-LTR framework using coordinate descent with theoretical Pareto efficiency guarantees for balancing CTR and GMV. Alibaba e-commerce production system.

Scales REINFORCE to YouTube production with millions of items; introduces top-K off-policy correction for learning from biased logged feedback. Seminal industry RL paper.

Value-based RL decomposition for slate recommendations; long-term slate value decomposes into tractable item-wise LTVs. YouTube production deployment solving combinatorial slate optimization.

Actor-Critic framework capturing inter-item relationships in list recommendations; addresses sequential decision-making for e-commerce. JD.com application.

Model-based RL using world models for recommendation; addresses data efficiency challenge of learning from limited online interactions.

Foundational framework for exposure-based fairness; develops efficient algorithms maximizing user utility while satisfying fairness constraints on content provider exposure. Applicable to job platforms, e-commerce, any two-sided marketplace.

Studies relevance-fairness trade-offs in two-sided marketplaces; proposes counterfactual evaluation for balancing consumer and supplier objectives. Spotify production research.

Learning-to-rank enforcing merit-based fairness guarantees to item groups while accounting for selection bias in implicit feedback. Addresses fairness maintenance over time.

Comprehensive study of how recommendations affect content providers; examines supplier-side impacts and multi-stakeholder trade-offs. Spotify research.

Counterfactual estimators for offline evaluation correlating with online A/B tests; addresses bias-variance tradeoffs in off-policy evaluation. Criteo production system enabling faster iteration.

Demonstrates RMSE optimization doesn't translate to top-N accuracy; establishes methodology for evaluating recommendation quality beyond prediction error metrics. Netflix context.

Critical analysis showing many neural approaches don't outperform well-tuned baselines; emphasizes reproducibility and rigorous evaluation methodology. Essential for experimental design.