Recommendation and Learning to Rank

Ranking determines which items a user or query encounters. Clicks and purchases reflect both preferences and what was shown, so learning from them requires models of relevance and exposure. This review covers ranking objectives, personalized recommendation and evaluation. Text encoders and retrievers belong in NLP.

Learning ranking functions

  • RankNet - Learns document scores from pairwise preferences. A logistic function of score differences estimates which item should rank higher, and cross-entropy trains the scorer without requiring absolute relevance values. Burges et al. 2005.
    • LambdaMART - Fits boosted trees to ranking gradients weighted by the change in a ranking metric when two items swap. Extends RankNet through LambdaRank to prioritize swaps that affect measures such as NDCG. Burges 2010.

Implicit feedback and sequential recommendation

  • Confidence-weighted matrix factorization - Separates binary preference targets from confidence derived from interaction counts. Weighted least squares fits observed and unobserved user-item pairs, giving missing interactions low confidence instead of treating them as dislikes. Hu et al. 2008.
  • BPR - Optimizes personalized rankings from implicit feedback by preferring observed interactions over unobserved items. Its pairwise objective trains matrix-factorization or neighborhood models directly for ranking rather than rating prediction. Rendle et al. 2009.
  • SASRec - Uses causal self-attention over a user’s interaction history to predict the next item. Beats tested Markov, convolutional and recurrent baselines on sparse and dense datasets. Evaluation ranks the held-out item against 100 sampled negatives. Kang and McAuley 2018.
  • LightGCN - Learns user and item embeddings through linear propagation on the interaction graph and combines layers. Removes feature transformations and nonlinearities from graph convolution for collaborative filtering. He et al. 2020.

Exposure and position bias

  • Selection-aware recommendation - Weights observed ratings by inverse observation probabilities to estimate loss under selective feedback. Unbiasedness requires correct, nonzero propensities. Biased rating samples can otherwise favor the wrong recommender. Schnabel et al. 2016.
  • Learning to rank from biased clicks - Weights clicks by inverse examination propensity to correct position bias in ranking loss. A randomized intervention estimates these propensities. The guarantee depends on the examination model, not on treating unclicked items as irrelevant. Joachims et al. 2017.

Comparative evidence and evaluation

  • Tuned recommendation baselines - Of 18 neural recommendation methods reviewed, seven could be reproduced with reasonable effort. Six were often beaten by tuned neighborhood or graph heuristics. The seventh did not consistently beat a linear model. Ferrari Dacrema et al. 2019.
  • Sampled ranking metrics - Ranking a held-out item against sampled negatives can reverse algorithm comparisons relative to the full catalogue, even in expectation. Small samples make top-heavy metrics behave more like AUC, changing what the evaluation rewards. Rendle 2019.