implicit feedback
In plain wordsClues about what people like that come from what they do, such as clicking, buying, watching or skipping, rather than from ratings or reviews they deliberately give.
User-preference data inferred from behavior (clicks, purchases, dwell and watch time, plays, skips, add-to-carts) rather than stated directly, as in explicit feedback such as star ratings. Implicit feedback is abundant, costs users nothing and covers nearly every interaction, so it supplies most of the training data for modern recommender systemsrecommender system, n.: Software that guesses which songs, products, posts or jobs you are most likely to want and shows those first, as when a video app lines up the next clip or a shop suggests an add-on. and search ranking. It is harder to interpret: it is mostly positive-only, since an unclicked item may be disliked or simply unseen; noisy, since a click is not satisfaction; and biased by what the system chose to show and where (position and exposure bias). Methods therefore treat observed interactions as confidence-weighted positives and unobserved ones as weak negatives, as in weighted matrix factorizationmatrix factorization, n.: Breaking a big table of numbers, such as which users liked which films, into two smaller tables of hidden traits, so missing entries can be filled in by multiplying them back together.; optimize pairwise ranking objectives such as Bayesian personalized ranking; sample negatives; and apply counterfactual corrections such as inverse propensity weighting. Optimizing a proxy such as clicks or watch time can diverge from users' long-term interests (engagement optimizationengagement optimization, n.: Designing apps and AI to keep people clicking, watching or chatting for as long as possible because that earns money, even when it is not what is good for them.).
We have almost no star ratings, so the model trains on implicit feedback: completed listens count as positives and early skips as negatives.
Origin
Studied in information retrieval and human–computer interaction in the 1990s, where reading time, scrolling and bookmarking were tested as interest signals; Douglas Oard and Jinmook Kim surveyed its use for recommenders in 1998. Yifan Hu, Yehuda Koren and Chris Volinsky (2008) and the Bayesian personalized ranking of Steffen Rendle and colleagues (2009) established the standard modeling approaches.
Related
- recommender systemrecommender system, n.: Software that guesses which songs, products, posts or jobs you are most likely to want and shows those first, as when a video app lines up the next clip or a shop suggests an add-on.
- collaborative filteringcollaborative filtering, n.: Recommending things based on what people with similar tastes liked, without needing to know anything about the things themselves, as in 'people who bought this also bought…'.
- matrix factorizationmatrix factorization, n.: Breaking a big table of numbers, such as which users liked which films, into two smaller tables of hidden traits, so missing entries can be filled in by multiplying them back together.
- learning to ranklearning to rank, n.: Training a model to put items in the best order — search results, products, recommendations — by learning from examples of which items people judged more relevant or chose more often.
- engagement optimizationengagement optimization, n.: Designing apps and AI to keep people clicking, watching or chatting for as long as possible because that earns money, even when it is not what is good for them.
- label noiselabel noise, n.: Mistakes in the answers attached to training or test examples — a dog photo marked 'cat' — which can mislead what a model learns and distort how good it seems.
Sources
- Oard, D. W. & Kim, J. (1998). Implicit Feedback for Recommender Systems. Proc. AAAI Workshop on Recommender Systems.
- Hu, Y., Koren, Y. & Volinsky, C. (2008). Collaborative Filtering for Implicit Feedback Datasets. Proc. 8th IEEE International Conference on Data Mining (ICDM).
- Rendle, S., Freudenthaler, C., Gantner, Z. & Schmidt-Thieme, L. (2009). BPR: Bayesian Personalized Ranking from Implicit Feedback. Proc. UAI 2009.
Mentioned in
recommender systemrecommender system, n.: Software that guesses which songs, products, posts or jobs you are most likely to want and shows those first, as when a video app lines up the next clip or a shop suggests an add-on., cold start problemcold start problem, n.: The trouble a recommendation system has with brand-new users or items: with no history to learn from, it has little basis for good suggestions., collaborative filteringcollaborative filtering, n.: Recommending things based on what people with similar tastes liked, without needing to know anything about the things themselves, as in 'people who bought this also bought…'., matrix factorizationmatrix factorization, n.: Breaking a big table of numbers, such as which users liked which films, into two smaller tables of hidden traits, so missing entries can be filled in by multiplying them back together.