Optimizing top-n collaborative filtering via dynamic negative item sampling

Abstract

Collaborative filtering techniques rely on aggregated user preference data to make personalized predictions. In many cases, users are reluctant to explicitly express their preferences and many recommender systems have to infer them from implicit user behaviors, such as clicking a link in a webpage or playing a music track. The clicks and the plays are good for indicating the items a user liked (i.e., positive training examples), but the items a user did not like (negative training examples) are not directly observed. Previous approaches either randomly pick negative training samples from unseen items or incorporate some heuristics into the learning model, leading to a biased solution and a prolonged training period. In this paper, we propose to dynamically choose negative training samples from the ranked list produced by the current prediction model and iteratively update our model. The experiments conducted on three large-scale datasets show that our approach not only reduces the training time, but also leads to significant performance gains.

Optimizing Factorization Machines for Top-N Context-Aware Recommendations

“…This is because the top ranked unobserved items hurt the ranking performance more than those with lower ranked positions [27,28]. With the intuitive observation and above analysis, we devise two simple yet effective sampling schemes to further optimize RankingFM for top-N item ranking.…”

“…(16). For typical IR tasks, the candidate documents for a query in training datasets have usually been limited to a small size (e.g., 1000) because of query filtering [27]. However, for recommendation with implicit feedback, the size of candidate items is usually very huge (e.g., 10 million) as all unobserved items should be considered as candidates.…”

Irgan

“…us, it is natural to leverage GAN to distil a generative retrieval model to t such an unknown conditional data distribution using the observed user feedback data. Second, the uni ed training scheme of two schools of IR models o ers a potential of obtaining be er retrieval models because (i) the generative retrieval adaptively provides di erent negative samples to the discriminative retrieval training, which is strategically diverse compared with the static negative sampling [3,35] or dynamic negative sampling using the discriminative retrieval model itself [4,49]; and (ii) the reward signal from the discriminative retrieval model provides strategic guidance of training the generative retrieval model, which is otherwise unavailable in traditional generative retrieval model training. From the generative retrieval's perspective, IR-GAN is superior to traditional maximum likelihood estimation [18].…”

“…It is proved that the gradient of SCE matches the one of MLE, with no prerequisite of in nite noise samples, which is a very promising property of SCE learning. Dynamic negative item sampling work [34,45,49] in top-N item recommendation with implicit feedback turns to be a practical use case of SCE, with speci c solution of e cient sampling strategies.…”

BoostFM

“…The reason is to calculate 6NDCG(c) i j of different item pairs, the recommender has to score all the items in ζ c to find the rank of i and j. In IR tasks, the candidate documents returned by the retrieval model have already been reduced to a small size [34]. However, in recommendation scenario, since there is no query to filter candidate items, all unobserved items have to be considered as candidates (i.e., |ζ c | = |I|).…”