Öneri sistemleri için ilgili bir görev kullanılarak negatif örneklerin yaratılması

Recommender systems benefit both the consumers by helping them navigate alarge selection of items and the providers by increasing their profits. Majority of thesesystems are trained based on consumer preferences such as product ratings/reviewsand purchase history. The former is considered as explicit data and the latter isconsidered as implicit data. Explicit data is easier to work with but is not easy toobtain from each user. Implicit data is immediately available when a user purchasesan item and as such majority of the recommender systems use this type of data.The drawback of using implicit data is that it only includes the observed (positive)user-item pairs and the resulting system performance suffer from lack of negativeexamples. By only looking at the implicit data, it is nearly impossible to interpretwhy a user did not purchase an item. It may be because the user does not like thatitem, which implies a genuine negative example, or because the user is simply unawareof it. There are two main approaches to tackle this issue. First one is to considerall the unobserved user-item pairs as missing which leads to a highly sparse data set.The second one is to consider all or a weighted subset of unobserved pairs as negativewhich causes miss-interpretation by losing user-item pairs that may be positive.In this work, we propose a novel approach to explicitly deal with this ambiguity,by using a related task, namely click through rate (CTR) prediction. We train a deepneural network based CTR model and use it to generate negative samples for anotherdeep recommendation model. We call our approach Related Task Sampling (RTS).We use real-world datasets from a large tourism company to train both the CTRand the recommendation models in multiple conditions and compare with commonbaselines. Our main results show that the recommendation models trained by RTSnegatives outperform the baselines by more than 10% on average.