Deep Feedback Network for Recommendation

Abstract

<jats:p>Both explicit and implicit feedbacks can reflect user opinions on items, which are essential for learning user preferences in recommendation. However, most current recommendation algorithms merely focus on implicit positive feedbacks (e.g., click), ignoring other informative user behaviors. In this paper, we aim to jointly consider explicit/implicit and positive/negative feedbacks to learn user unbiased preferences for recommendation. Specifically, we propose a novel Deep feedback network (DFN) modeling click, unclick and dislike behaviors. DFN has an internal feedback interaction component that captures fine-grained interactions between individual behaviors, and an external feedback interaction component that uses precise but relatively rare feedbacks (click/dislike) to extract useful information from rich but noisy feedbacks (unclick). In experiments, we conduct both offline and online evaluations on a real-world recommendation system WeChat Top Stories used by millions of users. The significant improvements verify the effectiveness and robustness of DFN. The source code is in https://github.com/qqxiaochongqq/DFN.</jats:p>

Modeling Users' Contextualized Page-wise Feedback for Click-Through Rate Prediction in E-commerce Search

“…Recently, some pioneering work (DFN [33], DSTN [25]) highlight the importance of modeling both users' positive and negative feedback for CTR prediction. Besides the clicked behaviors, there are abundant exposure data in users' historical behaviors called unclicked behaviors (i.e., items which are impressed to the user but are not clicked).…”

“…The results show that the click sequence is essential for performance but is still insufficient because they omit the user's complete interaction history. • The methods DFN [33] and DSTN [25] are the previous stateof-the-art methods for modeling both the positive and negative feedback. They outperform YoutubeNet and DIN by modeling the click and unclick sequence separately.…”

A Peep into the Future: Adversarial Future Encoding in Recommendation

“…Matrix factorization (MF) [15] and Factorization machine (FM) [18] are classical recommendation models. Recently, neural models [3,9,11,16,28,30] have been successfully verified in modeling (high-order) feature interactions. AutoInt [20] and AFN [4] further bring in self-attention and logarithmic transformation layer to capture useful feature interactions.…”

MC$^2$-SF: Slow-Fast Learning for Mobile-Cloud Collaborative Recommendation

“…to denote the representations of either clicked or real-time exposed items S 𝐹 𝑢 . Note that, the reason that we take the exposed items into account, is that they are also informative as pointed in [4,29]. In the fast component, we use two GRUs to separately encode the clicked parts and the exposed parts of H 𝐹 , termed as H 𝐹 𝑝 and H 𝐹 𝑛 respectively, and then fuse them by a target-aware attention mechanism.…”