Repetition and Exploration in Sequential Recommendation

In several recommendation scenarios, including next basket recommendation, the importance of repetition and exploration has been discovered and studied. Sequential recommenders (SR) aim to infer a user's preferences and suggest the next item for them to interact with based on their historical interaction sequences. There has not been a systematic analysis of sequential recommenders from the perspective of repetition and exploration. As a result, it is unclear how these models, that are typically optimized for accuracy, perform in terms of repetition and exploration, as well as the potential drawbacks of deploying them in real applications. In this paper, we examine whether repetition and exploration are important dimensions in the sequential recommendation scenario. We consider this generalizability question both from a user-centered and an item-centered perspective. Towards the latter, we define item repeat exposure and item explore exposure and examine the recommendation performance of sequential recommendation models in terms of both accuracy and exposure from the perspective of repetition and exploration. We find that (i) there is an imbalance in accuracy and difficulty w.r.t. repetition and exploration in SR scenarios, (ii) using the conventional average overall accuracy with a significance test does not fully represent a model's recommendation accuracy, and (iii) accuracy-oriented sequential recommendation models may suffer from less/zero item explore exposure issue, where items are mostly (or even only) recommended to their repeat users and fail to reach their potential new users. To analyze our findings, we remove repeat samples from the dataset, which often act as easy shortcuts, and focus on a pure exploration SR scenario. We find that (i) removing the repetition shortcut increases recommendation novelty and helps users who prefer to consume novel items next, (ii) neural-based models fail to learn the basic characteristics of this pure exploration scenario and suffer from an inherent repetitive bias issue, (iii) using shared item embeddings in the prediction layer may skew recommendations to repeat items, and (iv) removing all repeat items by post-processing recommendation results leads to a substantial improvement on top of several SR methods.