Research on Generating Explainable Serendipity-Oriented Recommendations based on Knowledge Graph

Recommender systems have popularly been used in many Web applications to eliminate users’ information overload, with recommended items/information being tailored to users’ personal interests. However, they have often been criticized as “reinforcement of the same, relatively limited set of information.” Some commercial companies (such as Facebook) have even ad- mitted that their algorithms have suffered from “echo chambers” problem (that is also called “overspecialization,” “filter bubbles,” or “blind spots”), because users have often been trapped in a subspace of options that are too similar to their profile, and hence lose the opportunity to explore outside alternatives. In this project, we will be engaged in developing a serendipity- oriented recommending method to solve this problem, which is indeed motivated by a recent user survey we conducted on an industrial e-commerce platform (Mobile Taobao1) that reveals the significant impact of serendipitous recommendation on improving users’ satisfaction and purchase intention. To be specific, this project will have the following novelties.

1. Our serendipity-oriented recommender system will be targeted to capture users’ hidden and deep preference. For this objective, we will particularly take advantage of knowledge graph (KG), given it provides a rich semantic and heterogeneous network scenario to link up different types of entities (e.g., users, items, attributes) with each other, so that it may be feasible to discover users’ latent, indirect preference for items. Though recently there is increasing attention paid to KG-aware recommendation, little has in-depth exploited its ability in supporting users to explore unexpected (but still relevant) item space.

2. We will further consider users’ curiosity value, as a personal psychological trait, to dy- namically adjust the recommendation’s serendipity degree. Actually, most of related methods simply optimize serendipity for every one, but neglect the effect of users’ personal characteris- tics, especially curiosity, on their propensity towards accepting the serendipitous recommenda- tion. As motivated by the psychological theory that curiosity can reflect a person’s willingness to actively seek out new knowledge and experiences, we will aim to design a curiosity-based adjusting strategy for achieving personalized serendipity.

3. The third contribution will lie on further strengthening the explainability of our recom- mender system. Because the serendipitous recommendation is supposed to be unexpected and surprising to users, it should be meaningful to develop a dedicated explanation method that is capable of disclosing the relevance of such recommendation to users, so they may be more likely to accept it in the practical situation.

4. Last but not least, in order to judge the actual performance of our serendipity-oriented recommending algorithm, we will not only verify the existing metrics based on a large-scale user feedback dataset that we have collected in the recent survey, but also propose a potentially more precise metric that is able to unify multiple components of serendipity (i.e., relevance, timeliness, and unexpectedness) into an atomic measurement.

We believe this work will provide a timely study on boosting recommender systems to be more powerful, in terms of supporting users to have unexpected exploration and discovery. With supports of industrial partners such as Alibaba and Wisers2, we will develop prototype systems to be applicable in both e-commerce and digital media domains. We will conduct a series of empirical user studies and offline simulations to measure our systems. Meanwhile, we will release our serendipity dataset to the public, which, to the best of our knowledge, is the first dataset containing a large amount of real users’ feedback on recommendations in terms of their perceived serendipity.