Content area
Abstract
Producing personalized predictions, i.e., filtering redundant information and locating desired parts as predictions of a specific user’s information needs according to his/her historical interactions, plays an indispensable role in online applications. In related literature, personalized predictions are popularly produced by modeling within the framework of recommender systems such as traditional recommendation and sequential recommendation that model users’ static and dynamic preferences towards items, respectively. Recently, personalized predictions of individual elements have been extended to the set level where subsequent sets (each with arbitrary number of elements) are predicted given previous sets, namely Temporal Sets Prediction (TSP). However, most of the existing models that propose to provide personalized predictions are built with the purpose of shrinking the discrepancy between predictions and targets (i.e., improving the accuracy of predictions), leading the provided suggestions to be monotonous in terms of topics or categories. To enrich the suggestions and broaden users’ horizons, diversifying personalized predictions has become a trending research topic. This thesis presents our efforts in incorporating diversification and personalization by means of the widely adopted elegant probabilistic models — Determinantal Point Processes (DPPs).
First, an intuitive design of adopting the maximum a posteriori (MAP) inference over DPP distribution for generating diverse and relevant suggestions is implemented in traditional recommender systems. We directly formulate the inherent goal of diversity promoting recommendation as a two-objective optimization problem by simultaneously minimizing recommendation errors and maximizing diversity, which are integrated into Generative Adversarial Nets (GAN). In addition to the common practice of generating DPP-distributed diverse item subsets using MAP inference in the training process, a conditional DPP samples generation algorithm is applied in the serving (evaluation) procedure to ensure that the learned diversity-promoting recommendation model is capable of providing users with expected services.
Next, the dependencies among items that constitute the DPP-distributed subsets are noticed and explored. Based on which, we propose two objective functions for sequential recommendation that are both dependency-aware and diversity-aware, utilizing DPP-distributed likelihoods of standard DPPs and conditional DPPs, respectively. In contrast to commonly used objective functions (e.g., binary cross-entropy and pairwise ranking BPR), which primarily prioritize accuracy and are usually formulated under the assumption of independence among estimated items, our proposed objective functions for sequential recommendation stand out because they explicitly consider dependencies and incorporate diversity as a guiding factor.
Then, an interesting discovery of ranking interpretation behind the probabilities of k-DPP distributed subsets inspires the proposal of set-level k-DPP probability based ranking optimization for recommendation. In this thesis, we formalize the set-level relevance and diversity ranking objectives on basis of DPP kernel decomposition. The probabilities of subsets with cardinality k learned from the specialized k-DPP distribution over a fixed size ground set enable us to compare and rank multiple items at the set level with consideration of both relevance and diversity. We implement the k-DPP based ranking in the context of both Matrix Factorization (MF) and neural networks approaches on three real-world datasets, obtaining improved relevance and diversity performances. The k-DPP ranking approach is broadly applicable, and when applied to seminal CF models it also yields strong performance improvements, suggesting that this technique holds significant value to the field of recommender systems.
Last, associating the set-level predictions in the field of TSP with structures of structured DPP (SDPP), which are both comprised of arbitrary number of elements, we present a Structured Temporal Sets Prediction framework (STSP), which explores and exploits three types of underlying relationships (i.e., preference correlation, cooccurrence pattern, and temporal sets dependency). STSP makes two significant contributions to this field. First, elements in a set of temporal sequence are directly modeled as a SDPP structure, in which a DPP distribution is modeled over collections of structures (temporal sets) instead of over subsets of individual items. Second, all trainable representations and underlying relationships are integrated into a dedicated sets-level objective function with the intention of enhancing the probability of selecting the desired subset of structures as a DPP. Accordingly, we treat temporal sets in a broader perspective compared to existing combination manners, enabling the underlying relationships to be explicitly represented.
Altogether, we explore an interesting research task of producing diversified and personalized predictions, which is implemented in varied contexts of item-level recommendation and set-level TSP. To achieve diversification and meanwhile boost accuracy performance, the repulsion character of DPP is exploited from multiple aspects, as well as dependencies among elements are captured. The intuitive design of dual consideration of diversity in both training and serving procedures, the noticed dependency across training sequence instances, the interesting discovery of k-DPP probability specified ranking implication, and the association between set-level predictions and SDPP structures hold significant value for further studies.
