Abstract

Recommender systems have been extensively applied in various big data scenarios with the rapid development of online platforms. However, big data brings three challenges for recommender systems: cold-start, data sparsity, and low efficiency. Hashing is a promising technique for efficient online recommendations by mapping representing users and items with binary codes in a shared Hamming subspace, where users’ preferences over items can be efficiently estimated by the fast XOR operations of their hash codes.

However, existing hashing-based recommender systems provide efficient recommendations at the price of sacrificing accuracy. To improve accuracy, we explore the following four hashing-based recommendations to cope with challenges caused by approximate discrete algorithms, imbalanced data, cold-start, and sparse issues.

• We propose Discrete Ranking-based Matrix Factorization (DRMF) to improve the accuracy of discrete collaborative filtering. That is formulated with a cross-entropy loss between each user’s actual pairwise preference ranking and predicted pairwise preference ranking. Due to the non-linearity of cross-entropy loss, we seek a local quadratic upper bound to formulate the upper bound into a series of binary quadratic programming problems and solve them with semi-definite programming solvers.

• To capture preferences over items from imbalanced categories for collaborative filtering, we develop a Discrete Scale-invariant Metric Learning (DSML), which captures users’ preferences over items from categories with different intra-class variations. Specifically, we propose a scale-invariant margin based on angles at negative item points in the shared Hamming subspace. Then, we derive a scale-invariant triple hinge loss based on the margin. Finally, we integrate a pairwise ranking loss into the scale-invariant loss to capture more preference information.

• By considering the marketing application of recommendation, we propose another discrete collaborative filtering technique, Collaborative Generated Hashing (CGH). That can be applied to find potential users or items in marketing applications, where the generative network is designed with the principle of Minimum Description Length (MDL) for learning compact and informative binary codes. Besides, we explore evaluation metrics for evaluating the performance of marketing.

• We propose a discrete content-aware recommendation approach, Discrete Pairwise Hashing (DPH), to solve cold-start and sparse issues. The user-item interactive data and item content information are unified to learn effective representations. Specifically, we first pretrain robust item representations from item content data by a Denoising Auto-encoder(DAE); then, we formulate a collaborative model by combining the learned DAE and a pairwise loss; finally, we optimize with the alternating optimization strategy.

Details

Title
Discrete Techniques for Efficient Recommendations
Author
Zhang, Yan
Publication year
2022
Publisher
ProQuest Dissertations & Theses
ISBN
9798380480482
Source type
Dissertation or Thesis
Language of publication
English
ProQuest document ID
2878257611
Copyright
Database copyright ProQuest LLC; ProQuest does not claim copyright in the individual underlying works.