Design and evaluation of factorization-based algorithms for user preference analysis
- Publication Type:
- Thesis
- Issue Date:
- 2019
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Recommendation systems (RecSys) are valuable for both industry and customers in many fields, including e-commerce and social media. Despite the great demand for such effective systems, many challenges still exist. A major obstruction is the sparsity and poor quality of data that hinder the learning of a satisfactory RecSys. Another obstacle lies in the open nature of RecSys: this poses a threat to their safety in applications.
In this thesis, we work towards meeting these two challenges. To improve RecSys performance, we study how to exploit information from user reviews, constraints on user behaviors and user/items demographic features. Three approaches are proposed: 1) we develop a privileged matrix factorization model that exploits reviews for the learning of both user/item factors; 2) we build a collaborative allocation model that investigates the geometric constraint on the user-preference matrix; 3) given that the features might be noisy in reality, we propose an approach to identifying noisy information and selecting useful side features. Driven by concern for the security of RecSys, our first consideration is to develop an evaluation method for testing the robustness of target models before proposing an approach to improve their resistance to malicious attacks. The target model is evaluated by measuring the minimal number of features required to mis-predict a user’s preference. To enhance the robustness of target models, we inject noise in the training phase to enforce resistance to perturbations. Target models are further guided by standard networks through the distillation of generalized knowledge to avoid performance degeneration. This way, the target model becomes more resistant to adversarial perturbations while still achieving similar performances to standard models.
We conclude the thesis by outlining main contributions and indicating primary results.
Please use this identifier to cite or link to this item: