Modeling Heterogeneous Users Behaviors in Online Systems Public

Downloadable Content

open in viewer

Many online systems such as e-commerce, music/video streaming platforms have been proliferating in recent decades, creating dramatic changes in people’s shopping experiences by providing accessibility to incredible volumes of products, and enabling millions of users to sell/purchase online commodities. In such systems, understanding behavior of both product sellers and customers, two main objects of the online systems, is important to the systems’ prosperity. Thus, this dissertation makes three unique contributions as followings: First, we focus on understanding and characterizing the product delivery activities of the product owners. This task has played an essential role in maintaining not only the trust between the consumers and the product owners but also the trust between these two objects and the platform providers. Un- fortunately, in the literature, little is known to address the problem. In this direction, we extract novel features that reveal factors, which influence to the product delivery phase of the product owners. As a result, we build predictive models for on-time product delivery identification and delivery duration time estimation in the crowdfunding platforms. Second, we investigate the problem of modeling consumer behaviors with global constraints. The problem is crucial in many applications like basket- based shopping platforms, video/music streaming services (i.e. Spotify, YouTube, etc.). This is mainly because a consumer preference is often decided by the general taste of all products she/he preferred so far in her/his current session. In this line, we present a Matrix Factorization (MF) based recommender with several constraints on global similar product embeddings and global similar consumer embeddings. Due to the fact that MF based methods are intrinsic to a linear nature and dot product operator in MF based methods do not convey the crucial triangle inequality, we further proposed three novel metric learning-based neural recommenders to encode complex preferences of customers over products better. Moreover, we improve the robustness of our models by applying adversarial personalized ranking and customizing it with a flexible noise. Finally, we study the task of modeling consumer behaviors with both long- term and short-term interest dependencies. In many e-commerce platforms like Amazon, Netflix and Yelp, encoding a consumer long-term preference dependency based on all of her interacted products so far is not enough. The main reason is that her preferred next product can have a strong correlation with her current interest, which is reflected by her recently preferred items. To address the task, we present signed distance-based neural recommenders. Furthermore, we go beyond the Euclidean representation space and present our Quaternion-based recommenders that introduce the benefits of Quaternion space in modeling the consumer preferences with both long-term and short-term dependencies.

  • etd-4666
Defense date
  • 2020
Date created
  • 2020-11-21
Resource type
Rights statement


In Collection:


Permanent link to this page: