UC Santa Cruz

As online shopping becomes to be popular, the recommender system in e-commerce sites is an increasingly popular business tool to increase sales. Researchers and industry practitioners are looking for all possible approaches to improve recommendation performance. Even a minor improvement could lead to a big business return.

Traditional approaches of recommender systems include content-based methods and collaborative filtering methods. For example, if a user viewed some cameras in the website, the system learns that the user is interested in cameras and recommends more similar items to the user. Yet it might not match with the user's true purchase intention. In real-world applications, recommender systems could leverage more information from the user, both information within a single session and information across sessions. To solve this problem, we propose to investigate the session-aware recommender system in e-commerce. Such system can understand a user's short-term goal and long-term preference, in order to recommend appropriate items accordingly.

We first explore how to integrate the complementary information (e.g. the user's purchase information, search information and so on) within a single session to build a textit{unified recommender system}. We analyze the available information for the unified model, including the user's history-related information, the search-related information and the product's marketing-related information. Three unified models are proposed and compared to integrate different pieces of information.

To go beyond making recommendations within a single session, we then study how to make better recommendations across sessions. To make recommendations based on a user's previous behavior in earlier sessions, we need to understand how users make purchase decisions across sessions. Earlier research in economics and marketing indicates that a consumer usually makes purchase decision(s) based on the product's marginal net utility (i.e., the marginal utility minus the product price). Utility is defined as the satisfaction or pleasure a user gets when purchasing the corresponding product. A rational consumer chooses the product to purchase in order to maximize the total net utility. To better match users' purchase decisions in the real world, we explore how to recommend products with the highest marginal net utility in e-commerce sites. Inspired by the Cobb-Douglas utility function in consumer behavior theory, we propose a novel utility-based recommendation framework. The framework can be utilized to revamp a family of existing recommendation algorithms.

To further incorporate the time interval between sessions into the system, we propose and study a new problem: how to recommend the right product at the right time? We adapt the proportional hazards model in survival analysis and propose the new opportunity model in e-commerce. The new model estimates the joint probability of a user making a follow-up purchase of a particular product at a particular time. This joint purchase probability can be leveraged by recommender systems in various scenarios, including a zero-query pull-based scenario or a proactive push-based email promotion scenario.

We evaluate the soundness of our models with multiple metrics. Experimental results with a real-world e-commerce website (shop.com) show that they have decent predictability of the user's purchase behavior within a session and across sessions. In addition, the models significantly improve the conversion rate in pull-based systems and the user satisfaction/utility in push-based systems.

In this dissertation, we first introduce the motivation of the work.

Secondly, we report some state-of-art related work of this topic.

Thirdly, models are proposed to tackle the problem, followed by the experimental results. Then we summarize our contribution in both the research field of recommender systems and the e-commerce domain.