From Passive Browsing to Adaptive Interaction: A Hybrid Online Shopping Framework
Abstract:
In the traditional online shopping paradigm, e-commerce platforms retrieve and rank products given a search query and present the results. Users can browse various items to explore and may rewrite another query to repeat the process. With this widely accepted setting, e-commerce platforms passively provide what users ask for, while the query may not accurately express the user's intention. With advancements in Natural Language Processing, an increasing number of research papers have proposed conversational recommendation system, where platforms actively engage users with targeted questions to deliver more accurate recommendations. However, entirely conversational recommendation systems that require users to provide details over multiple rounds may demand excessive user attention, and a sudden shift from the traditional shopping paradigm to a conversation-based system could be difficult for users to adapt to.
In this thesis, we propose a new online shopping paradigm that allows users to search and explore within a familiar and conventional setting, while asking critical questions only when necessary. Realizing this vision entails addressing several key research challenges. First, what questions should be asked based on the user's browsing behavior. Second, how the user's behavior can be more effectively interpreted. Third, when to prompt the user with questions, which requires understanding their intentions early even when the behavior sequence is short. Next, we further elaborate how these problems are tackled.
Given that the user has browsed for a while, the system needs to learn to ask the most critical questions to adjust the recommendation results. We design a novel utility score that measures if a clicked item carries essential attributes that lead to the target, then a Selection Net chooses the most important attributes to clarify and a Ranking Net rerank the results according to user's response. While the proposed Dual-SI (dual-learning model that hybrids the best from session aware recommendation and interactive methods) performs well and effectively improves the shopping efficiency, it could be further enhanced by improving the behavior sequence understanding.
To enable deep understanding of session data, we pre-train a general-purpose User Behavior Model (UBM), comprising an Interaction Encoder and a Session Encoder, using large-scale data. The UBM improves the interpretation of behavior sequences by exploiting complex intra-item semantic relationships, inter-item connections, and inter-interaction dependencies through textual behavior sequences. We employ a two-stage pre-training scheme that leverages various augmentation strategies and contrastive learning objectives to enable self-supervised learning. The pre-trained UBM is versatile and can be applied to a range of downstream tasks.
Now that we know how to precisely interpret a user’s behavior sequence and what questions to ask, the last question would be finding out when is the best time to engage the user. This requires early understanding of user’s intention given only the short behavior sequence at the early stage. However, short sequential data carries limited information and is difficult to interpret. We propose SARAF, a review-guided sequence augmentation framework that leverages Large Language Models with structured reasoning and incorporates AI feedback for quality refinement. Experiments confirmed SARAF’s potential in sparse behavior settings, with extensive analysis offering valuable insights.
To summarize, this thesis presents a framework for next-generation online shopping, enabling users to browse and explore freely while the system actively clarifies critical questions only when necessary. We identify three key research challenges essential to both realizing this framework and advancing the field: determining what critical questions to ask, improving user behavior understanding, and identifying the right moment to intervene through early intent recognition from short behavior sequences. Extensive experiments confirmed the effectiveness of our methods. Beyond facilitating a smooth transition from passive browsing to adaptive engagement, this thesis also lays a strong foundation for future research, such as incorporating multi-modal data to further enhance user understanding and improving explainability by providing transparent reasoning behind each question to foster trust.