Modeling User Viewing Flow Using Large Language Models for Article Recommendation

TL;DR

The paper addresses information overload in article recommendation by introducing SINGLE, a framework that decomposes user interest into constant viewing flow (general preferences inferred from past clicks via LLM prompting) and instant viewing flow (time-sensitive interests captured through attention over the click history). It represents and encodes articles with title, summarized body, and attributes using BERT, and leverages LLMs to summarize bodies and extract user characteristics, followed by an attention-based mechanism to model instant interest. The approach shows consistent improvements over strong baselines on ATA and MIND, with an online UVCTR gain of 2.4% in Alibaba's ATA deployment, and ablation analyses confirm the contributions of summarization, instant/constant flows, and the flow gate. The work demonstrates the practical value of combining LLM-driven user profiling with dynamic attention over historical interactions for tailored, diverse article recommendations in real-world settings.

Abstract

This paper proposes the User Viewing Flow Modeling (SINGLE) method for the article recommendation task, which models the user constant preference and instant interest from user-clicked articles. Specifically, we first employ a user constant viewing flow modeling method to summarize the user's general interest to recommend articles. In this case, we utilize Large Language Models (LLMs) to capture constant user preferences from previously clicked articles, such as skills and positions. Then we design the user instant viewing flow modeling method to build interactions between user-clicked article history and candidate articles. It attentively reads the representations of user-clicked articles and aims to learn the user's different interest views to match the candidate article. Our experimental results on the Alibaba Technology Association (ATA) website show the advantage of SINGLE, achieving a 2.4% improvement over previous baseline models in the online A/B test. Our further analyses illustrate that SINGLE has the ability to build a more tailored recommendation system by mimicking different article viewing behaviors of users and recommending more appropriate and diverse articles to match user interests.

Figures (6)

• Figure 1: The Architecture of User Viewing Flow Modeling (SINGLE) Method. It contains user constant viewing flow modeling and user instant viewing flow modeling methods.
• Figure 2: The Framework of our User Viewing Flow Modeling (SINGLE) Method.
• Figure 3: The Topic Distribution of User Clicked Articles. We randomly choose three different users from the ATA dataset to show the topic distribution of 20 clicked articles. The articles that are annotated with the purple and blue indicate the constant interest and instant interest of users, respectively.
• Figure 4: The Similarity Scores between Representations of Users and Previously Clicked Articles. We calculate the average cosine similarity scores between user representations and clicked articles at different time steps. The user instant and constant representations are conducted by our instant and constant viewing flow modeling methods.
• Figure 5: The Pipeline of Online System. We deploy SINGLE on the ATA online website of Alibaba.