Why Not Together? A Multiple-Round Recommender System for Queries and Items

TL;DR

This paper introduces MAGUS, a novel multi-round recommender framework that jointly leverages user queries and items by constructing a word-based relational graph that connects query and item representations. An offline initializer from any existing recommender provides initial node scores, while a label-propagation updater and a feature-propagation weight trainer iteratively refine predictions as users provide feedback across rounds. MAGUS can be integrated with a wide range of base recommender systems and demonstrates improved efficiency in identifying user-aligned items across three real-world datasets, outperforming item-only baselines and competitive conversational approaches. The approach supports practical deployment, extensibility to top-N recommendations, and potential integration with large language models to enhance query generation and user interaction in exploratory search settings.

Abstract

A fundamental technique of recommender systems involves modeling user preferences, where queries and items are widely used as symbolic representations of user interests. Queries delineate user needs at an abstract level, providing a high-level description, whereas items operate on a more specific and concrete level, representing the granular facets of user preference. While practical, both query and item recommendations encounter the challenge of sparse user feedback. To this end, we propose a novel approach named Multiple-round Auto Guess-and-Update System (MAGUS) that capitalizes on the synergies between both types, allowing us to leverage both query and item information to form user interests. This integrated system introduces a recursive framework that could be applied to any recommendation method to exploit queries and items in historical interactions and to provide recommendations for both queries and items in each interaction round. Empirical results from testing 12 different recommendation methods demonstrate that integrating queries into item recommendations via MAGUS significantly enhances the efficiency, with which users can identify their preferred items during multiple-round interactions.

Figures (9)

• Figure 1: An illustrated example of MAGUS recommending both queries (e.g., query B) and items (e.g., items D, E, and F), as shown in (b). This framework effectively leverages both queries (e.g., query A) and items (e.g., items A, B, and C) present in the browsing logs, serving to alleviate the data sparsity issue of user feedback, as shown in (a).
• Figure 2: An illustrated example of our relational graph organizing individual words and certain combinations of words (as shown in (a)). To maintain clarity, we have opted not to explicitly illustrate the edges representing mutual inhibition and mutual independence. In each user session, MAGUS collaborates with an offline-tuned recommendation method to initialize the node scores, where the propagation direction is from the nodes representing the items towards the other nodes (as shown in (b)). Subsequently, when a user provides her preference on a recommendation, MAGUS updates the relevant nodes by a label propagation algorithm whose propagation direction now operates from the recommended node towards the nodes representing items (as shown in (c)).
• Figure 3: Performance comparisons of MAGUS with different maximum numbers of rounds (i.e., $K_\mathtt{MAX}$) in terms of RA@$K_\mathtt{MAX}$ and SA@$K_\mathtt{MAX}$ with the FM recommendation base.
• Figure 4: Performance comparisons of MAGUS using different lengths of the recommendation list in terms of RA@3 and SA@3 with the FM recommendation base.
• Figure 5: An illustrated example of extending a single-round baseline method to our multiple-round setting in the masonry layout, where we introduce a recurrent neural network to facilitate the flow of information across rounds.

Theorems & Definitions (3)

• definition 1: Multiple-round Recommender Systems Supporting both Queries and Items
• definition 2: Words and Relational Graph
• definition 3: Conversational Agent and Human User Interactions