When Box Meets Graph Neural Network in Tag-aware Recommendation

TL;DR

BoxGNN tackles tag-aware recommendation by representing users, items, and tags as high-dimensional boxes, enabling uncertainty-aware modeling. Through two logical operators (intersection and union) and a type-aware GNN, it captures high-order signals in the collaborative tag graph, while a Gumbel-based volume objective provides differentiable guidance for box intersections. The approach yields superior results on three datasets, with ablations confirming the importance of high-order propagation and the volume-based training objective. This work advances TRS by unifying geometric embeddings with graph-based message passing to leverage multi-hop relations and tag semantics in recommendation tasks.

Abstract

Last year has witnessed the re-flourishment of tag-aware recommender systems supported by the LLM-enriched tags. Unfortunately, though large efforts have been made, current solutions may fail to describe the diversity and uncertainty inherent in user preferences with only tag-driven profiles. Recently, with the development of geometry-based techniques, e.g., box embedding, diversity of user preferences now could be fully modeled as the range within a box in high dimension space. However, defect still exists as these approaches are incapable of capturing high-order neighbor signals, i.e., semantic-rich multi-hop relations within the user-tag-item tripartite graph, which severely limits the effectiveness of user modeling. To deal with this challenge, in this paper, we propose a novel algorithm, called BoxGNN, to perform the message aggregation via combination of logical operations, thereby incorporating high-order signals. Specifically, we first embed users, items, and tags as hyper-boxes rather than simple points in the representation space, and define two logical operations to facilitate the subsequent process. Next, we perform the message aggregation mechanism via the combination of logical operations, to obtain the corresponding high-order box representations. Finally, we adopt a volume-based learning objective with Gumbel smoothing techniques to refine the representation of boxes. Extensive experiments on two publicly available datasets and one LLM-enhanced e-commerce dataset have validated the superiority of BoxGNN compared with various state-of-the-art baselines. The code is released online

Figures (5)

• Figure 1: An example of collaborative tag graph that users, items and tags are interconnected with each other. It illustrates that users may buy items due to the diverse interests.
• Figure 2: Toy example of message aggregation via intersection operation, where $Box(\cdot)$ represents the box embedding of nodes, $u$, $i$ and $t$ denotes user, item and tag, respectively. Notably, we fail to obtain the 2nd-order user box embedding, as $t_0^{(1)}$ and $i_0^{(1)}$ are disjoint.
• Figure 3: The overall framework of the proposed BoxGNN. Note the Intersection and Union are logical operations including Min/Max Offset and Centralization, where Min Offset is to obtain the minimum offset among boxes in an element-wise manner.
• Figure 4: Effect of the parameter $\beta$ on all datasets.
• Figure 5: The visualization of centroid points of Box and node embeddings of LFGCF on MovieLens dataset.

Theorems & Definitions (2)

• definition 1
• definition 2