Orthogonal Hyper-category Guided Multi-interest Elicitation for Micro-video Matching

TL;DR

OPAL addresses the challenge of micro-video matching under users’ multiple evolving interests by introducing orthogonal hyper-category guided multi-interest embeddings. It employs a two-stage training regime—pre-training to learn soft, category-aligned interests and fine-tuning to induce exclusive, evolving hard interests—coupled with a future-item prediction objective and regularizers that enforce heterogeneity and balance. Key contributions include the orthogonal hyper-category embeddings with an orthogonality constraint, a uniformity loss to prevent trivial assignments, and a unique loss to enforce hard-category exclusivity, all integrated into an efficient Faiss-based retrieval framework. Empirical results on two real-world datasets show OPAL outperforms six baselines in Recall@K and HitRate@K and yields richer, more diverse recommendations, demonstrating the value of structured, stage-wise disentanglement for large-scale micro-video matching.

Abstract

Watching micro-videos is becoming a part of public daily life. Usually, user watching behaviors are thought to be rooted in their multiple different interests. In the paper, we propose a model named OPAL for micro-video matching, which elicits a user's multiple heterogeneous interests by disentangling multiple soft and hard interest embeddings from user interactions. Moreover, OPAL employs a two-stage training strategy, in which the pre-train is to generate soft interests from historical interactions under the guidance of orthogonal hyper-categories of micro-videos and the fine-tune is to reinforce the degree of disentanglement among the interests and learn the temporal evolution of each interest of each user. We conduct extensive experiments on two real-world datasets. The results show that OPAL not only returns diversified micro-videos but also outperforms six state-of-the-art models in terms of recall and hit rate.

Figures (4)

• Figure 1: An example: multiple interests and interest evolution
• Figure 2: The architecture of OPAL, where $k=6$.
• Figure 3: The impact of multiple interests on performance
• Figure 4: The impact of multiple interests on diversity