How to Learn Item Representation for Cold-Start Multimedia Recommendation?

Abstract

The ability of recommending cold items (that have no behavior history) is a core strength of multimedia recommendation compared with behavior-only collaborative filtering. To learn effective item representation, a key challenge lies in the discrepancy between training and testing, since the cold items only exist in the testing data. This means that the signal used to represent an item varies during training and testing-in the training stage, we can represent an item with both collaborative embedding and content embedding; whereas in the testing stage, we represent a cold item with content embedding only. Nevertheless, existing learning frameworks omit this critical discrepancy, resulting in suboptimal item representation for multimedia recommendation. In this work, we pay special attention to cold items in multimedia recommender training. To address the discrepancy, we use two vectors to represent an item: one follows the traditional way that combines collaborative embedding and content embedding, and the other assumes that the item is cold by replacing the collaborative embedding with zero vector. We then propose a Multi-Task Pairwise Ranking (MTPR) framework for model training, which enforces the observed interactions ranking higher than unobserved ones even if the item is assumed to be cold. As a general learning framework, Our MTPR is agnostic to the choice of the collaborative (and/or content) encoder. We demonstrate it on VBPR, a representative multimedia recommendation model based on matrix factorization. Extensive experiments on three datasets of diverse domains validate MTPR, which leads to better representation for both cold and non-cold items in the testing stage, thus improving the overall performance of multimedia recommendation.

Contrastive Learning for Cold-Start Recommendation

“…Note that, F that are learnable, thus differs from X that are learned by pretrained feature extractors [25,38]. Previous works either treat the feature representation and collaborative embeddings individually [7,31,36] or apply a regularizer term to encourage their dimensionwise similarity [2,22,30]. Unlike these methods, we optimize the feature encoder by encouraging the feature representations to (1) encode collaborative signals from collaborative embeddings and (2) preserve affinities with users who have interacted with before.…”

“…It alleviates the cold-start problem by improving the robustness of the model. • MTPR [7] Inspired by the counterfactual thinking, this method defines counterfactual representation to replace the collaborative embedding with the all-zero vector. • CB2CF [3] It performs a deep neural multiview model to represent the rich content information.…”

“…Such coalition effects memorized in CF models allow us to transfer collaborative knowledge to cold-start items. In this research line, extensive studies have been conducted, which can be roughly categorized into two types, robustness-based [7,31,36] and constraint-based [3,22,30,47]. Despite success, each type of methods suffers from some inherent limitations:…”

“…• From the viewpoint of robust learning, robustness-based methods frame the cold-start items as the corrupted forms of warm items whose interactions are missing. Technically, they augment the training data by randomly corrupting warm items' collaborative embeddings [7,36] or manually crafting new items [21,35], so as to enhance the generalization ability to unseen items. However, no function is designed to explicitly consider the relationships between item content and collaborative embeddings, thereby hardly refining useful collaborative signals for cold-start items.…”

Multi-trends Enhanced Dynamic Micro-video Recommendation

“…In recent years, the amount of searchable micro-videos has increased dramatically and exacerbated the need for recommender systems that can effectively mine users' preference and identify potentially interested micro-videos in a personalized manner. Due to the powerful representation learning capacity, the rapid development of deep learning techniques has nourished the research field of recommendation [17,24,33,41,42,57,58,62,65,67,68,70,73,74]. Such a development also gives rise to diverse models for video recommendation, which can be roughly categorized to collaborative filtering [2,29], content-based filtering [11,16,44,48,77], and hybrid ones [5,6,72].…”