Contrastive Collaborative Filtering for Cold-Start Item Recommendation

Abstract

The cold-start problem is a long-standing challenge in recommender systems. As a promising solution, content-based generative models usually project a cold-start item's content onto a warm-start item embedding to capture collaborative signals from item content so that collaborative filtering can be applied. However, since the training of the cold-start recommendation models is conducted on warm datasets, the existent methods face the issue that the collaborative embeddings of items will be blurred, which significantly degenerates the performance of cold-start item recommendation. To address this issue, we propose a novel model called Contrastive Collaborative Filtering for Cold-start item Recommendation (CCFCRec), which capitalizes on the co-occurrence collaborative signals in warm training data to alleviate the issue of blurry collaborative embeddings for cold-start item recommendation. In particular, we devise a contrastive collaborative filtering (CF) framework, consisting of a content CF module and a co-occurrence CF module to generate the content-based collaborative embedding and the co-occurrence collaborative embedding for a training item, respectively. During the joint training of the two CF modules, we apply a contrastive learning between the two collaborative embeddings, by which the knowledge about the co-occurrence signals can be indirectly transferred to the content CF module, so that the blurry collaborative embeddings can be rectified implicitly by the memorized co-occurrence collaborative signals during the applying phase. Together with the sound theoretical analysis, the extensive experiments conducted on real datasets demonstrate the superiority of the proposed model. The codes and datasets are available on https://github.com/zzhin/CCFCRec.

Preference Aware Dual Contrastive Learning for Item Cold-Start Recommendation

“…The Thirty-Eighth AAAI Conference on Artificial Intelligence (AAAI-24) (Wei et al 2021) employ generative adversarial network (GAN) (Goodfellow et al 2020) and contrastive learning, respectively, to align the collaborative feature and the content feature of the same warm item. Whereas CFCCRec (Zhou, Zhang, and Yang 2023) samples two warm items with cooccurrence users, and aligns the content feature of one item to the collaborative feature of the other item. However, since cold items have no collaborative feature and co-occurrence users with warm items, these item-level alignment strategy based methods did not take this condition into consideration in the training stage, leading to limited performance for the cold items that firstly appear in the test stage, especially for the low similarity cold items, i.e., the cold items whose multimedia information is less similar to the warm items.…”

“…In this paper, we focus on the second category, i.e., zeroshot item cold-start problem, where cold items are newly coming with no historical interactions. To address this problem, existing methods introduce the item's multimedia information to obtain the content feature as the representation of the cold item (Volkovs, Yu, and Poutanen 2017;Du et al 2020;Wang et al 2021;Wei et al 2021;Zhou, Zhang, and Yang 2023). Some of these methods integrate the item's collaborative feature into its content feature, and attempt to infer the integrated feature via randomly corrupting the collaborative feature of a subset of warm items in the training stage from the perspective of robust learning (Volkovs, Yu, and Poutanen 2017;Du et al 2020).…”

“…GAR (Chen et al 2022) generates an item's content feature with a similar distribution as its collaborative feature, by training the generator and the recommender, adversarially. CLCRec (Wei et al 2021) and CFCCRec (Zhou, Zhang, and Yang 2023) obtain the item's content feature by maximizing the mutual information with its own collaborative feature, or another item's collaborative feature with co-occurrence signal. However, these item-level alignment methods can not fully consider the condition of cold items in the training stage, resulting in limited performance.…”

“…We analyze the effectiveness of the two group-level user preference calculation methods in our PAD-CLRec, where PAD-CLRec/A and PAD-CLRec/W denote our PAD-CLRec with the Average method and the Weighted Sum method, respectively. To make the results more convincing, an additional item-level aligning method based on co-occurrence relation in study (Zhou, Zhang, and Yang 2023) is introduced to replace the L P in our joint objective function and form the Sampling method named PAD-CLRec/S for comparison. Results are given in Table 3.…”

Robust Basket Recommendation via Noise-tolerated Graph Contrastive Learning

“…CMP-PSP [50] effectively leveraged contrastive multi-view learning and pseudo-siamese networks to mitigate data sparsity and noisy interactions. CCFCRec [58] adopts contrastive collaborative filtering for cold-start item recommendation which applies contrastive learning to transfer the co-occurrence signals to the content CF module. KACL [40] performs contrastive learning across the user-item interaction view and KG view to include the knowledge graph in the recommendation while eliminating the noise it may introduce.…”