Personalized micro video recommendation aims to recommend the micro videos tailored to user preference based on the user’s interaction history with the micro videos, which has drawn increasing attention from both the academic and industrial communities. Existing solutions primarily concentrate on video-level interactions between users and micro videos to model their preferences, and cannot distinguish the finer-grained users’ interactions with various modalities. Ignoring modality-level interactions prevents the full understanding of the user’s true and subtle preferences on micro videos. To this end, in this article, we propose a Contrastive Multimodal Interaction Graph Learning ( MVideoRec ) model to automatically and explicitly learn the modality-level interaction between users and micro videos for recommendations. Specifically, we designed a graph structure learning module with a sparsification strategy to infer modality-level interaction graph, which will be dynamically and iteratively updated based on the node representations obtained from the node representation learning module. Furthermore, to address the lack of ground truth labels, we propose to generate teacher view from video-level interaction graph and student view from modality-level interaction graph, as well as construct intra-modality and inter-modality contrastive pairwise instances to provide self-supervised signals. Extensive experiments on three real-world micro video datasets validate the effectiveness of MVideoRec.

Cross-domain recommender systems could potentially improve the recommendation performance by means of transferring abundant knowledge from the auxiliary domain to the target domain. They could help address some key challenges in recommender systems, such as data sparsity and cold start. However, most existing cross-domain recommendation approaches represent the user preferences based on a single kind of user’s feature or behavior and fail to explore the hidden interaction effects of different kinds of features or behaviors. In this article, we propose the S equential and G raphical Cross -Domain Recommendations with a Multi-View Hierarchical Transfer Gate (SGCross) to transfer user representations from multiple perspectives. The SGCross model constructs a user profile by learning the personal preference from a personal view, the dynamic preference from a temporal view, as well as the collaborative preference from a collaborative view. Specifically, a Multi-view Hierarchical Gate (MHG) is designed to transfer the informative representations of user knowledge on different views from the auxiliary domain separately, aiming to enhance the user representations. Furthermore, a two-stage attentive fusion module is designed to integrate transferred information at two levels: the domain level and the view level. Extensive experiments on the Amazon dataset and the Douban dataset have demonstrated that SGCross effectively improves the accuracy of cross-domain recommendations and outperforms the state-of-the-art baseline models.

Extreme Multi-label text Classification (XMC) aims to find the most relevant labels (i.e., the positives) for a document from an extremely large label set.The remaining labels are regarded as the negatives.Recently, the deep learning-based methods have been widely used to solve XMC, most of which use sigmoid as the activation function of output layer and use binary cross entropy loss as the learning objective.However, the existing methods suffer from the following limitations.First, the score of each label is predicted independently, where the label rank-missing problem is ignored.Second, the cardinalities of the positives and the negatives are extremely unbalanced in XMC, which makes the classifier more biased towards the majority one.In this paper, we use label group to denote the positive and the negative labels and propose a novel XMC model leveraging Group-wise label Ranking (X-GRank) to address those limitations.Specifically, X-GRank uses the newly proposed GRank loss to rank the label groups.Then, X-GRank solves the label imbalance problem by constraining the backward gradient amplitude between label groups.Extensive experiments show that X-GRank outperforms the state-of-the-art methods on five widely used datasets.

Next basket recommendation aims to provide users a basket of items on the next visit by considering the sequence of their historical baskets. However, since a user’s purchase interests vary over time, historical baskets often contain many irrelevant items to his/her next choices. Therefore, it is necessary to denoise the sequence of historical baskets and reserve the indeed relevant items to enhance the recommendation performance. In this work, we propose a Hierarchical Reinforcement Learning framework for next Basket recommendation, named HRL4Ba, which learns the personalized inter-basket and intra-basket contexts of the user for dynamic denoising. Specifically, the high-level and the low-level agent in the denoising module perform hierarchical decisions, i.e., revise baskets and remove items; the recommendation module serves as the environment to give feedback to agents and recommends the next basket. Extensive experiments on two e-commerce datasets show the HRL4Ba outperforms existing state-of-the-art methods, and our ablation studies further show the effectiveness of each component in HRL4Ba.

Social recommendation (SR) aims to enhance the performance of recommendations by incorporating social information. However, such information is not always reliable, e.g., some of the friends may share similar preferences with the user on a specific item, while others may be irrelevant to this item due to domain differences. Therefore, modeling all of the user’s social relationships without considering the relevance of friends will introduce noises to the social context. To address this issue, in this work, we propose a Denoisingguided deep Reinforcement Learning framework for Social recommendation (DRL4So). Specifically, the agent (i.e., social denoiser) in our framework automatically masks the user’s friends who are irrelevant to the target item; Then, the environment (i.e., recommender) is designed to give rewards to the agent for social denoising without supervised signals; Finally, the two components are jointly trained by DPG to ensure that social denoising correctly guides the recommendation. We conduct extensive experiments on three public datasets, and the results show that DRL4So outperforms existing state-of-the-art SR methods (improving 38.67% and 19.81% in terms of HR@10 and NDCG@10, respectively).

Predicting the popularity of shopping guide micro-videos incorporating merchandise is crucial for online advertising. What are the significant factors affecting the popularity of the micro-video? How to extract and effectively fuse multiple modalities for the micro-video popularity prediction? This is a question that needs to be urgently answered to better provide insights for advertisers. In this paper, we propose a Multimodal and Temporal Attention Fusion (MTAF) framework to represent and combine multi-modal features. Specifically, we first explore the importance of the micro-video content-agnostic factors using two existing tree-based ensemble methods. Furthermore, we employ three state-of-the-art pre-trained models, BERT, VGGish and ResNet152, to obtain high-level multimodal content representations, including uploaders’ description of products, vocal emotion, facial attractiveness, respectively. In addition, a bi-directional GRU is used to learn early popularity trend characteristics of the micro-video. Finally, a multimodal and temporal attention mechanism layer is designed to combine all features from the multiple sources. Comprehensive experiments are conducted on TikTok e-commerce micro-video dataset to evaluate the effectiveness of our model and different modalities.

Personalized recommendation of micro-videos is crucial for many content sharing platforms. Micro-videos typically com-prise fruitful multimodal contents. Existing methods have limitations in learning multimodal representations: (i) They suffer from data sparsity problems as only rely on the inter-action prediction loss to train the whole model. (ii) They do not consider the correlation among multimodal contents into the representation learning. To tackle these limitations, we propose a cross-Modal context enhanced method via Mutual infoMax for Recommendation, termed M3Rec. Specifically, we design the cross-modal graph neural network to inter-change multimodal information to generate modality-aware representations. Then, based on the coupled modality con-text, we design the masked modality prediction (MMP) with three self-supervised objectives to learn the correlation among visual, textual, and acoustic contents empowered by the mu-tual information maximization principle. Finally, we enhance representations via self-supervised pre-training to boost rec-ommendation. Extensive experiments demonstrate that our approach outperforms existing state-of-the-art methods.

Cross-domain recommendation (CDR) technology is proved to be an effective way to tackle the difficulties encountered by traditional recommender technology (e.g. CF), such as data sparsity and cold-start. However, on account of the heterogeneity, it is difficult to enhance the representation of user preferences with the informative knowledge of shared user learned from auxiliary domain. In this paper, we propose a CDR method with co-capsule networks based knowledge transfer to implement the recommendation for the cold-start users. Concretely, the model captures the preference of users with a two-tier structure, the attentive GRU is employed to learn the primary intent from item level and the capsule network is used to further refer the user interests in feature level. After studying the mapping matrix, NeuMF is adapted to execute the recommendation task. We conduct extensive experiments on public datasets and the results demonstrate that the proposed model outperforms many state-of-the-art models.