Cold-start item recommendation is a significant challenge in recommendation systems, particularly when new items are introduced without any historical interaction data. While existing methods leverage multi-modal content to alleviate the cold-start issue, they often neglect the inherent multi-view structure of modalities, namely the distinction between shared and modality-specific features. In this paper, we propose Multi-Modal Multi-View Variational AutoEncoder (M²VAE), a generative model that addresses the challenges of modeling common and unique views in attribute and multi-modal features, as well as user preferences over single-typed item features. Specifically, we generate type-specific latent variables for item IDs, categorical attributes, and image features, and use Product-of-Experts (PoE) to derive a common representation. A disentangled contrastive loss decouples the common view from unique views while preserving feature informativeness. To model user inclinations, we employ a user-aware hierarchical Mixture-of-Experts (MoE) to adaptively fuse representations. We further incorporate co-occurrence signals via contrastive learning, eliminating the need for pretraining. Extensive experiments on real-world datasets validate the effectiveness of our approach.

In advertising recommendation systems, click-through rate (CTR) prediction is a critical task. Capturing users' interests from their rich historical behaviors is key to improving prediction results. Although traditional deep learning methods can capture users' interests to some extent, they fail to account for the local and global interests reflected in users' historical behaviors and the dynamic relationships between them. In this paper, we propose a novel architecture - Dual-Interest Adaptive Network (DIAN), which adaptively extracts both local and global interests of users. Specifically, to better explore users' interests in depth, we propose an Adaptive Interest Extraction Block applied to users' historical behavior sequences. By introducing an attention mechanism, this module can flexibly allocate weights to users' local and global interests after decoupling user behaviors. Additionally, to capture complex feature interactions, our model introduces two feature extractors: one combines a Multi-Layer Perceptron (MLP) with a Cross Network for high-order feature extraction, and the other incorporates an Attention Factorization Machine (AFM) for low-order feature extraction. We conducted extensive experiments on the Movielens-1M and Amazon Electronics datasets, validating the effectiveness of DIAN.

Sequential recommendation (SR) aims to predict the next purchasing item according to users' dynamic preference learned from their historical user-item interactions. To improve the performance of recommendation, learning dynamic heterogeneous cross-type behavior dependencies is indispensable for recommender system. However, there still exists some challenges in Multi-Behavior Sequential Recommendation (MBSR). On the one hand, existing methods only model heterogeneous multi-behavior dependencies at behavior-level or item-level, and modeling interaction-level dependencies is still a challenge. On the other hand, the dynamic multi-grained behavior-aware preference is hard to capture in interaction sequences, which reflects interaction-aware sequential pattern. To tackle these challenges, we propose a Multi-Grained Preference enhanced Transformer framework (M-GPT). First, M-GPT constructs an interaction-level graph of historical cross-typed interactions in a sequence. Then graph convolution is performed to derive interaction-level multi-behavior dependency representation repeatedly, in which the complex correlation between historical cross-typed interactions at specific orders can be well learned. Secondly, a novel multifaceted transformer architecture equipped with multi-grained user preference extraction is proposed to encode the interaction-aware sequential pattern enhanced by capturing temporal behavior-aware multi-grained preference . Experiments on the real-world datasets indicate that our method M-GPT consistently outperforms various state-of-the-art recommendation methods. Our code is available at: https://github.com/hchchchchchchc/MGPT.

Learning effective image representation and constructing a suitable metric space are two main challenges in few-shot image classification. Existing methods normally consider the joint characteristic distribution of the image to improve the image representation ability, but it also brings high computational cost and produces high-dimensional embedding vectors, which limit their real-world applicability. In this paper, in order to reduce computational complexity and embedding dimension, we propose an effectively low-dimensional feature representation module (LDFR), which introduces a window mechanism to make the model focus on the correlation between local channels by using Brownian Distance Covariance. Furthermore, we combine LDFR with Hybrid Attention (LDFR-HA) to solve the problem of unbalanced sample features in few-shot image classification. Specifically, weighted average is adopted in the Hybrid Attention to construct the metric space and perform classification. Numerous experiments are conducted on two standard few-shot image classification benchmarks, i.e., general object recognition and fine-grained categorization. Extensive evaluations demonstrate that LDFR-HA significantly outperforms existing approaches. On popular datasets miniImageNet and CUB, LDFR-HA achieves 3.52 percentage point(pp)/2.29pp and 2.1pp/1.45pp gains over the state-of-the-art method on 5-way 1-shot/5-shot tasks, respectively.

Click-Through Rate(CTR) prediction aims to predict the possibility of users clicking on products, which has become the core task of advertising recommendation systems. Due to the richness of user historical behavior, a key to making effective prediction is to capture users' diverse interests from historical behavior. An efficient way to do this is to perform dot product of behavior and target embedding with attentive neural networks. To better model the users' diverse interests, our proposed disentangled interest extraction block decouples the unary terms modeling the impact of user behavior sequence from pairwise interactions. Specifically, the decoupled pairwise term can learn the pure pairwise interactions, whereas the unary term models the impact of behavior sequence on each target items. Meanwhile, our model emphasizes both high- and low-order feature interactions by combining Attentional Factorization Machines(AFMs) with deep learning. This work intends to accomplish our goal by proposing a novel architecture Deep Disentangled Interest Network(DDIN). We conduct comprehensive experiments on Movielens dataset and Amazon electronic dataset. The results demonstrate the effectiveness of DDIN which is superior to some state-of-art models by up to 26.271 %.