In the context of information explosion, multimodal recommender systems (MMRS) have demonstrated great potential in capturing users' complex preferences and enhancing recommendation performance by integrating multimodal data such as images and text. However, multimodal data inherently suffers from semantic inconsistency, which can introduce information conflicts or noise. Moreover, users' reliance on different modalities varies dynamically with context and time (multimodal dynamic preferences). These challenges may lead to truth deviation and deep semantic mismatch, ultimately degrading recommendation performance. To address these issues, we propose Dual-Path Multimodal Recommendation (DPRec), a novel model that improves precision and robustness of the recommendation through cross-modal information compensation and dynamic modal preference learning. Specifically, DPRec first employs a cross-modal attention mechanism to dynamically model inter-modal correlations, effectively exploring complementary and shared features for robust user and item representations. Second, it integrates feature projection, modality alignment, and dynamic weighting mechanisms to adaptively adjust modality importance based on user context, ensuring flexibility in handling preference dynamics. Lastly, a modality contrastive loss is utilized to maximize mutual information between modalities, mitigating semantic mismatch by enhancing deep collaborative representations. Extensive experiments on three public datasets show that DPRec consistently outperforms state-of-the-art (SOTA) methods, achieving average improvements of 3.94% in Recall@20 and 3.84% in NDCG@20. Our code is publicly available at: https://anonymous.4open.science/r/DPRec-4D15.

Composed image retrieval is designed to more accurately retrieve target images that align with user intentions by using a combination of reference images and descriptive modification texts. However, existing methods primarily focus on designing complex feature fusion networks while neglecting the prevalent issues of noise and inconsistent sample quality in training data, leading to insufficient cross-modal semantic alignment and sample relevance modeling. To address this, we propose an innovative Tripartite Alignment Network (TAN) that introduces a momentum distillation mechanism, leveraging the historical knowledge of a teacher network as additional super-vision to guide the optimization of the student network. During the feature encoder fine-tuning stage, we design response-based knowledge distillation and feature-based knowledge distillation techniques, explicitly strengthening modal alignment through composed-target contrastive learning and implicitly promoting modal fusion via composed-target matching learning. In the combiner training stage, we incorporate a lightweight combiner network and employ a cross-entropy-based matching loss function, encouraging high matching scores for relevant image-text pairs and low scores for irrelevant pairs. Extensive experiments on the FashionIQ and Shoes datasets demonstrate that TAN exhibits superior performance compared to existing state-of-the-art methods, with notable improvements in R@10 of +14.03% and +13.09%, respectively. These results affirm the effectiveness of momentum distillation in multimodal learning. Access the source code at https://github.com/Maserhe/TAN.

In recent years, spiking neural networks (SNNs), which originated from the theoretical basis of neuroscience, have attracted neuromorphic computing and brain-like computing due to their advantages, such as neural dynamics and coding mechanism, which are similar to biological neurons. SNNs have become one of the mainstream frameworks in the field of brain-like computing. However, most of the Leaky Integrate-and-Fire (LIF) neuron models currently used by SNNs based on direct training of backpropagation (BP) do not consider the changes in the recurrent connections and the dynamic strength of neuron connections over time. This study presented the LIF neuron model with recurrent connections and a method for dynamically changing the presynaptic currents. Recurrent LIF neurons have an additional cyclic connection compared with classic LIF neurons. Their postsynaptic current stimulates a change in membrane potential at the next time point. Their dynamics were more similar to the activities of biological neurons. We also proposed an efficient and flexible BP training method for recurrent LIF neurons. On the basis of the above methods, we proposed the recurrent SNN with dynamic presynaptic currents based on backpropagation (RDS-BP). We test the proposed RDS-BP on three image data sets (MNIST, Fashion-MNIST and CIFAR-10) and two text data sets (IMDB and TREC). The results showed that the performance of RDS-BP not only exceeded the naive SNN models based on BP but also exceeded the SNN methods proposed in previous studies in recent years, which had excellent performance in previous experiments. Our work provides a new LIF neuron model with a recurrent connection and dynamic presynaptic current and a BP training arrangement for the proposed neuron, which could merit developments with neuromorphic and brain-like computing.

Embedding items directly into a common feature space, and then measuring the similarity by calculating the feature distance in this space, has become the main method for similarity learning in current fashion retrieval tasks. The method is simple and efficient, but it ignores the correlation among fashion attributes and the impact of these correlations on the feature space, thereby reducing the accuracy of retrieval. Since the number of fashion attributes is large and the semantic granularity is also different, how to capture the relationship between fashion attributes and perform refined embedding to accurately represent fashion items is a challenge. In this paper, by constructing an attribute tree, we propose a hierarchical attribute embedding method for representing fashion items to enhance the relationship between attributes and use masking technology to disentangle different attributes. Based on these modules, we propose a hierarchical attribute-aware embedding network (HAEN) which takes images and attributes as input, learns multiple attribute-specific embedding spaces, and measures fine-grained similarity in the corresponding spaces. The extensive experimental result on two fashion-related public datasets FashionAI and DARN shows the superiority (+5.11% and +3.09% in MAP, respectively) of our proposed HAEN compared with state-of-the-art methods.

Personalized online recommendations strive to adapt their services to individual users by making use of both item and user information. Despite recent progress, the issue of balancing exploitation-exploration (EE) [1] remains challenging. In this paper, we model the personalized online recommendation of e-commence as a two-phase multi-armed bandit problem. This is the first time that “big arm” and “small arm” are introduced into multi-armed bandit (MAB), and a two-stage strategy is adopted to provide target users with the most suitable recommendation list. In the first phase, MAB is used to obtain an item subset that users may be interested in from a large number of items. We use item categories as arms instead of individual items in existing related models to control the arm scale and reduce computational complexity. In the second phase, we directly use the items generated in the first phase as arms of MAB and obtain rewards through fine-grained implicit feedback from users. Empirical studies on three real-world datasets show that our proposed method TPBandit performs better than state-of-the-art bandit-based recommendation methods in several evaluation metrics such as Precision, Recall, and Hit Ratio. Moreover, the two-phase method improves the recommendation performance by nearly 50% compared to the one-phase method in the best case.