Most recent paradigms of generative model-based recommendation still face challenges related to the cold-start problem. Existing models addressing cold item recommendations mainly focus on acquiring more knowledge to enrich embeddings or model inputs. However, many models do not assess the efficiency with which they utilize the available training knowledge, leading to the extraction of significant knowledge that is not fully used, thus limiting improvements in cold-start performance. To address this, we introduce the concept of epistemic uncertainty (which refers to uncertainty caused by a lack of knowledge of the best model) to indirectly define how efficiently a model uses the training knowledge. Since epistemic uncertainty represents the reducible part of the total uncertainty, we can optimize the recommendation model further based on epistemic uncertainty to improve its performance. To this end, we propose a Cold-Start Recommendation based on Epistemic Uncertainty (CREU) framework. Additionally, CREU is inspired by Pairwise-Distance Estimators (PaiDEs) to efficiently and accurately measure epistemic uncertainty by evaluating the mutual information between model outputs and weights in high-dimensional spaces. The proposed method is evaluated through extensive offline experiments on public datasets, which further demonstrate the advantages and robustness of CREU. The source code is available at https://github.com/EsiksonX/CREU.

Cross-domain recommendation aims to leverage heterogeneous information to transfers knowledge from a data-sufficient domain (source domain) to a data-scarce domain (target domain). Existing approaches mainly focus on learning single-domain user preferences and then employ a transferring module to obtain cross-domain user preferences, but ignore the modeling of users' domain specific preferences on items. We argue that incorporating domain-specific preferences from the source domain will introduce irrelevant information that fails to the target domain. Additionally, directly combining domain-shared and domain-specific information may hinder the target domain's performance. To this end, we propose C^2DR, a novel approach that disentangles domain-shared and domain-specific preferences from a causal perspective. Specifically, we formulate a causal graph to capture the critical causal relationships based on the underlying recommendation process, explicitly identifying domain-shared and domain-specific information as causal irrelevant variables. Then, we introduce disentanglement regularization terms to learn distinct representations of the causal variables that obey the independence constraints in the causal graph. Remarkably, our proposed method enables effective intervention and transfer of domain-shared information, thereby improving the robustness of the recommendation model. We evaluate the efficacy of C^2DR through extensive experiments on three real-world datasets, demonstrating significant improvements over state-of-the-art baselines.

The cross-domain recommendation (CDR) model addresses challenges such as data sparsity, the long tail distribution of user-item interactions, and the cold start of items or users. However, solely transferring domain-shared knowledge based on the co-occurrence patterns, without considering user preferences, leads to negative transfer in CDR. To overcome these limitations, we propose an advanced deep learning CDR model called the Domain Adversarial Deep Interest Network (DADIN) aims to facilitate smooth knowledge transfer from the source domain to the target domain and effectively alleviate negative transfer. Firstly, the joint distribution alignment of user preference in DADIN is realized by introducing a skip-connection-based domain agnostic layer, and then the domain classifier is artificially designed to distinguish between the information coming from the source domain or the target domain. Additionally, DADIN combines prediction loss, global domain confusion loss, and intra-class domain confusion losses through the Min-Max game and gradient reverse layer to achieve collaborative optimization. Two real-world experiments show the area under curve (AUC) of DADIN is 0.78 on the Huawei dataset, and it outperforms its competitors by 0.71% on the Amazon dataset, showcasing its state-of-the-art performance. Moreover, our ablation studies further demonstrate that domain adversarial technique increases the AUC by 2.34% on the Huawei dataset and 16.67% on the Amazon dataset, respectively.

Recommendation system algorithm based on multi-task learning (MTL) is the major method for Internet operators to understand users and predict their behaviors in the multi-behavior scenario of platform. Task correlation is an important consideration of MTL goals, traditional models use shared-bottom models and gating experts to realize shared representation learning and information differentiation. However, The relationship between real-world tasks is often more complex than existing methods do not handle properly sharing information. In this paper, we propose an Different Expression Parallel Heterogeneous Network (DEPHN) to model multiple tasks simultaneously. DEPHN constructs the experts at the bottom of the model by using different feature interaction methods to improve the generalization ability of the shared information flow. In view of the model's differentiating ability for different task information flows, DEPHN uses feature explicit mapping and virtual gradient coefficient for expert gating during the training process, and adaptively adjusts the learning intensity of the gated unit by considering the difference of gating values and task correlation. Extensive experiments on artificial and real-world datasets demonstrate that our proposed method can capture task correlation in complex situations and achieve better performance than baseline models.

There are two fundamental problems in applying deep learning/machine learning methods to disease classification tasks, one is the insufficient number and poor quality of training samples; another one is how to effectively fuse multiple source features and thus train robust classification models. To address these problems, inspired by the process of human learning knowledge, we propose the Feature-aware Fusion Correlation Neural Network (FaFCNN), which introduces a feature-aware interaction module and a feature alignment module based on domain adversarial learning. This is a general framework for disease classification, and FaFCNN improves the way existing methods obtain sample correlation features. The experimental results show that training using augmented features obtained by pre-training gradient boosting decision tree yields more performance gains than random-forest based methods. On the low-quality dataset with a large amount of missing data in our setup, FaFCNN obtains a consistently optimal performance compared to competitive baselines. In addition, extensive experiments demonstrate the robustness of the proposed method and the effectiveness of each component of the model.