Federated recommendation (FR) facilitates collaborative training by aggregating local models from massive devices, enabling client-specific personalization while ensuring privacy. However, we empirically and theoretically demonstrate that server-side aggregation can undermine client-side personalization, leading to suboptimal performance, i.e., the aggregation bottleneck. This issue stems from the inherent heterogeneity across numerous clients in FR, which drives the global model to deviate from local optima. To this end, we propose FedEM, which elastically merges the global and local models to compensate for impaired personalization. Unlike existing personalized federated recommendation (pFR) methods, FedEM (1) investigates the aggregation bottleneck in FR through theoretical insights, rather than relying on heuristic analysis; (2) leverages off-the-shelf local models rather than designing additional mechanisms to boost personalization. Extensive experiments demonstrate that our method preserves client personalization during collaborative training, outperforming state-of-the-art baselines.

Causal Inference (CI) plays a crucial role in building unbiased recommender systems. However, most current CI-based debiasing methods only pay attention on either popularity bias or conformity bias. This paper presents a Disentangled Counterfactual Reasoning framework to alleviate dual biases in recommendation, so called DCR. Concretely, we consider the impact of both item popularity and user conformity during training, and separate their indirect effects by disentangling user and item embeddings into biased and unbiased components. In the inference stage, we perform counterfactual reasoning to simultaneously mitigate the indirect and direct effects of bias factors. Experimental results demonstrate the effectiveness of our DCR.

Large language models (LLMs) have been widely applied in recommender systems, achieving remarkable success. However, LLM-based recommendation (LR) suffers from more severe popularity bias than conventional recommendation (CR), stemming from both training and inference stages. In this paper, we propose a novel debiasing method for LR, which performs debiasing in such two stages, so termed as Dual Debiasing in LR (D²LR). Concretely, in the training stage, we conduct token-wise inverse propensity score weighting to force the LLM to pay more attention on unpopular tokens. In the inference stage, we train a more biased CR model by increasing the weights of popular items, which adjusts the generation probability of corresponding tokens according to its scores for items, hoping to suppress the excessive generation of popular tokens. Experiments conducted on three real-world datasets validate the effectiveness of our D²LR in mitigating popularity bias in LR.

Hashing-based recommendation (HR) methods, whose core idea is mapping users and items into hamming space, are common practice to improve item retrieval efficiency. However, existing HR fails to align optimization objective (i.e., Bayesian Personalized Ranking) and evaluation metric (i.e., Recall), leading to suboptimal performance. In this paper, we propose a smooth recall loss (termed as SRLoss), which targets Recall as the optimization objective. Due to the existence of discrete constraints, the optimization problem is NP-hard. To this end, we propose an approximation-adjustable gradient estimator to solve our problem. Experimental Results demonstrate the effectiveness of our proposed method.

Review-aware Rating Regression (RaRR) suffers the severe challenge of extreme data sparsity as the multi-modality interactions of ratings accompanied by reviews are costly to obtain. Although some studies of semi-supervised rating regression are proposed to mitigate the impact of sparse data, they bear the risk of learning from noisy pseudo-labeled data. In this article, we propose a simple yet effective paradigm, called co-training-teaching ( CoT 2 ), for integrating the merits of both co-training and co-teaching toward robust semi-supervised RaRR. CoT 2 employs two predictors trained with different feature sets of textual reviews, each of which functions as both “labeler” and “validator.” Specifically, one predictor (labeler) first labels unlabeled data for its peer predictor (validator); after that, the validator samples reliable instances from the noisy pseudo-labeled data it received and sends them back to the labeler for updating. By exchanging and validating pseudo-labeled instances, the two predictors are reinforced by each other in an iterative learning process. The final prediction is made by averaging the outputs of both the refined predictors. Extensive experiments show that our CoT 2 considerably outperforms the state-of-the-art recommendation techniques in the RaRR task, especially when the training data is severely insufficient.

To perform online inference efficiently, hashing techniques, devoted to encoding model parameters as binary codes, play a key role in reducing the computational cost of content-aware recommendation (CAR), particularly on devices with limited computation resource. However, current hashing methods for CAR fail to align their learning objectives (e.g., squared loss) with the ranking-based metrics (e.g., Normalized Discounted Cumulative Gain (NDCG)), resulting in suboptimal recommendation accuracy. In this article, we propose a novel ranking-based CAR hashing method based on Factorization Machine (FM), called Discrete Listwise FM (DLFM), for fast and accurate recommendation. Concretely, our DLFM is to optimize NDCG in the Hamming space for preserving the listwise user-item relationships. We devise an efficient algorithm to resolve the challenging DLFM problem, which can directly learn binary parameters in a relaxed continuous solution space, without additional quantization. Particularly, our theoretical analysis shows that the optimal solution to the relaxed continuous optimization problem is approximately the same as that of the original discrete optimization problem. Through extensive experiments on two real-world datasets, we show that DLFM consistently outperforms state-of-the-art hashing-based recommendation techniques.

In recommender systems (RSs), inverse propensity score (IPS) has been a key technique to mitigate popularity bias by decreasing the contribution of popular items in modeling user-item interactions. However, conventional IPS treats all users equally, which tends to over-debias the popularity-insensitive (PI) users and under-debias the popularity-sensitive (PS) users. Furthermore, in such a treatment, IPS only performs slightly well on the debiased test while does not work on the normal biased test. To this end, we propose a user-dependent IPS (UDIPS in short) method, which adaptively conducts propensity estimation for each user-item pair based on the user's sensitivity to item popularity. Like IPS, our theoretical analysis validates the unbiasedness of UDIPS. Remarkably, our solution is model-agnostic and can be easily used to upgrade current unbiased recommenders. We implemented it in four state-of-the-art models for unbiased recommendation, and experimental results on two benchmark datasets demonstrate the effectiveness of our method in both unbiased and normal biased test.

Federated recommender system (FRS), which enables many local devices to train a shared model jointly without transmitting local raw data, has become a prevalent recommendation paradigm with privacy-preserving advantages. However, previous work on FRS performs similarity search via inner product in continuous embedding space, which causes an efficiency bottleneck when the scale of items is extremely large. We argue that such a scheme in federated settings ignores the limited capacities in resource-constrained user devices ( i.e. , storage space, computational overhead, and communication bandwidth), and makes it harder to be deployed in large-scale recommender systems. Besides, it has been shown that transmitting local gradients in real-valued form between server and clients may leak users’ private information. To this end, we propose a lightweight federated recommendation framework with privacy-preserving matrix factorization, LightFR , that is able to generate high-quality binary codes by exploiting learning to hash technique under federated settings, and thus enjoys both fast online inference and economic memory consumption. Moreover, we devise an efficient federated discrete optimization algorithm to collaboratively train model parameters between the server and clients, which can effectively prevent real-valued gradient attacks from malicious parties. Through extensive experiments on four real-world datasets, we show that our LightFR model outperforms several state-of-the-art FRS methods in terms of recommendation accuracy, inference efficiency and data privacy.

We address the efficiency problem of personalized ranking from implicit feedback by hashing users and items with binary codes, so that top-N recommendation can be fast executed in a Hamming space by bit operations. However, current hashing methods for top-N recommendation fail to align their learning objectives (such as pointwise or pairwise loss) with the benchmark metrics for ranking quality (e.g. Average Precision, AP), resulting in sub-optimal accuracy. To this end, we propose a Discrete Listwise Personalized Ranking (DLPR) model that optimizes AP under discrete constraints for fast and accurate top-N recommendation. To resolve the challenging DLPR problem, we devise an efficient algorithm that can directly learn binary codes in a relaxed continuous solution space. Specifically, theoretical analysis shows that the optimal solution to the relaxed continuous optimization problem is exactly the same as that of the original discrete DLPR problem. Through extensive experiments on two real-world datasets, we show that DLPR consistently surpasses state-of-the-art hashing methods for top-N recommendation.

Deep neural networks (DNNs) have been a key technique for click-through rate (CTR) estimation, yet existing DNNs-based CTR models neglect the inconsistency between their optimization objectives (e.g., Binary Cross Entropy, BCE) and CTR ranking metrics (e.g., Area Under the ROC Curve, AUC). It is noteworthy that directly optimizing AUC by gradient-descent methods is difficult due to the non-differentiable Heaviside function built-in AUC. To this end, we propose a smooth approximation of AUC, called smooth-AUC (SAUC), towards the rank-based CTR prediction. Specifically, SAUC relaxes the Heaviside function via sigmoid with a temperature coefficient (aiming at controlling the function sharpness) in order to facilitate the gradient-based optimization. Furthermore, SAUC is a plug-and-play objective that can be used in any DNNs-based CTR model. Experimental results on two real-world datasets demonstrate that SAUC consistently improves the recommendation accuracy of current DNNs-based CTR models.

Combining matrix factorization (MF) with network embedding (NE) has been a promising solution to social recommender systems. However, such a scheme suffers from the online predictive efficiency issue due to the ever-growing users and items. In this paper, we propose a novel hashing-based social recommendation model, called semi-discrete socially embedded matrix factorization (S2MF), which leverages the dual advantages of social information for recommendation effectiveness and hashing trick for online predictive efficiency. Experimental results demonstrate the advantages of S2MF over state-of-the-art discrete recommendation models and its real-valued competitors.