Generative recommendation is emerging as a transformative paradigm by directly generating recommended items, rather than relying on matching. Building such a system typically involves two key components: (1) optimizing the tokenizer to derive suitable item identifiers, and (2) training the recommender based on those identifiers. Existing approaches often treat these components separately—either sequentially or in alternation—overlooking their interdependence. This separation can lead to misalignment: the tokenizer is trained without direct guidance from the recommendation objective, potentially yielding suboptimal identifiers that degrade recommendation performance. To address this, we propose BLOGER, a Bi-Level Optimization for GEnerative Recommendation framework, which explicitly models the interdependence between the tokenizer and the recommender in a unified optimization process. The lower level trains the recommender using tokenized sequences, while the upper level optimizes the tokenizer based on both the tokenization loss and recommendation loss. We adopt a meta-learning approach to solve this bi-level optimization efficiently, and introduce gradient surgery to mitigate gradient conflicts in the upper-level updates, thereby ensuring that item identifiers are both informative and recommendation-aligned. Extensive experiments on multiple real-world datasets demonstrate that BLOGER consistently outperforms state-of-the-art generative recommendation methods while maintaining practical efficiency with no significant additional computational overhead, effectively bridging the gap between item tokenization and autoregressive generation. We release our code at https://github.com/Ten-Mao/BLOGER.

Collaborative filtering (CF) is the basic method for recommendation with implicit feedback. Recently, various state-of-the-art CF integrates graph neural networks. However, they often suffer from popularity bias, causing recommendations to deviate from users' genuine preferences. Additionally, several contrastive learning methods based on the in-batch sample strategy have been proposed to train the CF model effectively, but they are prone to suffering from sample bias. To address this problem, debiased contrastive loss has been employed in the recommendation, but instead of personalized debiasing, it treats each user equally. In this paper, we propose a popularity-aware debiased contrastive loss for CF, which can adaptively correct the positive and negative scores based on the popularity of users and items. Our approach aims to reduce the negative impact of popularity and sample bias simultaneously. We theoretically analyze the effectiveness of the proposed method and reveal the relationship between popularity and gradient, which justifies the correction strategy. We extensively evaluate our method on three public benchmarks over balanced and imbalanced settings. The results demonstrate its superiority over the existing debiased strategies, not only on the entire datasets but also when segmenting the datasets based on item popularity.

Self-attention, which allows transformers to capture deep bidirectional contexts, plays a vital role in BERT-like pre-trained language models. However, the maximum likelihood pre-training objective of BERT may produce an anisotropic word embedding space, which leads to biased attention scores for high-frequency tokens, as they are very close to each other in representation space and thus have higher similarities. This bias may ultimately affect the encoding of global contextual information. To address this issue, we propose TOCOL, a TOken-Level COntrastive Learning framework for improving the contextual representation of pre-trained language models, which integrates a novel self-supervised objective to the attention mechanism to reshape the word representation space and encourages PLM to capture the global semantics of sentences. Results on the GLUE Benchmark show that TOCOL brings considerable improvement over the original BERT. Furthermore, we conduct a detailed analysis and demonstrate the robustness of our approach for low-resource scenarios.

Centrality measures are widely used to map each node to its importance in a network. For many practical applications, vital nodes bearing high centrality scores have superior positions over other nodes. To benefit from the positive impact of becoming a vital node, the problem of improving the centrality of the target node has attracted increasing attention. Many existing studies attack this problem by directly increasing the centrality score of the target node on the premise of knowing the network structure. However, these methods suffer from privacy issues due to their dependence on the network structure and may lose their effectiveness because other nodes can simultaneously increase the scores. Therefore, in this paper, we explore the following question: given a black-box network whose structure is unknown, is it possible to improve the centrality ranking (rather than the score) of a target node by implementing certain strategies? We provide an affirmative answer to this question. First, to avoid relying on the network structure for promotion, we propose strategies that freeze the original graph while appending nodes and edges just around the target node. Second, to guide strategies for effectively boosting centrality, we devise two principles that provide the target node with either the maximum gain or the minimum loss of centrality scores over other nodes. We prove that a strategy meeting the proposed principles is guaranteed to upgrade the target node's ranking. Extensive experiments were conducted to verify the effectiveness of the proposed strategies on black-box networks.

As one of the core components of customer service bot, User Intent Prediction (UIP) aims at predicting users? intents (usually represented as predefined user questions) before they ask, and has been widely applied in real applications. However, when developing a machine learning system for this problem, two critical issues, i.e., the problem of feature drift and class imbalance, may emerge and seriously deprave the system performance. Moreover, various scenarios may arise due to business demands, making the aforementioned problems much more severe. To address these two problems, we propose an attention-based Deep Multi-instance Sequential Cross Network (aDMSCN) to deal with the UIP task. On the one hand,the UIP task can be subtly formalized as multi-instance learning(MIL) task with an attention-based method proposed to alleviate the influences of feature drift. To the best of our knowledge, this is the first attempt to model the problem from a MIL perspective.On the other hand, a ratio-sensitive loss is also developed in our model, which can mitigate the negative impact of class imbalance. Extensive experiments on both offline real-world datasets and on-line A/B testing show that our proposed framework significantly out performs other state-of-art methods for the UIP task.

Currently the collaborative filtering based recommender system has become more and more indispensable due to its capability in providing users with personalised suggestions. Despite its advances in term of efficiency, easy implementation and robustness, traditional collaborative filtering techniques suffer from several challenges such as cold-start and data sparsity. To overcome these limitations, external information is expected to help improve the overall effectiveness. Among the diverse context information, trust relationships is a widely utilised mechanism. Meanwhile, researchers also found distrust relationships is unavoidable in social network and recommender systems can benefit from distrust information. However, most existed distrusted oriented methods do not take the property of multi-facets in distrust relationships into consideration. In this paper, we exploit distrust relationships in a multi-faceted perspective and proposed a matrix factorization based model with integration of different distrust relationship of quality user between different people. Experimental study on well-known dataset has shown promising result and it is expected that this work could provide insight for researchers in this domain to further discuss the distrust in recommender systems.