Chung-ju Huang, Huiqiang Zhao, Yuanpeng He, Lijian Li, Wenpin Jiao, Zhi Jin, Peixuan Chen, Leye Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

In privacy-preserving machine learning with vertically distributed data, private entity alignment methods are used to securely match and utilize features of the same samples. However, existing methods not only risk exposing sample intersections and introducing unnecessary samples but also face a gap in adapting to multi-party scenarios. To address these limitations, we propose Peafowl, a novel multi-party private entity alignment protocol. Peafowlachieves entity alignment among multiple parties through a mapping from original datasets to intersections, termed permutation. This method mitigates intersection disclosure and sample redundancy concerns by avoiding direct use of the intersection. The proposed protocol leverages a cloud server that utilizes secret-shared shuffle to protect the privacy of the permutation, in case of colluding data providers reconstructing intersections. Further, by integrating a seed homomorphic pseudorandom generator, Peafowlavoids the intensive communication of secret sharing and achieves superior runtime performance. Additionally, an offline/online variant is introduced to ensure a linear growth in communication and computation complexity relative to the dataset size by pre-computing permutation calculations. Implemented on a real PPML framework, the protocol demonstrates practical efficiency in various multi-party settings. Experimental results indicate that Peafowl’s overhead is less than 1% of the total training cost, while the offline/online variant achieves approximately a 50% reduction in online runtime. Overall, Peafowloffers an efficient and straightforward solution for multi-party PPML, making it an attractive option for implementation and future improvements.

Session-based recommendation (SBR) methods often rely on user behavior data, which can struggle with the sparsity of session data, limiting performance. Researchers have identified that beyond behavioral signals, rich semantic information in item descriptions is crucial for capturing hidden user intent. While Large Language Models (LLMs) offer new ways to leverage this semantic data, the challenges of session anonymity, short-sequence nature, and high LLM training costs have hindered the development of a lightweight, efficient LLM framework for SBR. To address the above challenges, we propose an LLM-enhanced SBR framework that integrates semantic and behavioral signals from multiple views. This two-stage framework leverages the strengths of both LLMs and traditional SBR models while minimizing training costs. In the first stage, we use multi-view prompts to infer latent user intentions at the session semantic level, supported by an intent localization module to alleviate LLM hallucinations. In the second stage, we align and unify these semantic inferences with behavioral representations, effectively merging insights from both large and small models. Extensive experiments on two real datasets demonstrate that the LLM4SBR framework can effectively improve model performance. We release our codes along with the baselines at https://github.com/tsinghua-fib-lab/LLM4SBR .

Precisely evaluating the effect of new policies (e.g. ad-placement models, recommendation functions, ranking functions) is one of the most important problems for improving interactive systems. The conventional policy evaluation methods rely on online A/B tests, but they are usually extremely expensive and may have undesirable impacts. Recently, Inverse Propensity Score (IPS) estimators are proposed as alternatives to evaluate the effect of new policy with offline logged data that was collected from a different policy in the past. They tend to remove the distribution shift induced by past policy. However, they ignore the distribution shift that would be induced by the new policy, which results in imprecise evaluation. Moreover, their performances rely on accurate estimation of propensity score, which can not be guaranteed or validated in practice. In this paper, we propose a non-parametric method, named Focused Context Balancing (FCB) algorithm, to learn sample weights for context balancing, so that the distribution shift induced by the past policy and new policy can be eliminated respectively. To validate the effectiveness of our FCB algorithm, we conduct extensive experiments on both synthetic and real world datasets. The experimental results clearly demonstrate that our FCB algorithm outperforms existing estimators by achieving more precise and robust results for offline policy evaluation.

Most of previous machine learning algorithms are proposed based on the i.i.d. hypothesis. However, this ideal assumption is often violated in real applications, where selection bias may arise between training and testing process. Moreover, in many scenarios, the testing data is not even available during the training process, which makes the traditional methods like transfer learning infeasible due to their need on prior of test distribution. Therefore, how to address the agnostic selection bias for robust model learning is of paramount importance for both academic research and real applications. In this paper, under the assumption that causal relationships among variables are robust across domains, we incorporate causal technique into predictive modeling and propose a novel Causally Regularized Logistic Regression (CRLR) algorithm by jointly optimize global confounder balancing and weighted logistic regression. Global confounder balancing helps to identify causal features, whose causal effect on outcome are stable across domains, then performing logistic regression on those causal features constructs a robust predictive model against the agnostic bias. To validate the effectiveness of our CRLR algorithm, we conduct comprehensive experiments on both synthetic and real world datasets. Experimental results clearly demonstrate that our CRLR algorithm outperforms the state-of-the-art methods, and the interpretability of our method can be fully depicted by the feature visualization.