Large Language Models (LLMs) have achieved impressive reasoning abilities, but struggle with temporal understanding, especially when questions involve multiple entities, compound operators, and evolving event sequences. Temporal Knowledge Graphs (TKGs), which capture vast amounts of temporal facts in a structured format, offer a reliable source for temporal reasoning. However, existing TKG-based LLM reasoning methods still struggle with four major challenges: maintaining temporal faithfulness in multi-hop reasoning, achieving multi-entity temporal synchronization, adapting retrieval to diverse temporal operators, and reusing prior reasoning experience for stability and efficiency. To address these issues, we propose MemoTime, a memory-augmented temporal knowledge graph framework that enhances LLM reasoning through structured grounding, recursive reasoning, and continual experience learning. MemoTime decomposes complex temporal questions into a hierarchical Tree of Time, enabling operator-aware reasoning that enforces monotonic timestamps and co-constrains multiple entities under unified temporal bounds. A dynamic evidence retrieval layer adaptively selects operator-specific retrieval strategies, while a self-evolving experience memory stores verified reasoning traces, toolkit decisions, and sub-question embeddings for cross-type reuse. Comprehensive experiments on multiple temporal QA benchmarks show that MemoTime achieves overall state-of-the-art results, outperforming the strong baseline by up to 24.0%. Furthermore, MemoTime enables smaller models (e.g., Qwen3-4B) to achieve reasoning performance comparable to that of GPT-4-Turbo.

The inclusion of the images opens up a security vulnerability of visually-aware recommender systems (VARSs). It can be exploited by unscrupulous parties to upload well-crafted adversarial images for certain malicious purposes (e.g., promoting their own products for profits). Some studies have focused on attacking VARSs to gain insights into their robustness, while they are still far from practical, i.e., the attacks often 1) lack diversity in perturbations, 2) are easily perceived and 3) have limited transferability, which may lead to overestimation of defenses in practice. To tackle the problems, we propose to perturb the style of the product, which is an unnoticeable but important property of visual recommendations. Specifically, we propose a novel Style perturbation-based Practical Attack Framework (SPAF). Unlike existing attacks that change pixels within l∞ -norm constraints, SPAF interferes with styles in latent feature space so that the attack becomes unbounded in the pixel space to reflect possible actual perturbations. SPAF formulates attack objectives as an optimization problem and adopts an adaptive adversarial style transfer network to solve it so that transferable and imperceptible attacks can be generated. Comprehensive experiments on real-world datasets demonstrate that SPAF significantly outperforms state-of-the-art attacks.

The robustness of large language models (LLMs) becomes increasingly important as their use rapidly grows in a wide range of domains.Retrieval-Augmented Generation (RAG) is considered as a means to improve the trustworthiness of text generation from LLMs.However, how the outputs from RAG-based LLMs are affected by slightly different inputs is not well studied.In this work, we find that the insertion of even a short prefix to the prompt leads to the generation of outputs far away from factually correct answers.We systematically evaluate the effect of such prefixes on RAG by introducing a novel optimization technique called Gradient Guided Prompt Perturbation (GGPP).GGPP achieves a high success rate in steering outputs of RAG-based LLMs to targeted wrong answers.It can also cope with instructions in the prompts requesting to ignore irrelevant context.We also exploit LLMs' neuron activation difference between prompts with and without GGPP perturbations to give a method that improves the robustness of RAG-based LLMs through a highly effective detector trained on neuron activation triggered by GGPP generated prompts.Our evaluation on open-sourced LLMs demonstrates the effectiveness of our methods.

The field of generating recommendations within the framework of causal inference has seen a recent surge.This approach enhances insights into the influence of recommendations on user behavior and helps in identifying the underlying factors. Existing research has often leveraged propensity scores to mitigate bias, albeit at the risk of introducing additional variance. Others have explored the use of unbiased data from randomized controlled trials, although this comes with assumptions that may prove challenging in practice. In this paper, we first present the causality-aware interpretation of recommendations and reveal how the underlying exposure mechanism can bias the maximum likelihood estimation (MLE) of observational feedback. Recognizing that confounders may be elusive, we propose a contrastive self-supervised learning to minimize exposure bias, employing inverse propensity scores and expanding the positive sample set. Building on this foundation, we present a novel contrastive counterfactual learning method (CCL) that incorporates three unique positive sampling strategies grounded in estimated exposure probability or random counterfactual samples. Through extensive experiments on two real-world datasets, we demonstrate that our CCL outperforms the state-of-the-art methods.

Recent studies have shown that recommender systems (RSs) are highly vulnerable to data poisoning attacks. Understanding attack tactics helps improve the robustness of RSs. We intend to develop efficient attack methods that use limited resources to generate high-quality fake user profiles to achieve 1) transferability among black-box RSs 2) and imperceptibility among detectors. In order to achieve these goals, we introduce textual reviews of products to enhance the generation quality of the profiles. Specifically, we propose a novel attack framework named R-Trojan, which formulates the attack objectives as an optimization problem and adopts a tailored transformer-based generative adversarial network (GAN) to solve it so that high-quality attack profiles can be produced. Comprehensive experiments on real-world datasets demonstrate that R-Trojan greatly outperforms state-of-the-art attack methods on various victim RSs under black-box settings and show its good imperceptibility.

As the deployment of artificial intelligence (AI) is changing many fields and industries, there are concerns about AI systems making decisions and recommendations without adequately considering various ethical aspects, such as accountability, reliability, transparency, explainability, contestability, privacy, and fairness. While many sets of AI ethics principles have been recently proposed that acknowledge these concerns, such principles are high-level and do not provide tangible advice on how to develop ethical and responsible AI systems. To gain insight on the possible implementation of the principles, we conducted an empirical investigation involving semi-structured interviews with a cohort of AI practitioners. The salient findings cover four aspects of AI system design and development, adapting processes used in software engineering: (i) high-level view, (ii) requirements engineering, (iii) design and implementation, (iv) deployment and operation.

With the widespread accumulation of observational data, researchers obtain a new direction to learn counterfactual effects in many domains (e.g., health care and computational advertising) without Randomized Controlled Trials (RCTs). However, observational data suffer from inherent missing counterfactual outcomes and distribution discrepancy between treatment and control groups due to behaviour preference. Motivated by recent advances in representation learning in domain adaptation, we propose a novel framework based on Cycle-Balanced REpresentation learning for counterfactual inference (CBRE) to solve the above problems. Specifically, we realize a robust and balanced representation for different groups using adversarial training. Meanwhile, we construct an information loop that preserves original data properties cyclically, reducing information loss when transforming data into latent representation space. Experimental results on three real-world datasets demonstrate that CBRE matches/outperforms the state-of-the-art methods, and it has a great potential to be applied to counterfactual inference.

Deep reinforcement learning enables an agent to capture users' interest through dynamic interactions with the environment. It uses a reward function to learn user's interest and to control the learning process, attracting great interest in recommendation research. However, most reward functions are manually designed; they are either too unrealistic or imprecise to reflect the variety, dimensionality, and non-linearity of the recommendation problem. This impedes the agent from learning an optimal policy in highly dynamic online recommendation scenarios. To address the above issue, we propose a generative inverse reinforcement learning approach that avoids the need of defining an elaborative reward function. In particular, we model the recommendation problem as an automatic policy learning problem. We first generate policies based on observed users' preferences and then evaluate the learned policy by a measurement based on a discriminative actor-critic network. We conduct experiments on an online platform, VirtualTB, and demonstrate the feasibility and effectiveness of our proposed approach via comparisons with several state-of-the-art methods.

Deep learning frameworks generally require sufficient training data to generalize well while fail to adapt on small or few-shot datasets. Meta-learning offers an effective means of tackling few-shot scenarios and has drawn increasing attention in recent years. Meta-optimization aims to learn a shared set of parameters across tasks for meta-learning while facing challenges in determining whether an initialization condition can be generalized to tasks with diverse distributions. In this regard, we propose a meta-gradient boosting framework that can fit diverse distributions based on a base learner (which learns shared information across tasks) and a series of gradient-boosted modules (which capture task-specific information). We evaluate the model on several few-shot learning benchmarks and demonstrate the effectiveness of our model in modulating task-specific meta-learned priors and handling diverse distributions.

A common challenge for most current recommender systems is the cold-start problem. Due to the lack of user-item interactions, the fine-tuned recommender systems are unable to handle situations with new users or new items. Recently, some works introduce the meta-optimization idea into the recommendation scenarios, i.e. predicting the user preference by only a few of past interacted items. The core idea is learning a global sharing initialization parameter for all users and then learning the local parameters for each user separately. However, most meta-learning based recommendation approaches adopt model-agnostic meta-learning for parameter initialization, where the global sharing parameter may lead the model into local optima for some users. In this paper, we design two memory matrices that can store task-specific memories and feature-specific memories. Specifically, the feature-specific memories are used to guide the model with personalized parameter initialization, while the task-specific memories are used to guide the model fast predicting the user preference. And we adopt a meta-optimization approach for optimizing the proposed method. We test the model on two widely used recommendation datasets and consider four cold-start situations. The experimental results show the effectiveness of the proposed methods.

Sharing a pre-trained machine learning model, particularly a deep neural network via prediction APIs, is becoming a common practice on machine learning as a service (MLaaS) platforms nowadays. Although deep neural networks (DNN) have shown remarkable successes in many tasks, they are also criticized for the lack of interpretability and transparency. Interpreting a shared DNN model faces two additional challenges compared with interpreting a general model. (1) Limited training data can be disclosed to users. (2) The internal structure of the models may not be available. These two challenges impede the application of most existing interpretability approaches, such as saliency maps or influence functions, for DNN models. Case-based reasoning methods have been used for interpreting decisions; however, how to select and organize the data points under the constraints of shared DNN models is not discussed. Moreover, simply providing cases as explanations may not be sufficient for supporting instance level interpretability. Meanwhile, existing interpretation methods for DNN models generally lack the means to evaluate the reliability of the interpretation. In this article, we propose a framework named Shared Model INTerpreter (SMINT) to address the above limitations. We propose a new data structure called a boundary graph to organize training points to mimic the predictions of DNN models. We integrate local features, such as saliency maps and interpretable input masks, into the data structure to help users to infer the model decision boundaries. We show that the boundary graph is able to address the reliability issues in many local interpretation methods. We further design an algorithm named hidden-layer aware p-test to measure the reliability of the interpretations. Our experiments show that SMINT is able to achieve above 99% fidelity to corresponding DNN models on both MNIST and ImageNet by sharing only a tiny fraction of training data to make these models interpretable. The human pilot study demonstrates that SMINT provides better interpretability compared with existing methods. Moreover, we demonstrate that SMINT is able to assist model tuning for better performance on different user data.

Despite outperforming humans in many tasks, deep neural network models are also criticized for the lack of transparency and interpretability in decision making. The opaqueness results in uncertainty and low confidence when deploying such a model in model sharing scenarios, where the model is developed by a third party. For a supervised machine learning model, sharing training process including training data is a way to gain trust and to better understand model predictions. However, it is not always possible to share all training data due to privacy and policy constraints. In this paper, we propose a method to disclose a small set of training data that is just sufficient for users to get the insight into a complicated model. The method constructs a boundary tree using selected training data and the tree is able to approximate the complicated deep neural network models with high fidelity. We show that data point pairs in the tree give users significantly better understanding of the model decision boundaries and paves the way for trustworthy model sharing.

Over the past years, frameworks such as MapReduce and Spark have been introduced to ease the task of developing big data programs and applications. However, the jobs in these frameworks are roughly defined and packaged as executable jars without any functionality being exposed or described. This means that deployed jobs are not natively composable and reusable for subsequent development. Besides, it also hampers the ability for applying optimizations on the data flow of job sequences and pipelines. In this paper, we present the Hierarchically Distributed Data Matrix (HDM) which is a functional, strongly-typed data representation for writing composable big data applications. Along with HDM, a runtime framework is provided to support the execution of HDM applications on distributed infrastructures. Based on the functional data dependency graph of HDM, multiple optimizations are applied to improve the performance of executing HDM jobs. The experimental results show that our optimizations can achieve improvements of between 10% to 60% of the Job-Completion-Time for different types of operation sequences when compared with the current state of art, Apache Spark.