Large language models (LLMs) have revolutionized the role of AI, yet also pose potential risks of propagating unethical content. Alignment technologies have been introduced to steer LLMs towards human preference, gaining increasing attention. Despite notable breakthroughs in this direction, existing methods heavily rely on high-quality positive-negative training pairs, suffering from noisy labels and the marginal distinction between preferred and dispreferred response data. Given recent LLMs' proficiency in generating helpful responses, this work pivots towards a new research focus: achieving alignment using solely human-annotated negative samples, preserving helpfulness while reducing harmfulness. For this purpose, we propose Distributional Dispreference Optimization (D$^2$O), which maximizes the discrepancy between the generated responses and the dispreferred ones to effectively eschew harmful information. We theoretically demonstrate that D$^2$O is equivalent to learning a distributional instead of instance-level preference model reflecting human dispreference against the distribution of negative responses. Besides, D$^2$O integrates an implicit Jeffrey Divergence regularization to balance the exploitation and exploration of reference policies and converges to a non-negative one during training. Extensive experiments demonstrate that our method achieves comparable generation quality and surpasses the latest baselines in producing less harmful and more informative responses with better training stability and faster convergence.
Graph Signal Processing (GSP) based recommendation algorithms have recently attracted lots of attention due to its high efficiency. However, these methods failed to consider the importance of various interactions that reflect unique user/item characteristics and failed to utilize user and item high-order neighborhood information to model user preference, thus leading to sub-optimal performance. To address the above issues, we propose a frequency-aware graph signal processing method (FaGSP) for collaborative filtering. Firstly, we design a Cascaded Filter Module, consisting of an ideal high-pass filter and an ideal low-pass filter that work in a successive manner, to capture both unique and common user/item characteristics to more accurately model user preference. Then, we devise a Parallel Filter Module, consisting of two low-pass filters that can easily capture the hierarchy of neighborhood, to fully utilize high-order neighborhood information of users/items for more accurate user preference modeling. Finally, we combine these two modules via a linear model to further improve recommendation accuracy. Extensive experiments on six public datasets demonstrate the superiority of our method from the perspectives of prediction accuracy and training efficiency compared with state-of-the-art GCN-based recommendation methods and GSP-based recommendation methods.
Large Language Models (LLMs) have demonstrated considerable advances, and several claims have been made about their exceeding human performance. However, in real-world tasks, domain knowledge is often required. Low-resource learning methods like Active Learning (AL) have been proposed to tackle the cost of domain expert annotation, raising this question: Can LLMs surpass compact models trained with expert annotations in domain-specific tasks? In this work, we conduct an empirical experiment on four datasets from three different domains comparing SOTA LLMs with small models trained on expert annotations with AL. We found that small models can outperform GPT-3.5 with a few hundreds of labeled data, and they achieve higher or similar performance with GPT-4 despite that they are hundreds time smaller. Based on these findings, we posit that LLM predictions can be used as a warmup method in real-world applications and human experts remain indispensable in tasks involving data annotation driven by domain-specific knowledge.
Click Through Rate (CTR) prediction plays an essential role in recommender systems and online advertising. It is crucial to effectively model feature interactions to improve the prediction performance of CTR models. However, existing methods face three significant challenges. First, while most methods can automatically capture high-order feature interactions, their performance tends to diminish as the order of feature interactions increases. Second, existing methods lack the ability to provide convincing interpretations of the prediction results, especially for high-order feature interactions, which limits the trustworthiness of their predictions. Third, many methods suffer from the presence of redundant parameters, particularly in the embedding layer. This paper proposes a novel method called Gated Deep Cross Network (GDCN) and a Field-level Dimension Optimization (FDO) approach to address these challenges. As the core structure of GDCN, Gated Cross Network (GCN) captures explicit high-order feature interactions and dynamically filters important interactions with an information gate in each order. Additionally, we use the FDO approach to learn condensed dimensions for each field based on their importance. Comprehensive experiments on five datasets demonstrate the effectiveness, superiority and interpretability of GDCN. Moreover, we verify the effectiveness of FDO in learning various dimensions and reducing model parameters. The code is available on \url{https://github.com/anonctr/GDCN}.
Click-through rate (CTR) prediction is widely used in academia and industry. Most CTR tasks fall into a feature embedding \& feature interaction paradigm, where the accuracy of CTR prediction is mainly improved by designing practical feature interaction structures. However, recent studies have argued that the fixed feature embedding learned only through the embedding layer limits the performance of existing CTR models. Some works apply extra modules on top of the embedding layer to dynamically refine feature representations in different instances, making it effective and easy to integrate with existing CTR methods. Despite the promising results, there is a lack of a systematic review and summarization of this new promising direction on the CTR task. To fill this gap, we comprehensively summarize and define a new module, namely \textbf{feature refinement} (FR) module, that can be applied between feature embedding and interaction layers. We extract 14 FR modules from previous works, including instances where the FR module was proposed but not clearly defined or explained. We fully assess the effectiveness and compatibility of existing FR modules through comprehensive and extensive experiments with over 200 augmented models and over 4,000 runs for more than 15,000 GPU hours. The results offer insightful guidelines for researchers, and all benchmarking code and experimental results are open-sourced. In addition, we present a new architecture of assigning independent FR modules to separate sub-networks for parallel CTR models, as opposed to the conventional method of inserting a shared FR module on top of the embedding layer. Our approach is also supported by comprehensive experiments demonstrating its effectiveness.
Large Language Models (LLMs) have made unprecedented breakthroughs, yet their increasing integration into everyday life might raise societal risks due to generated unethical content. Despite extensive study on specific issues like bias, the intrinsic values of LLMs remain largely unexplored from a moral philosophy perspective. This work delves into ethical values utilizing Moral Foundation Theory. Moving beyond conventional discriminative evaluations with poor reliability, we propose DeNEVIL, a novel prompt generation algorithm tailored to dynamically exploit LLMs' value vulnerabilities and elicit the violation of ethics in a generative manner, revealing their underlying value inclinations. On such a basis, we construct MoralPrompt, a high-quality dataset comprising 2,397 prompts covering 500+ value principles, and then benchmark the intrinsic values across a spectrum of LLMs. We discovered that most models are essentially misaligned, necessitating further ethical value alignment. In response, we develop VILMO, an in-context alignment method that substantially enhances the value compliance of LLM outputs by learning to generate appropriate value instructions, outperforming existing competitors. Our methods are suitable for black-box and open-source models, offering a promising initial step in studying the ethical values of LLMs.
The recommendation ecosystem involves interactions between recommender systems(Computer) and users(Human). Orthogonal to the perspective of recommender systems, we attempt to utilize LLMs from the perspective of users and propose a more human-central recommendation framework named RAH, which consists of Recommender system, Assistant and Human. The assistant is a LLM-based and personal proxy for a human to achieve user satisfaction. The assistant plays a non-invasion role and the RAH framework can adapt to different recommender systems and user groups. Subsequently, we implement and evaluate the RAH framework for learning user personalities and proxy human feedback. The experiment shows that (1) using learn-action-critic and reflection mechanisms can lead more aligned personality and (2) our assistant can effectively proxy human feedback and help adjust recommender systems. Finally, we discuss further strategies in the RAH framework to address human-central concerns including user control, privacy and fairness.
Sequential recommendation demonstrates the capability to recommend items by modeling the sequential behavior of users. Traditional methods typically treat users as sequences of items, overlooking the collaborative relationships among them. Graph-based methods incorporate collaborative information by utilizing the user-item interaction graph. However, these methods sometimes face challenges in terms of time complexity and computational efficiency. To address these limitations, this paper presents AutoSeqRec, an incremental recommendation model specifically designed for sequential recommendation tasks. AutoSeqRec is based on autoencoders and consists of an encoder and three decoders within the autoencoder architecture. These components consider both the user-item interaction matrix and the rows and columns of the item transition matrix. The reconstruction of the user-item interaction matrix captures user long-term preferences through collaborative filtering. In addition, the rows and columns of the item transition matrix represent the item out-degree and in-degree hopping behavior, which allows for modeling the user's short-term interests. When making incremental recommendations, only the input matrices need to be updated, without the need to update parameters, which makes AutoSeqRec very efficient. Comprehensive evaluations demonstrate that AutoSeqRec outperforms existing methods in terms of accuracy, while showcasing its robustness and efficiency.
Recommender systems are important for providing personalized services to users, but the vast amount of collected user data has raised concerns about privacy (e.g., sensitive data), security (e.g., malicious data) and utility (e.g., toxic data). To address these challenges, recommendation unlearning has emerged as a promising approach, which allows specific data and models to be forgotten, mitigating the risks of sensitive/malicious/toxic user data. However, existing methods often struggle to balance completeness, utility, and efficiency, i.e., compromising one for the other, leading to suboptimal recommendation unlearning. In this paper, we propose an Interaction and Mapping Matrices Correction (IMCorrect) method for recommendation unlearning. Firstly, we reveal that many collaborative filtering (CF) algorithms can be formulated as mapping-based approach, in which the recommendation results can be obtained by multiplying the user-item interaction matrix with a mapping matrix. Then, IMCorrect can achieve efficient recommendation unlearning by correcting the interaction matrix and enhance the completeness and utility by correcting the mapping matrix, all without costly model retraining. Unlike existing methods, IMCorrect is a whitebox model that offers greater flexibility in handling various recommendation unlearning scenarios. Additionally, it has the unique capability of incrementally learning from new data, which further enhances its practicality. We conducted comprehensive experiments to validate the effectiveness of IMCorrect and the results demonstrate that IMCorrect is superior in completeness, utility, and efficiency, and is applicable in many recommendation unlearning scenarios.