Personalized Text Generation with Contrastive Activation Steering

Personalized text generation aims to infer users' writing style preferences from their historical texts and generate outputs that faithfully reflect these stylistic characteristics. Existing solutions primarily adopt two paradigms: retrieval-augmented generation (RAG) and parameter-efficient fine-tuning (PEFT). While these approaches have advanced the field, they suffer from two critical limitations: (1) the entanglement of content semantics and stylistic patterns in historical texts impedes accurate modeling of user-specific writing preferences; and (2) scalability challenges arising from both RAG's inference latency by retrieval operations and PEFT's parameter storage requirements for per user model. To overcome these limitations, we propose StyleVector, a training-free framework that disentangles and represents personalized writing style as a vector in LLM's activation space, enabling style-steered generation during inference without requiring costly retrieval or parameter storage. Comprehensive experiments demonstrate that our framework achieves a significant 8% relative improvement in personalized generation while reducing storage requirements by 1700 times over PEFT method.

Reveal the Mystery of DPO: The Connection between DPO and RL Algorithms

With the rapid development of Large Language Models (LLMs), numerous Reinforcement Learning from Human Feedback (RLHF) algorithms have been introduced to improve model safety and alignment with human preferences. These algorithms can be divided into two main frameworks based on whether they require an explicit reward (or value) function for training: actor-critic-based Proximal Policy Optimization (PPO) and alignment-based Direct Preference Optimization (DPO). The mismatch between DPO and PPO, such as DPO's use of a classification loss driven by human-preferred data, has raised confusion about whether DPO should be classified as a Reinforcement Learning (RL) algorithm. To address these ambiguities, we focus on three key aspects related to DPO, RL, and other RLHF algorithms: (1) the construction of the loss function; (2) the target distribution at which the algorithm converges; (3) the impact of key components within the loss function. Specifically, we first establish a unified framework named UDRRA connecting these algorithms based on the construction of their loss functions. Next, we uncover their target policy distributions within this framework. Finally, we investigate the critical components of DPO to understand their impact on the convergence rate. Our work provides a deeper understanding of the relationship between DPO, RL, and other RLHF algorithms, offering new insights for improving existing algorithms.

Self-supervised Hierarchical Representation for Medication Recommendation

Medication recommender is to suggest appropriate medication combinations based on a patient's health history, e.g., diagnoses and procedures. Existing works represent different diagnoses/procedures well separated by one-hot encodings. However, they ignore the latent hierarchical structures of these medical terms, undermining the generalization performance of the model. For example, "Respiratory Diseases", "Chronic Respiratory Diseases" and "Chronic Bronchiti" have a hierarchical relationship, progressing from general to specific. To address this issue, we propose a novel hierarchical encoder named HIER to hierarchically represent diagnoses and procedures, which is based on standard medical codes and compatible with any existing methods. Specifically, the proposed method learns relation embedding with a self-supervised objective for incorporating the neighbor hierarchical structure. Additionally, we develop the position encoding to explicitly introduce global hierarchical position. Extensive experiments demonstrate significant and consistent improvements in recommendation accuracy across four baselines and two real-world clinical datasets.

CoRA: Collaborative Information Perception by Large Language Model's Weights for Recommendation

Involving collaborative information in Large Language Models (LLMs) is a promising technique for adapting LLMs for recommendation. Existing methods achieve this by concatenating collaborative features with text tokens into a unified sequence input and then fine-tuning to align these features with LLM's input space. Although effective, in this work, we identify two limitations when adapting LLMs to recommendation tasks, which hinder the integration of general knowledge and collaborative information, resulting in sub-optimal recommendation performance. (1) Fine-tuning LLM with recommendation data can undermine its inherent world knowledge and fundamental competencies, which are crucial for interpreting and inferring recommendation text. (2) Incorporating collaborative features into textual prompts disrupts the semantics of the original prompts, preventing LLM from generating appropriate outputs. In this paper, we propose a new paradigm, CoRA (an acronym for Collaborative LoRA), with a collaborative weights generator. Rather than input space alignment, this method aligns collaborative information with LLM's parameter space, representing them as incremental weights to update LLM's output. This way, LLM perceives collaborative information without altering its general knowledge and text inference capabilities. Specifically, we employ a collaborative filtering model to extract user and item embeddings, converting them into collaborative weights with low-rank properties through the collaborative weights generator. We then merge the collaborative weights into LLM's weights, enabling LLM to perceive the collaborative signals and generate personalized recommendations without fine-tuning or extra collaborative tokens in prompts. Extensive experiments confirm that CoRA effectively integrates collaborative information into LLM, enhancing recommendation performance.

Graph Representation Learning via Causal Diffusion for Out-of-Distribution Recommendation

Graph Neural Networks (GNNs)-based recommendation algorithms typically assume that training and testing data are drawn from independent and identically distributed (IID) spaces. However, this assumption often fails in the presence of out-of-distribution (OOD) data, resulting in significant performance degradation. In this study, we construct a Structural Causal Model (SCM) to analyze interaction data, revealing that environmental confounders (e.g., the COVID-19 pandemic) lead to unstable correlations in GNN-based models, thus impairing their generalization to OOD data. To address this issue, we propose a novel approach, graph representation learning via causal diffusion (CausalDiffRec) for OOD recommendation. This method enhances the model's generalization on OOD data by eliminating environmental confounding factors and learning invariant graph representations. Specifically, we use backdoor adjustment and variational inference to infer the real environmental distribution, thereby eliminating the impact of environmental confounders. This inferred distribution is then used as prior knowledge to guide the representation learning in the reverse phase of the diffusion process to learn the invariant representation. In addition, we provide a theoretical derivation that proves optimizing the objective function of CausalDiffRec can encourage the model to learn environment-invariant graph representations, thereby achieving excellent generalization performance in recommendations under distribution shifts. Our extensive experiments validate the effectiveness of CausalDiffRec in improving the generalization of OOD data, and the average improvement is up to 10.69% on Food, 18.83% on KuaiRec, 22.41% on Yelp2018, and 11.65% on Douban datasets.

Efficient Prompt Tuning by Multi-Space Projection and Prompt Fusion

Prompt tuning is a promising method to fine-tune a pre-trained language model without retraining its large-scale parameters. Instead, it attaches a soft prompt to the input text, whereby downstream tasks can be well adapted by merely learning the embeddings of prompt tokens. Nevertheless, existing methods still suffer from two challenges: (i) they are hard to balance accuracy and efficiency. A longer (shorter) soft prompt generally leads to a better (worse) accuracy but at the cost of more (less) training time. (ii) The performance may not be consistent when adapting to different downstream tasks. We attribute it to the same embedding space but responsible for different requirements of downstream tasks. To address these issues, we propose an Efficient Prompt Tuning method (EPT) by multi-space projection and prompt fusion. Specifically, it decomposes a given soft prompt into a shorter prompt and two low-rank matrices, whereby the number of parameters is greatly reduced as well as the training time. The accuracy is also enhanced by leveraging low-rank matrices and the short prompt as additional knowledge sources to enrich the semantics of the original short prompt. In addition, we project the soft prompt into multiple subspaces to improve the performance consistency, and then adaptively learn the combination weights of different spaces through a gating network. Experimental experiments on 13 natural language processing downstream tasks show that our method significantly and consistently outperforms 11 comparison methods with the relative percentage of improvements up to 28.8%, and training time decreased by 14%.

Towards Unified Modeling for Positive and Negative Preferences in Sign-Aware Recommendation

Recently, sign-aware graph recommendation has drawn much attention as it will learn users' negative preferences besides positive ones from both positive and negative interactions (i.e., links in a graph) with items. To accommodate the different semantics of negative and positive links, existing works utilize two independent encoders to model users' positive and negative preferences, respectively. However, these approaches cannot learn the negative preferences from high-order heterogeneous interactions between users and items formed by multiple links with different signs, resulting in inaccurate and incomplete negative user preferences. To cope with these intractable issues, we propose a novel \textbf{L}ight \textbf{S}igned \textbf{G}raph Convolution Network specifically for \textbf{Rec}ommendation (\textbf{LSGRec}), which adopts a unified modeling approach to simultaneously model high-order users' positive and negative preferences on a signed user-item interaction graph. Specifically, for the negative preferences within high-order heterogeneous interactions, first-order negative preferences are captured by the negative links, while high-order negative preferences are propagated along positive edges. Then, recommendation results are generated based on positive preferences and optimized with negative ones. Finally, we train representations of users and items through different auxiliary tasks. Extensive experiments on three real-world datasets demonstrate that our method outperforms existing baselines regarding performance and computational efficiency. Our code is available at \url{https://anonymous.4open.science/r/LSGRec-BB95}.

Repeated Padding as Data Augmentation for Sequential Recommendation

Sequential recommendation aims to provide users with personalized suggestions based on their historical interactions. When training sequential models, padding is a widely adopted technique for two main reasons: 1) The vast majority of models can only handle fixed-length sequences; 2) Batching-based training needs to ensure that the sequences in each batch have the same length. The special value \emph{0} is usually used as the padding content, which does not contain the actual information and is ignored in the model calculations. This common-sense padding strategy leads us to a problem that has never been explored before: \emph{Can we fully utilize this idle input space by padding other content to further improve model performance and training efficiency?} In this paper, we propose a simple yet effective padding method called \textbf{Rep}eated \textbf{Pad}ding (\textbf{RepPad}). Specifically, we use the original interaction sequences as the padding content and fill it to the padding positions during model training. This operation can be performed a finite number of times or repeated until the input sequences' length reaches the maximum limit. Our RepPad can be viewed as a sequence-level data augmentation strategy. Unlike most existing works, our method contains no trainable parameters or hyperparameters and is a plug-and-play data augmentation operation. Extensive experiments on various categories of sequential models and five real-world datasets demonstrate the effectiveness and efficiency of our approach. The average recommendation performance improvement is up to 60.3\% on GRU4Rec and 24.3\% on SASRec. We also provide in-depth analysis and explanation of what makes RepPad effective from multiple perspectives. The source code will be released to ensure the reproducibility of our experiments.

Stealthy Attack on Large Language Model based Recommendation

Recently, the powerful large language models (LLMs) have been instrumental in propelling the progress of recommender systems (RS). However, while these systems have flourished, their susceptibility to security threats has been largely overlooked. In this work, we reveal that the introduction of LLMs into recommendation models presents new security vulnerabilities due to their emphasis on the textual content of items. We demonstrate that attackers can significantly boost an item's exposure by merely altering its textual content during the testing phase, without requiring direct interference with the model's training process. Additionally, the attack is notably stealthy, as it does not affect the overall recommendation performance and the modifications to the text are subtle, making it difficult for users and platforms to detect. Our comprehensive experiments across four mainstream LLM-based recommendation models demonstrate the superior efficacy and stealthiness of our approach. Our work unveils a significant security gap in LLM-based recommendation systems and paves the way for future research on protecting these systems.

ID Embedding as Subtle Features of Content and Structure for Multimodal Recommendation

Multimodal recommendation aims to model user and item representations comprehensively with the involvement of multimedia content for effective recommendations. Existing research has shown that it is beneficial for recommendation performance to combine (user- and item-) ID embeddings with multimodal salient features, indicating the value of IDs. However, there is a lack of a thorough analysis of the ID embeddings in terms of feature semantics in the literature. In this paper, we revisit the value of ID embeddings for multimodal recommendation and conduct a thorough study regarding its semantics, which we recognize as subtle features of content and structures. Then, we propose a novel recommendation model by incorporating ID embeddings to enhance the semantic features of both content and structures. Specifically, we put forward a hierarchical attention mechanism to incorporate ID embeddings in modality fusing, coupled with contrastive learning, to enhance content representations. Meanwhile, we propose a lightweight graph convolutional network for each modality to amalgamate neighborhood and ID embeddings for improving structural representations. Finally, the content and structure representations are combined to form the ultimate item embedding for recommendation. Extensive experiments on three real-world datasets (Baby, Sports, and Clothing) demonstrate the superiority of our method over state-of-the-art multimodal recommendation methods and the effectiveness of fine-grained ID embeddings.