Lifelong Personalized Low-Rank Adaptation of Large Language Models for Recommendation (2408.03533v2)

Abstract: We primarily focus on the field of LLMs for recommendation, which has been actively explored recently and poses a significant challenge in effectively enhancing recommender systems with logical reasoning abilities and open-world knowledge. Current mainstream efforts mainly center around injecting personalized information from recommendation models into LLMs by customizing input templates or aligning representations between semantic and recommendation spaces at the prediction layer. However, they face three significant limitations: (1) LoRA is mostly used as a core component in existing works, but personalization is not well established in LoRA parameters as the LoRA matrix shared by every user may not cater to different users' characteristics, leading to suboptimal performance. (2) Although lifelong personalized behavior sequences are ideal for personalization, their use raises effectiveness and efficiency issues since LLMs require escalating training and inference time to extend text lengths. (3) Existing approaches aren't scalable for large datasets due to training efficiency constraints. Thus, LLMs only see a small fraction of the datasets (e.g., less than 10%) instead of the whole datasets, limiting their exposure to the full training space. To address these problems, we propose RecLoRA. This model incorporates a Personalized LoRA module that maintains independent LoRAs for different users and a Long-Short Modality Retriever that retrieves different history lengths for different modalities, significantly improving performance while adding minimal time cost. Furthermore, we design a Few2Many Learning Strategy, using a conventional recommendation model as a lens to magnify small training spaces to full spaces. Extensive experiments on public datasets demonstrate the efficacy of our RecLoRA compared to existing baseline models.

Lifelong Personalized Low-Rank Adaptation of LLMs for Recommendation

The paper "Lifelong Personalized Low-Rank Adaptation of LLMs for Recommendation" presents a novel approach to enhancing recommender systems by introducing a personalized low-rank adaptation (LoRA) framework for LLMs. The proposed framework, RecLoRA, addresses significant challenges in personalized recommendation by leveraging robust numerical results and innovative architectural components, such as the Personalized LoRA module and the Long-Short Modality Retriever.

Motivation

Recommender systems (RSs) play a crucial role in mitigating information overload by suggesting relevant items to users based on their preferences. While LLMs have made significant strides in NLP tasks due to their capability to understand and generate human-like text, integrating LLMs into RS poses several unique challenges. Current efforts largely center on injecting personalized information into LLMs. However, conventional methods face three pivotal limitations:

1. Existing works often employ a shared LoRA matrix for all users, which does not account for individual user characteristics, leading to suboptimal personalization.
2. Utilizing lifelong personalized behavior sequences increases training and inference time, challenging the efficiency of LLMs due to the extended text lengths required.
3. Scalability issues arise with large datasets as training efficiency constraints limit LLM exposure to complete datasets, thereby inhibiting full leveraging of training spaces.

Proposed Solution: RecLoRA

To address these limitations, the authors propose the RecLoRA framework, characterized by three main contributions:

1. Personalized LoRA Module: Unlike conventional approaches where a static LoRA matrix is shared, RecLoRA introduces a personalized LoRA module. This module maintains separate LoRA parameters for different users, thus achieving fine-grained personalization. Specifically, a set of parallel meta-LoRA weights is employed, with a soft routing mechanism guided by a classic recommendation model (CRM) like SIM to dynamically generate personalized LoRA matrices. This design enhances the recommendation capabilities by aligning them more closely with individual user behaviors.
2. Long-Short Modality Retriever: The issue of efficiency due to long behavior sequences is addressed by the Long-Short Modality Retriever, which retrieves different history lengths for different input modalities (ID and text). For CRM, longer sequences are used to comprehensively capture user behavior, whereas for LLM inputs, shorter sequences are utilized to balance processing time. This substantially improves effectiveness without a proportional increase in time cost.
3. Few2Many Learning Strategy: Recognizing the computational constraints for large datasets, RecLoRA employs a Few2Many learning strategy. A conventional recommendation model is initially trained on the complete dataset to learn extensive user-item interaction patterns. This model then serves to transform small training subsets into representations that effectively encapsulate the full data spectrum, thereby augmenting the LLM without escalating training times. This approach ensures the LLM’s receptive field spans the full training space, enhancing generalizability and performance efficiency.

Experimental Results

Extensive experiments conducted on public datasets, including MovieLens and GoodReads, demonstrate RecLoRA's significant improvement over baseline models. Key observations include:

• AUC and Log Loss: RecLoRA achieved superior AUC and lower Log Loss values in comparison to both ID-based traditional recommendation models (DeepFM, SASRec, SIM) and LM-based models (CTR-BERT, TallRec, ReLLa). For example, on the MovieLens-25M dataset, RecLoRA outperformed the state-of-the-art ReLLa with AUC improvements of up to 0.0063 and Log Loss reductions of up to 0.0110.
• Efficiency: With the Long-Short Modality Retriever, RecLoRA effectively handled long sequence retrieval on the ID side while maintaining short sequences on the text side, achieving an outstanding balance between performance and time efficiency.
• Sample Efficiency: The Few2Many learning strategy demonstrated exceptional sample efficiency. RecLoRA outperformed ReLLa significantly with fewer training samples, highlighting the efficacy of incorporating comprehensive CRM knowledge into LLM finetuning.

Implications and Future Work

The introduction of RecLoRA has several practical and theoretical implications. Practically, it offers a scalable and efficient methodology to refine LLMs for personalized recommendation, crucial for applications in large-scale systems like e-commerce or content streaming platforms. Theoretically, it advances the understanding of parameter-efficient finetuning and user behavior modeling, presenting a nuanced approach to personalization in LLMs.

Future developments could explore more sophisticated personalization mechanisms within LLMs and address fairness issues to ensure equitable recommendation across diverse user demographics. Additionally, integrating adaptive meta-learning algorithms could further enhance the dynamic and personalized adaptation of the LoRA matrices.

This paper delineates an innovative pathway for melding personalization with LLM capabilities, significantly enriching the landscape of recommender systems research.