Item-Language Model for Conversational Recommendation (2406.02844v1)

Abstract: Large-LLMs have been extremely successful at tasks like complex dialogue understanding, reasoning and coding due to their emergent abilities. These emergent abilities have been extended with multi-modality to include image, audio, and video capabilities. Recommender systems, on the other hand, have been critical for information seeking and item discovery needs. Recently, there have been attempts to apply LLMs for recommendations. One difficulty of current attempts is that the underlying LLM is usually not trained on the recommender system data, which largely contains user interaction signals and is often not publicly available. Another difficulty is user interaction signals often have a different pattern from natural language text, and it is currently unclear if the LLM training setup can learn more non-trivial knowledge from interaction signals compared with traditional recommender system methods. Finally, it is difficult to train multiple LLMs for different use-cases, and to retain the original language and reasoning abilities when learning from recommender system data. To address these three limitations, we propose an Item-LLM (ILM), which is composed of an item encoder to produce text-aligned item representations that encode user interaction signals, and a frozen LLM that can understand those item representations with preserved pretrained knowledge. We conduct extensive experiments which demonstrate both the importance of the language-alignment and of user interaction knowledge in the item encoder.

An Essay on "Item-LLM for Conversational Recommendation"

The paper, "Item-LLM for Conversational Recommendation," addresses the integration of collaborative filtering (CF) knowledge into LLMs to enhance their performance in conversational recommendation tasks. This work introduces the Item-LLM (ILM), a novel architecture designed to overcome several inherent challenges presented by traditional LLMs when applied to recommendation systems.

Overview

LLMs have demonstrated superior capabilities in various areas, such as dialogue understanding, reasoning, and coding, given their powerful emergent abilities. However, the utilization of LLMs in recommendation systems has not mirrored these advancements. One primary stumbling block is the discrepancy between the data on which LLMs are traditionally trained and the interaction signals in recommendation systems. Another challenge lies in the retention of LLMs’ original language and reasoning abilities post fine-tuning on specific recommendation data.

The authors propose the ILM framework to address these issues, comprising an item encoder and a frozen LLM. The core innovation lies in the item encoder, which generates text-aligned item representations encoding user interaction signals. This method allows the pre-trained knowledge of the LLM to be preserved, enhancing its ability to process diverse inputs.

Methodology

Model Architecture

The ILM follows a two-phase training approach:

1. Item-Language Representation Learning: In the initial phase, the authors utilize a Querying Transformer (Q-Former) as the item encoder. Inspired by the BLIP-2 model, this encoder bridges the modality gap by pre-training with item-text alignment tasks. Additionally, a novel item-item contrastive learning loss is introduced to regularize the model and encode co-watch information, thus improving item-language representations.
2. Item-LLM Training: In the second phase, the trained Q-Former is integrated with a frozen LLM. An adaptor layer is used to map Q-Former outputs to the LLM’s input dimension. The training in this phase is performed on conversational recommendation tasks, where only the Q-Former and adaptor parameters are updated, safeguarding the LLM’s pretrained capacities.

Experiments and Results

The empirical evaluation is conducted on two datasets: the ELM 24 tasks and the OpenP5 benchmark, covering a broad spectrum of conversational recommendation sub-tasks. Evaluative metrics include Semantic Consistency (SC) and log perplexity for the ELM tasks, while top-K hit rate (HR@K) and normalized discounted cumulative gain (NDCG@K) are utilized for OpenP5 tasks.

Results Insights

The ILM consistently outperforms baselines such as the CoLLM approach across all metrics and datasets. Notably, in the ELM tasks, the SC improved by 3.27% and log perplexity reduced by 12.12%. For the OpenP5 tasks, the ILM demonstrated superior performance for both seen and unseen test data, further establishing the efficacy of phase-1 training techniques. Ablation studies underscored the importance of phase-1 training and highlighted the regularizing effect of the item-item and user-item contrastive losses.

Implications and Future Directions

The ILM framework bridges the gap between collaborative filtering signals and LLM capabilities without compromising on the LLM’s inherent strengths. This method has significant practical implications for developing more adept conversational recommender systems that can seamlessly integrate complex user interaction signals.

Theoretically, this work paves the way for further exploration into multi-modal learning, where non-linguistic interaction data can be effectively incorporated into linguistic models. The authors' methodology can be extended and adapted to various domains beyond video and retail recommendations, incorporating different forms of user interaction signals.

Future research could explore the application of more sophisticated semantic id-based methods alongside the ILM framework to further enhance performance. Additionally, investigating the integration of other advanced user-interaction signals can provide a comprehensive understanding of user preferences, thus refining the recommendation accuracy.

Conclusion

The "Item-LLM for Conversational Recommendation" presents a substantial advancement in the field of recommender systems by adeptly integrating collaborative filtering signals into LLMs. The proposed ILM framework not only mitigates many existing challenges but also preserves the LLM's pre-trained knowledge, enhancing the model's robustness in conversational tasks. The results demonstrate the substantial gains afforded by this novel approach, charting a promising trajectory for future research and practical applications in AI-driven recommender systems.