Personalized Recommendations with Generative AI

Recommender systems are designed to help users find items that match their interests. This could include videos, products, or articles. Usually, these systems work with a fixed set of items that are already available. Recently, advancements in generative AI have made it possible to create new items based on user input, instead of just retrieving existing ones. This opens up a new challenge: how can we personalize these generative items for users when there could be an endless number of options?

This article discusses a framework to tackle this challenge, focusing on how to better connect users with the items they want by using prompts, or specific requests from users, to retrieve models that generate customized outputs. In our study, we present a new dataset containing thousands of images generated by various models using different prompts, and we explain how we can rank these generated items to best match user preferences.

#The Challenge of Personalization

Personalization in generative recommendations means understanding what individual users like. However, with numerous generative models available, it’s unrealistic for users to check each potential option one by one. Instead, we propose a solution that first narrows down the models based on user prompts and preferences. This process consists of two main steps: retrieving relevant models based on given prompts and ranking the items generated by those models.

#Framework Overview

The framework we propose includes two key stages:

1. Prompt-Model Retrieval: In this stage, we identify which generative models are most relevant to the user’s prompt. By using a fixed set of diverse prompts, we can visually assess how different models perform.

2. Generated Item Ranking: After narrowing the choices down, we then focus on ranking the generated items from these selected models based on user feedback. This feedback provides insights into user preferences.

#The GEMRec-18K Dataset

To support our framework, we created a dataset named GEMRec-18K. It consists of 18,000 images generated using 200 different generative models paired with 90 diverse prompts. This dataset is essential for improving generative recommendation systems as it allows us to analyze how well different models respond to various requests. The prompts were collected from various sources to ensure they cover a wide range of themes and styles.

#Importance of User Interaction

An effective recommendation system must facilitate user interaction. The proposed framework allows users to view and evaluate the generated images through an interactive interface. By doing this, users can express what they like or dislike, enabling the system to learn and improve over time. The first stage, Prompt-Model Retrieval, lets users compare outputs from different models. The second stage, Generated Item Ranking, allows users to arrange images based on their preferences, giving valuable feedback to help refine future recommendations.

#Exploring Generated Images

We analyzed the variety and quality of the images created by different models. By examining the differences between generated images for the same prompt, we can see how unique or similar the outputs are. This analysis helps in understanding which models produce diverse results and which tend to generate similar images.

#Limitations of Current Metrics

Evaluating the effectiveness of generative models is not straightforward. Current metrics often focus on popularity or accuracy, but these alone do not offer a complete picture. Popular models might yield similar results, which can limit diversity. Therefore, we need a more comprehensive metric that evaluates not just the quality of generated images, but also their variety.

#Evaluating with a New Metric

To address the limitations of existing evaluation techniques, we introduce a new metric called the Generative Recommendation Evaluation Score (GRE-Score). This score takes into account various factors, including how well the generated images match the prompts and the diversity of the outputs. By using this new metric, we can provide a more rounded assessment of each model’s performance.

#Future Directions

Our findings lay the groundwork for multiple future research opportunities. One direction is to expand the GEMRec dataset by including even more prompts and models. This will enhance the personalization aspect of our framework. Additionally, we aim to conduct studies with users to see how they interact with our system and gather data to further refine model recommendations.

Another important aspect is establishing standardized evaluation methods for generative recommendations. Understanding individual preferences is key, and we need to develop metrics that accurately reflect user tastes. Finally, while our study focuses on image generation, we believe the principles can be applied to other fields, such as text or music generation.

#Conclusion

The integration of generative AI into recommender systems presents exciting new possibilities for personalized recommendations. By proposing a structured framework that incorporates user feedback and innovative evaluation metrics, we can enhance the user experience in exploring generated items. Our work serves as a stepping stone in the ongoing journey to create more effective and personalized generative recommendation systems.