# Generating Synthetic Data for Conversational Music Recommendation Using Random Walks and Language Models

**Authors:** Megan Leszczynski, Ravi Ganti, Shu Zhang, Krisztian Balog, Filip Radlinski, Fernando Pereira, Arun Tejasvi Chaganty  
**arXiv:** [2301.11489](https://arxiv.org/abs/2301.11489)  
**Date:** 2023-01-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.11489v1/

## Key points

- Limitations of practical adoption of conversational recommendation systems (CRSs) due to lack of diverse training data
- CRSs allow users to provide natural language feedback for better recommendations
- Proposed method: generating synthetic training data by transforming curated item collections into item-seeking conversations
- Biased random walk algorithm used to generate slates (sets of item recommendations)
- Language model used to generate corresponding user utterances
- Created a conversational music recommendation dataset with over one million conversations
- Crowdsourced evaluation showed consistency between generated conversations and relevant recommendations
- CRS trained on synthetic data outperformed baseline models in offline and online evaluations
- Synthetic data generated through biased random walks and language models can improve performance of CRSs
- Innovative approach for addressing lack of diverse conversational training data for CRSs

## Summary

The paper titled "Generating Synthetic Data for Conversational Music Recommendation Using Random Walks and Language Models" addresses the limitations of practical adoption of conversational recommendation systems (CRSs) due to a lack of diverse training data. CRSs allow users to provide natural language feedback to control their recommendations, which improves upon traditional recommendation systems. To overcome this limitation, the authors propose a new method for generating synthetic training data by transforming curated item collections, such as playlists or movie watch lists, into item-seeking conversations. The proposed method involves using a biased random walk algorithm to generate a sequence of slates, which are sets of item recommendations. These slates are then used as input for a language model that generates corresponding user utterances. By applying this approach, the authors were able to create a conversational music recommendation dataset with over one million conversations. To evaluate the effectiveness of their synthetic data generation method, the authors conducted a crowdsourced evaluation and found that the generated conversations were consistent with relevant recommendations. They further trained a CRS using the synthetic data and compared its performance against standard retrieval baselines in offline and online evaluations. The results showed that the CRS trained on the synthetic data significantly outperformed the baseline models in both offline and online evaluations. This highlights the potential of using synthetic training data generated through biased random walks and language models to improve the performance of conversational recommendation systems. Overall, this paper presents an innovative approach for addressing the lack of diverse conversational training data for CRSs. The findings demonstrate that utilizing synthetic data can lead to improved recommendation performance, paving the way for wider adoption and enhanced user experiences in conversational recommendation systems.