Large Language Model - Based Conversational Recommender System for Personalizing Long-term Cryptocurrency Portfolio Recommendation

The growth of cryptocurrencies in the last 10 years as an investment option has increased. However, the volatile prices often make investors panic and lose money. The cryptocurrency portfolio recommender system built in previous research is still limited to historical prices and trends and does not focus on user preferences and understanding in natural language conversation. Therefore, we propose the development of a Conversational Recommender System (CRS) using a Large Language Model (LLM) to personalize long-term cryptocurrency portfolio recommendations. We use the prompt engineering and fine-tuning approach on LLM to match user preferences and make the CRS more tailored to the cryptocurrency domain. This method can overcome the system’s shortcomings in understanding user preferences, impacting recommendations’ personalization. Accuracy evaluation results show that our proposed system has an Item Coverage of 39.62%, an Expected Popularity Complement (EPC) value of 0.9405, an Average Recommendation Popularity (ARP) value of 0.0595, and a Fine-Tune Training Loss value of 0.4415. Meanwhile, our CRS has a positive response of 81% on satisfaction evaluation. The results show that our CRS has a deeper understanding of the conversation’s context and delivers more interactive recommendations for cryptocurrency portfolios, providing investors peace of mind when investing.