Avkoda smak: Att förbättra film- och bokrekommendationer med LLM-baserad personlighetsprofilering : Hur bra är språkmodeller på att estimera personlighet?

Recommendation systems are widely used to recommend for example movies, music and products. However, they often struggle to deliver good recommendations when data is sparse, such as with new users, new content or users who rarely provide feedback. Recent advances in automatic personality recognition have inspired the development of personality-aware recommendation systems to address these challenges. In this study, we suggest the use of large language models (LLMs) for personality estimation and the construction of personality-aware recommendation systems, since these models have shown a remarkable ability to handle complex and diverse tasks. We evaluate both prompted and finetuned LLMs on the Essays and PANDORA datasets for predicting the Big Five personality traits. In addition, we apply these models to user-generated movie and book reviews to estimate personality profiles that are integrated into a recommendation algorithm. The performance of this approach is compared to a traditional user-based collaborative filtering method. The results show that Claude 3.7 Sonnet and the finetuned GPT-4o achieved the highest classification accuracies, exceeding 60%. These performance levels are comparable to those reported for state-of-the-art machine learning models in personality classification, with no significant differences observed. Furthermore, personality-aware recommendation systems perform comparably to, and in some cases outperform, standard collaborative filtering approaches. In cold-start scenarios involving users with limited historical data, a personality-aware system achieved up to 30% higher recommendation precision. This opens the door to delivering high-quality book and movie recommendations even for completely new users. While the findings highlight the potential of LLMs in enhancing recommendation systems, future work should also evaluate this approach for larger datasets with more users for increased statistical significance.