[Paper Review] Improving Recommendation Systems with User Personality Inferred from Product Reviews

Xinyuan Lu, Min‐Yen Kan|arXiv (Cornell University)|Mar 9, 2023

Personality Traits and PsychologyPsychology3 citations

TL;DR

This paper proposes a method to automatically infer users' OCEAN personality traits from their product review texts using the Receptiviti API, integrating these inferred profiles into neural collaborative filtering (NCF) models to improve recommendation performance. The approach boosts performance by 3–28% on two new datasets (Amazon-beauty and Amazon-music), with openness and extroversion being most beneficial in music recommendation and conscientiousness in beauty product recommendation.

ABSTRACT

Personality is a psychological factor that reflects people's preferences, which in turn influences their decision-making. We hypothesize that accurate modeling of users' personalities improves recommendation systems' performance. However, acquiring such personality profiles is both sensitive and expensive. We address this problem by introducing a novel method to automatically extract personality profiles from public product review text. We then design and assess three context-aware recommendation architectures that leverage the profiles to test our hypothesis. Experiments on our two newly contributed personality datasets -- Amazon-beauty and Amazon-music -- validate our hypothesis, showing performance boosts of 3--28%.Our analysis uncovers that varying personality types contribute differently to recommendation performance: open and extroverted personalities are most helpful in music recommendation, while a conscientious personality is most helpful in beauty product recommendation.

Motivation & Objective

• To address the challenge of acquiring user personality data, which is expensive, time-consuming, and privacy-sensitive in recommendation systems.
• To explore whether user-generated review texts can serve as a viable, privacy-preserving source for inferring personality traits.
• To evaluate whether automatically inferred personality profiles improve recommendation performance in real-world recommendation scenarios.
• To construct new, publicly available datasets with personality scores inferred from Amazon reviews in the beauty and music domains.
• To analyze the differential impact of specific personality traits (OCEAN) on recommendation performance across domains.

Proposed method

• Leveraged the Receptiviti API to automatically infer OCEAN personality scores (Openness, Conscientiousness, Extroversion, Agreeableness, Neuroticism) from user-generated product reviews.
• Constructed two new datasets—Amazon-beauty and Amazon-music—by extending the Amazon review dataset and enriching it with inferred personality profiles.
• Designed three context-aware recommendation architectures that integrate personality embeddings into the neural collaborative filtering (NCF) framework.
• Used soft-labeling strategies to incorporate personality scores as auxiliary user features in the recommendation model.
• Evaluated performance using standard metrics: Hit Rate (HR) and Normalized Discounted Cumulative Gain (NDCG) across multiple settings.
• Analyzed the contribution of individual personality traits to recommendation performance using ablation studies and domain-specific analysis.

Figure 1 . The overall structure of our model. In this example, the user’s OCEAN score is {30,70,50,30,20}. The NCF + Most salient personality selects the personality with the highest score, i.e., conscientiousness as the personality embedding vector. NCF + Soft-labeled personality takes all five OC

Experimental results

Research questions

• RQ1 Can user personality traits be accurately inferred from product review text using commercial NLP APIs?
• RQ2 To what extent does incorporating inferred personality profiles improve recommendation system performance?
• RQ3 How do different personality traits (OCEAN) contribute differently to recommendation performance in distinct domains (e.g., music vs. beauty)?
• RQ4 What are the limitations of using coarse-grained personality models like OCEAN in recommendation systems?
• RQ5 How does the imbalance in personality distribution (e.g., high extroversion) in review data affect model generalization?

Key findings

• Inferred personality profiles from product reviews significantly improve recommendation performance, with gains ranging from 3% to 28% across datasets.
• In the music domain, openness and extroversion were the most influential traits, with extroversion showing the strongest positive impact on recommendation quality.
• In the beauty domain, conscientiousness emerged as the most beneficial personality trait, significantly enhancing recommendation performance.
• Agreeableness also contributed positively in music recommendation, though to a lesser extent than openness and extroversion.
• Neuroticism and openness showed minimal or negligible impact on performance in both domains, suggesting limited signal utility in the current setup.
• The study identified a potential bias in the data, with extroversion being overrepresented due to the nature of review-writing behavior, suggesting a need for alternative data sources like purchase history in future work.

Figure 2 . Distribution of personality traits in Amazon-beauty and Amazon-music datasets. The x-axis represents the score for each trait; the y-axis represents the number of users. The red line represents the median for each trait.