Introduction
Traditional group recommendation systems often rely on explicit social ties—friendships, followers, or network structures. However, these connections do not always capture the underlying psychological motivations that drive collective behavior. This research explores "Basic Human Values" as a more robust foundation for identifying cohesive groups. By focusing on values like hedonism, the authors propose a system that predicts real-world interests based on digital footprints rather than just social graphs.
Methodology
The researchers developed a dual-track approach to identify groups based on psychological alignment, specifically targeting the value of Hedonism (the pursuit of pleasure and gratification).
1. Graph-based Group Identification (GHV)
This method constructs a user graph where edges represent similarities in value scores derived from the IBM Personality Insight API. A Graph Neural Network (GNN) processes these connections to learn latent patterns, which are then partitioned using Spectral Clustering to find optimal groups.
2. Contextual-Psychological Identification (CPHV)
This pipeline utilizes deep learning and psycholinguistics. It employs BERT and Latent Semantic Analysis (LSA) to capture the contextual nuances of tweets, while Linguistic Inquiry and Word Count (LIWC) analyzes the psychological lexicon. Density-based clustering then identifies compact, psychologically aligned user groups.
Datasets and Validation
The study leveraged two distinct datasets scraped from Twitter:
- Identification Dataset: 1,140 unverified users with 3.2 million tweets used for discovering value-based patterns.
- Validation Dataset: 300 users categorized as "Movie Enthusiasts" or "Tech Enthusiasts" based on specific hashtags (e.g., #marvel, #gadgets). This served to verify if the identified "hedonist" groups actually shared these real-world hobbies.
Key Results and Insights
The CPHV model demonstrated superior performance, achieving a Silhouette Clustering Coefficient (SCC) of 76% and an Intra-cluster Correlation Coefficient (ICC) of 60%. These metrics confirm that the resulting groups are both distinct and internally cohesive compared to state-of-the-art models.
The research highlights that psychological traits, particularly hedonism, are highly visible in digital behavior. Users valuing hedonism frequently discuss travel, fashion, and entertainment, making them ideal candidates for group-based recommendations even in the absence of direct social links. The "hedonist" groups identified by the model showed a high average hedonism score (approx. 0.70), correlating strongly with the interests of the validation groups.
Conclusion and Future Work
While the model shows promise, challenges remain regarding data sparsity for less active users and the temporal shifts in human values over time. Future research aims to extend this work to multimodal analysis—incorporating image data from platforms like Instagram and professional data from LinkedIn—to further refine these behavioral profiles.
Summary updated