AI Uncovers Human Values Through Twitter Analysis

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Jul 26 2024

In an article published in the journal SN Computer Science, researchers explored the identification of basic human values through social media activity, particularly focusing on Twitter. They compiled a graph dataset to identify value priorities using graph clustering techniques and proposed a behavior-based group recommendation method. The study validated these groups through hypotheses, achieving significant intra-cluster correlation and silhouette clustering coefficient (SCC) scores among users.

Background

In the digital era, individuals frequently express their thoughts and preferences on social media, reflecting their basic human values, which are fundamental aspects of behavior influencing decisions and social interactions. Previous studies have utilized social media content to explore diverse human behaviors, including values, personality, and self-efficacy. However, existing research largely focuses on individual recommendations or group identification based on structural properties, social ties, and interests, lacking an emphasis on psychological attributes such as values.

This paper addressed this gap by identifying groups of users with similar value dimensions from Twitter interactions. Using the International Business Machines (IBM) personality insights application programming interface (API), the study employed graph neural networks (GNNs) and contextual-psychological analysis to discover and validate groups based on hedonism values. The proposed approach outperformed existing models, demonstrating better group identification and validation methods.

Methodology and Findings

The data collection process in this study involved two steps, collecting tweets from selected users to identify groups based on value dimensions and validating these groups with a focus on hedonism. The dataset for group identification consisted of 1,140 Twitter users and 3,275,417 tweets, primarily in English from the United States (US) and United Kingdom (UK). For validation, 840,508 tweets from 300 users interested in gadgets and movies were collected.

The methodology involved pre-processing tweets to remove noise and using the IBM personality insight API to compute users' value scores. Two approaches for group identification were employed, clustering based on hedonism scores and from contextual and psychological attributes. The study utilized GNNs and spectral clustering to enhance group accuracy, focusing on hedonism, self-enhancement, and self-transcendence dimensions.

Contextual analysis included topic modeling with latent semantic analysis (LSA) and bidirectional encoder representations from transformers (BERT) embeddings, revealing topics like music and fashion. The psychological analysis employed the linguistic inquiry and word count (LIWC) tool to refine user clusters. Validation involved lexical analysis of tweets from tech enthusiasts and movie lovers, confirming high hedonism scores correlated with interests in gadgets and movies.

The authors demonstrated that individuals with high hedonism scores often showed preferences for movies and technology, validated through IBM personality insight API scores and Z-test analysis, suggesting these users' tweets reflect their real-life preferences.

Performance Evaluation and Comparative Analysis 

The researchers presented a comparative analysis of the proposed hedonism-based group identification approaches, graph-based group identification from hedonism value scores (GHV), and contextual-psychological-based hedonism value scores (CPHV), against four state-of-the-art topic-based clustering models: COM, hierarchical Bayesian geographical model for group recommendation (HBGG), group recommendation model considering group cohesion (GGC), and hybrid-GGC (HGGC).

Both COM and HBGG utilized statistical topic modeling for group recommendation, GGC employed a latent Dirichlet allocation (LDA)-based clustering approach, and HGGC integrated content information with probabilistic matrix factorization. Hyperparameters for these models were set as per their respective studies, with initial topic numbers set at 20 and later tuned for robustness.

Performance evaluation was conducted using two metrics, SCC and intraclass correlation (ICC). SCC assessed cluster cohesion and separation, with values ranging from -1 to +1, where higher values indicated better clustering. ICC measured data similarity within clusters, with values between 0 and 1, where scores near 1 signified higher similarity.

Results showed that the CPHV model outperformed all other models in both SCC and ICC scores, with improvements of up to 21% and 16%, respectively. GHV achieved the best performance among hedonism-based approaches, while HGGC performed the best among the state-of-the-art models. The superior SCC and ICC scores of CPHV and GHV confirmed the effectiveness of the proposed methodologies in forming high-quality and similar clusters, validating their utility in hedonistic user group identification.

Insights and Implications

The researchers highlighted the effectiveness of group recommendation systems leveraging psychological factors and graph-based approaches. The integration of GNNs and spectral clustering enhanced node representation and clustering accuracy. The authors identified hedonistic users' preferences through social media interactions, revealing patterns in travel, nightlife, food, and relationships. Contextual and psychological filtering techniques, such as LSA and BERT, further refined group identification. Validation with tech enthusiasts and movie lovers showed strong cluster cohesion. Limitations included challenges in data availability and changing user values.

Conclusion

In conclusion, the researchers introduced innovative techniques for group identification and recommendation based on hedonistic values using Twitter data. By leveraging GNNs and contextual-psychological analysis, the researchers achieved superior performance in clustering accuracy and group validation compared to state-of-the-art models.

The findings highlighted that hedonistic users displayed distinct patterns in their social media interactions, which could be effectively captured using the proposed methods. Despite some limitations, such as data availability and evolving values, the authors paved the way for future research in value-based group recommendations, demonstrating significant improvements in cluster quality and validation.