@article{Wan2026, 
author = {Haojie Wan and Wei Yu and Zhiyu Li and Shijun Li and Fang Yu and Juncheng Yang},
title = {Trustworthy Evaluation Method for User Psychological Assessment Based on Large-Scale Heterogeneous Graph Networks},
year = {2026},
journal = {International Journal of Crowd Science},
volume = {10},
number = {3},
pages = {152-165},
keywords = {mental health identification, association mining, hierarchical network model, multi-latent variable features, temporal characteristics, result characteristics, anomaly detection},
url = {https://www.sciopen.com/article/10.26599/IJCS.2025.9100008},
doi = {10.26599/IJCS.2025.9100008},
abstract = {Online psychological assessment information system can quickly pre-screen users with potential mental health risks. Therefore, accurate and effective psychological assessment results are the foundation and guarantee of psychological health work. The process data, including response time and answer patterns, reflect the behavioral patterns of users during the assessment. However, the current predominant abnormal detection methods rely on constructing global or contextual data distributions, distances, or error calculation models. These methods fail to account for the influence of user personality traits and question characteristics on behavior patterns, making it challenging to capture personalized differences and the combined effects of multiple behaviors. To address these challenges, this paper proposes a knowledge-enhanced heterogeneous graph model based on large-scale data. It embeds a large number of factors affecting the answering process, including user profiles, question semantics, user relationships, question relationships, user-question interactions, contextual environments, and temporal dynamics, into a heterogeneous multi-graph network. It adaptively learns the importance of each node and edge through graph attention mechanisms and employs Transformer’s encoding layers to represent user personality traits, question knowledge features, contextual features, and multi-behavior features. The primary contribution of this approach lies in designing a unified multi-graph network model on a large-scale dataset for jointly learning the complex characteristics of users and questions. It explicitly distinguishes the abnormal characteristics of different users and questions in various contextual environments, effectively achieving personalized abnormal detection in user psychological assessments. Experimental results demonstrate that this method exhibits high accuracy and recall in identifying the trustworthiness of psychological assessments, enabling a fine-grained assessment of the credibility of each response.}
}