XAI IN ALTMETRICS ON X

EXPLAINING BEHAVIORS OF BOTS AND HUMANS

Abstract

This paper examines how Explainable Artificial Intelligence (XAI) can contribute to interpreting the behavioural patterns that differentiate bot and human accounts in the dissemination of scientific articles on platform X. Drawing on findings from a broader doctoral investigation, the study mobilises SHAP (SHapley Additive Explanations) values generated by a supervised XGBoost-based model trained on 46 predictive variables across five behavioural dimensions: activity, temporality, textuality, sentiment, and profile. The analysis is structured in three levels: the relative weight of each dimension in the model's global explanatory power; identification of thresholds for the most important variables; and variation of the explanatory profile across groups defined by thematic dispersion of accounts, considering 362,395 automatically classified accounts. Results indicate that the Activity dimension accounts for 39.1% of explanatory power, followed by Temporality (29.0%), Textuality (16.3%), Profile (12.8%), and Sentiment (2.7%). Bots post at a median volume eight times higher than humans, with more regular cadence, while humans exhibit more spontaneous and topic-specific behaviour. This explanatory hierarchy varies with the degree of automation by area: in low-automation contexts, volume is the dominant signal; in high-automation contexts, temporal regularity takes the lead. The findings show that XAI not only makes detection models more transparent, but also provides a foundation for a critical reading of altmetric indicators based on mention counts.