This study proposes a methodology for recalculating influencer account impact from a reliability-oriented perspective under conditions where sufficient ground-truth labels are unavailable, taking into account the prevalence of fake followers and manip...
This study proposes a methodology for recalculating influencer account impact from a reliability-oriented perspective under conditions where sufficient ground-truth labels are unavailable, taking into account the prevalence of fake followers and manipulated engagement signals in influencer marketing environments. Conventional influence evaluation methods have relied heavily on the magnitude of static indicators such as follower counts and engagement rates; however, such approaches are vulnerable to overestimation in practical settings where engagement purchasing, short-term boosting, and partial manipulation are common. Rather than framing the problem as one of binary fake account detection, this study reformulates it as the task of structurally identifying the valid contribution of observed signals to influence estimation.
To this end, the study constructs multivariate time-series data by aligning upload, engagement, and growth indicators on a daily basis, and summarizes account-level behavioral dynamics into latent embeddings using a CNN–BiLSTM–Attention–based behavioral representation learning model. These embeddings are designed to capture how engagement signals are generated, diffused, and converted over time, rather than their absolute magnitudes. Deep Embedded Clustering is then applied to structurally align the embedding space and to form behavioral type baselines by grouping accounts with similar temporal evolution patterns. In this process, clusters are interpreted not as labels indicating legitimacy or fraud, but as reference coordinate systems that enable relative comparison within the same behavioral type.
Based on the clustering results, this study defines a set of multi-dimensional structural indicators, including upload–engagement alignment, engagement–growth linkage, engagement concentration and isolated spikes, growth stability, and relative deviation within behavioral types. These indicators are combined to estimate the valid contribution ratio of observed signals. The estimated ratio is implemented as a refinement coefficient that continuously attenuates observed engagement and follower scale into valid signals, rather than excluding entire accounts. This approach enables influence evaluation that reflects only structurally reliable contributions, even when partial manipulation is present.
Using the refined valid engagement and valid follower measures, influence scores are recalculated and transformed into distribution-based grades. In addition, re-clustering based on valid signals allows account grouping grounded in substantive influence structures rather than superficial scale. Experimental results reveal a significant negative correlation between spike frequency and valid contribution ratios. Comparative analysis before and after refinement shows that accounts with large apparent engagement but strong structural abnormalities are conservatively re-evaluated, whereas accounts with smaller scale but stable temporal structures tend to be reassessed upward. These findings indicate that the proposed method does more than reorder rankings; it repositions accounts by reflecting temporal structure and growth linkage.
The primary contribution of this study lies in systematically improving the reliability of influence evaluation under label-scarce conditions without making deterministic judgments about authenticity. In particular, the proposed two-stage clustering framework—defining clusters as behavioral baselines rather than classification outcomes and linking them to valid signal reconstruction and grading —provides a methodological foundation that balances interpretability and practical applicability. By extending influencer evaluation from detection-oriented approaches to reliability-based signal refinement, this study offers a basis for future research involving multimodal data integration, personalized performance prediction, and adaptive long-term baseline modeling.