Impression Zombies: Characteristics Analysis and Classification of New Harmful Accounts on Social Media

TL;DR

The paper investigates Impression Zombies, a monetization-driven class of harmful accounts on X that inflate engagement. It provides the first quantitative characterization by comparing 9,909 replies from 101 posts to general users across account age, posting volume, profile cues, following patterns, and timing. It then presents a two-stage detection method that leverages contextual incongruity between parent posts and replies, achieving about 0.92 accuracy after fine-tuning embeddings with a metric-learning objective and classifying with an MLP. The work advances understanding of monetization-driven user behaviors on social platforms and offers a practical framework for improving information integrity, with future work including multimodal signals.

Abstract

``Impression Zombies'', a type of malicious account designed to artificially inflate engagement metrics, have recently emerged as a significant threat on X (formerly Twitter). These accounts disseminate a high volume of low-quality, irrelevant posts, which degrade the user experience. This study aims (1) to quantitatively characterize their behavioral patterns and (2) to develop a method for detecting such accounts. To address the first objective, we collected data from 9,909 accounts and compared the characteristics of Impression Zombies and general users within this dataset. We find that, Impression Zombies post more than three times the average total number of posts per day and tend to gather followers by using phrases such as ``follow back.'' Addressing the second objective, we constructed a classification model for Impression Zombies that leverages the contextual incoherence often observed between parent posts and the replies from Impression Zombies. Experimental results show that our model achieved approximately 92\% accuracy in detecting Impression Zombies. This study provides the first quantitative insights into Impression Zombies and offers a practical framework for detecting such accounts, contributing to platform transparency and the health of social media ecosystems.

Figures (6)

• Figure 1: Examples of replies lacking contextual connection to the parent post and the corresponding Impression Zombies.
• Figure 2: Histogram of account age measured from the account creation date to November 1, 2024. (a) shows the distribution of general users, and (b) shows the distribution of Impression Zombies. The x-axis represents elapsed days, and the y-axis represents the relative frequency. The red line indicates the Kernel Density Estimate (KDE) curve. The difference in distribution between general users and Impression Zombies were statistically significant, with a t-test yielding a p-value below 0.001.
• Figure 3: Scatter plot of the total number of posts versus account age (days from creation to November 1, 2024) for (a) general users and (b) Impression Zombies. The x-axis represents the elapsed days, and the y-axis represents the total number of posts. The red line indicates the regression line. The difference in distribution between general users and Impression Zombies were statistically significant, with a t-test yielding a p-value below 0.001.
• Figure 4: Distribution of the following-to-follower ratio for (a) general users and (b) Impression Zombies. The x-axis represents this ratio, and the y-axis represents the relative frequency. The red line indicates the KDE curve. The average following-to-follower ratio was 2.93 for general users, in contrast to 1.26 for Impression Zombies. The difference in distribution between general users and Impression Zombies were statistically significant, with a t-test yielding a p-value below 0.001.
• Figure 5: Heatmap showing the number of posts made at different times of the day for (a) general users and (b) Impression Zombies. The x-axis shows the time of day and the y-axis shows the day of the week. Color intensity represents high (purple) or low (yellow) activity.