Bots Don't Sit Still: A Longitudinal Study of Bot Behaviour Change, Temporal Drift, and Feature-Structure Evolution

📝 Abstract

Social bots are now deeply embedded in online platforms for promotion, persuasion, and manipulation. Most bot-detection systems still treat behavioural features as static, implicitly assuming bots behave stationarily over time. We test that assumption for promotional Twitter bots, analysing change in both individual behavioural signals and the relationships between them. Using 2,615 promotional bot accounts and 2.8M tweets, we build yearly time series for ten content-based meta-features. Augmented Dickey-Fuller and KPSS tests plus linear trends show all ten are non-stationary: nine increase over time, while language diversity declines slightly. Stratifying by activation generation and account age reveals systematic differences: second-generation bots are most active and link-heavy; short-lived bots show intense, repetitive activity with heavy hashtag/URL use; long-lived bots are less active but more linguistically diverse and use emojis more variably. We then analyse co-occurrence across generations using 18 interpretable binary features spanning actions, topic similarity, URLs, hashtags, sentiment, emojis, and media (153 pairs). Chi-square tests indicate almost all pairs are dependent. Spearman correlations shift in strength and sometimes polarity: many links (e.g. multiple hashtags with media; sentiment with URLs) strengthen, while others flip from weakly positive to weakly or moderately negative. Later generations show more structured combinations of cues. Taken together, these studies provide evidence that promotional social bots adapt over time at both the level of individual meta-features and the level of feature interdependencies, with direct implications for the design and evaluation of bot-detection systems trained on historical behavioural features.

💡 Analysis

Social bots are now deeply embedded in online platforms for promotion, persuasion, and manipulation. Most bot-detection systems still treat behavioural features as static, implicitly assuming bots behave stationarily over time. We test that assumption for promotional Twitter bots, analysing change in both individual behavioural signals and the relationships between them. Using 2,615 promotional bot accounts and 2.8M tweets, we build yearly time series for ten content-based meta-features. Augmented Dickey-Fuller and KPSS tests plus linear trends show all ten are non-stationary: nine increase over time, while language diversity declines slightly. Stratifying by activation generation and account age reveals systematic differences: second-generation bots are most active and link-heavy; short-lived bots show intense, repetitive activity with heavy hashtag/URL use; long-lived bots are less active but more linguistically diverse and use emojis more variably. We then analyse co-occurrence across generations using 18 interpretable binary features spanning actions, topic similarity, URLs, hashtags, sentiment, emojis, and media (153 pairs). Chi-square tests indicate almost all pairs are dependent. Spearman correlations shift in strength and sometimes polarity: many links (e.g. multiple hashtags with media; sentiment with URLs) strengthen, while others flip from weakly positive to weakly or moderately negative. Later generations show more structured combinations of cues. Taken together, these studies provide evidence that promotional social bots adapt over time at both the level of individual meta-features and the level of feature interdependencies, with direct implications for the design and evaluation of bot-detection systems trained on historical behavioural features.

📄 Content

Automated accounts, or social bots, play an increasingly prominent role on social media platforms. They are deployed at scale to amplify hashtags, push links, promote commercial products, and shape political and health-related narratives [Ferrara et al., 2016, Cresci, 2020]. Concerns about their impact have led to a large body of work on bot detection, much of which constructs feature-based models from network, content, and account metadata [Varol et al., 2017, Gilani et al., 2019, Wu et al., 2018, Loyola-González et al., 2019, Heidari et al., 2021].

Most of these systems treat behaviour as essentially static: features are computed from a snapshot of activity and used to train classifiers that are then assumed to remain valid over time. Yet both intuition and empirical work suggest that bots respond to platform policies, detection tools, and evolving user practices [Cresci et al., 2017, Pozzana and Ferrara, 2020, Lee et al., 2011]. If bot operators adapt their strategies, then features that once reliably distinguished bots from humans may drift, potentially degrading classifier performance.

Understanding whether bots change their behaviour over time is therefore crucial for both theory and practice. Prior work has examined bot and human behavioural differences [Gilani et al., 2019, Varol et al., 2017], long-term spam campaigns and content polluters [Lee et al., 2011], and the dynamics of coordinated botnets [Pozzana and Ferrara, 2020]. However, we still know relatively little about how the behavioural signatures of labelled promotional bots evolve over periods of many years.

Promotional bots-accounts that systematically advertise products, brands, events, or services-are of particular interest. They rely heavily on URLs, hashtags, and media to drive traffic, and often exploit sentiment, emojis, and stylistic cues to appear more human-like and engaging [Al-Rawi and Shukla, 2020, Puertas et al., 2019, Loyola-González et al., 2019, Phan et al., 2020]. If such accounts change how they post and what they post over time, then bot-detection models that rely on static assumptions may quickly become outdated. In this paper we focus on a simple but fundamental question:

Research question. Is there systematic evidence that promotional social bots change their behaviour over time, and if so, along which behavioural dimensions do these changes occur?

We operationalise this question in three parts:

• RQ1: Are core behavioural meta-features of promotional bots (e.g. tweeting, URLs, hashtags, sentiment, emojis) stationary over time, or do they exhibit non-stationary trends?

• RQ2: Do behavioural meta-features differ systematically across generations of bots introduced to the platform at different periods?

• RQ3: Do behavioural meta-features differ systematically across bots with different lifespans (short-, mid-, and long-lived accounts)?

To answer these questions we build on a longitudinal dataset of promotional Twitter bots collected between 2006 and 2021, focusing on a 12-year window (2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018)(2019)(2020) where full-year coverage is available. Using ten content-oriented meta-features extracted from 2.8 million tweets produced by 2,615 bot accounts, we construct yearly time series, perform stationarity tests, and analyse trends across generations and age classes.

The work makes three contributions:

1. We assemble and analyse a longitudinal dataset of promotional Twitter bots spanning 12 years, providing an empirical basis for studying the evolution of bot behaviour at scale.

2. We demonstrate that common behavioural meta-features used in bot-detection research are nonstationary for promotional bots, and we quantify their deterministic and stochastic trends.

3. We introduce generational and age-based stratifications of bot accounts, showing systematic differences between older and newer bots, and between short-lived and long-lived bots, across ten key behavioural dimensions.

4. We explore the nature of associations between behavioural feature pairs, distinguishing dependency from independence.

5. We analyse the dynamic changes in these relationships over time, including variations in strength (weak, moderate, strong) and direction (increasing, decreasing, or stable)

Together, these findings establish a novel framework for understanding bot adaptation strategies, providing a systematic approach to analysing how bots modify their behaviours in response to changing environments over time. The results underline the need to treat social bots as dynamic adversaries whose behaviour changes over time, and they highlight concrete dimensions along which this evolution can be observed.

Online social platforms have become key infrastructures for news consumption, political debate and everyday interaction. Alongside human users, they now host large numbers of automated and semi-automated accounts-commonly referred to as social bots-that generate, amplify, or curate co

This content is AI-processed based on ArXiv data.