Characterising Behavioural Families and Dynamics of Promotional Twitter Bots via Sequence-Based Modelling

This paper investigates whether promotional Twitter bots can be meaningfully grouped into behavioural families and whether members of a given family exhibit similar behavioural evolution over time. We analyse 2,798,672 tweets from 2,615 ground-truth promotional bot accounts active on Twitter between 2006 and 2021, restricting our analyses to complete calendar years from 2009 to 2020. We encode each bot’s activity as a sequence of symbolic blocks - a “digital DNA” representation - derived from seven categorical post-level behavioural features (posting action, URL, media, text duplication, hashtags, emojis, and sentiment), with the post timestamp retained only to preserve temporal ordering. Using non-overlapping blocks (equivalently k-mers with k = 7), we compute pairwise cosine similarities over block-frequency vectors, construct a similarity matrix, and apply hierarchical clustering to obtain bot families. We then characterise each family in terms of dominant feature usage, shared and unique blocks, and temporal changes across three life-cycle segments (beginning, middle, end). The results show that promotional bots can be clustered into four coherent families - Unique Tweeters, Duplicators with URLs, Content Multipliers, and Informed Contributors - that share common behavioural cores but differ systematically in engagement strategies and behavioural evolution.

Building on this family structure, we introduce a sequence-analysis framework for modelling bot behaviour change explicitly as behavioural mutations. Within each family, we align behavioural sequences using multiple sequence alignment (MSA), then detect and classify changes as insertions, deletions, substitutions, alterations, and identity (match) events. This allows us to quantify mutation frequencies, identify blocks most prone to change, analyse which behavioural features are most often substituted, and locate mutation hotspots along the sequences. The analysis shows that deletions and substitutions dominate while insertions are rare; that mutation patterns differ systematically across the four families; and that mutation hotspots concentrate in early sequence positions for some families while being widely distributed for others.

Finally, we evaluate whether family structure and mutation patterns support predictive reasoning about behavioural adaptation. We show that bots within the same family are more likely to share mutations than bots in different families, that closely related bots within a family share and propagate mutations more often than distant ones, and that behaviour around external trigger events (such as Christmas and Halloween) follows family-specific, partially predictable patterns. These findings support sequence-based family modelling of promotional bots together with a fine-grained, evolution-inspired account of how their behaviours adapt and evolve over time.