Predicting Post Virality with Temporal Cross-Attention over Trend Signals
Sarvagya Somvanshi Mohan Xu [email protected] [email protected] Rakhi Chadalavada Nathan Canera [email protected] [email protected] · 2026
Read Predicting Post Virality with Temporal Cross-Attention over Trend Signals by Sarvagya Somvanshi Mohan Xu [email protected] [email protected] Rakhi Chadalavada Nathan Canera [email protected] [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-02358.txt.
Introduction
Sarvagya Somvanshi, Mohan Xu, Rakhi Chadalavada, and Nathan Canera predict which Reddit posts will go viral by adding real-world trend signals to post features. Past models lean on static text and structure, ignoring how outside attention shifts over time. The authors ask whether exogenous signals can add predictive power beyond the post itself.
Their model, ViralityNet, fuses title and body embeddings, structural metadata, and learned subreddit embeddings. A temporal cross-attention block queries a daily sliding-window matrix of trend terms, built from the strongest Wikipedia pageview spikes in preceding days. A post counts as viral if it passes the ninetieth percentile of engagement for its subreddit and a minimum score. The design tests trend timing as a virality signal.
Catalog record
| CREATOR | Sarvagya Somvanshi Mohan Xu [email protected] [email protected] Rakhi Chadalavada Nathan Canera [email protected] [email protected] |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.02358
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Predicting Post Virality with Temporal Cross-Attention over Trend Signals Sarvagya Somvanshi Mohan Xu [email protected] [email protected] Rakhi Chadalavada Nathan Canera [email protected] [email protected] Abstract Current models for predicting social media virality rely heavily on static textual and structural features, effectively ignoring the highly dynamic nature of trend signals. We study whether real-world attention signals can improve the prediction of social-media virality beyond what post text alone reveals. We introduce ViralityNet , an architecture that predicts Reddit post virality by fusing internal platform representations with exogenous temporal signals derived from Wikipedia pageview spikes. We frame virality as a binary classification task that accounts for differences in subreddit scale, labeling posts as viral if they exceed the 90th percentile of per-subreddit engagement and a minimum absolute score threshold. ViralityNet combines four post-level streams—title embeddings, body embeddings, structural metadata, and learned subreddit embeddings—with a cross-attention block that queries a daily sliding-window trends matrix encoding the top-512 Wikipedia spike terms from the preceding seven days. Empirical results suggest that incorporating external attention signals yields consistent gains, outperforming text-only baselines by +0.015 AUC-PR and achieving an overall AUC-ROC of 0.836. Overall, we provide evidence that incorporating external attention signals yields measurable improvements over text-only baselines, highlighting the importance of real-world dynamics in shaping online virality. \nipsfinalcopy 1 Introduction Understanding what makes online content go viral has attracted sustained research attention from human-computer interaction, computational social science, and natural language processing communities cheng2014can ; szabo2010predicting . Most prior work has focused on structural features including follower counts, posting time, network centrality, and textual semantics tan2014effect . However, a key limitation of purely text-based models is that they are temporally blind : a post about a topic that was intensely discussed last month may carry very different viral potential from an otherwise identical post written during a week when that topic is spiking in public attention. This limitation matters because virality is not determined by content alone, but also by how well that content aligns with what the public cares about at a given moment. Models that ignore these temporal dynamics risk mischaracterizing the drivers of engagement, limiting both predictive accuracy and scientific understanding of online behaviour. We hypothesise that internet-wide attention signals, measured via daily Wikipedia pageview spikes, serve as a proxy for collective public interest that can inform virality prediction on Reddit. When a topic is rapidly gaining attention (reflected in a sudden surge of Wikipedia reads), Reddit posts about that topic are more likely to find a receptive audience. Contributions. We make three contributions: • We introduce a post-level trend alignment signal derived from Wikipedia pageview spikes and show it is statistically significant at the individual post level. • We propose ViralityNet , a cross-attention architecture that conditions post representations on daily sliding-window trend matrices, together with learned year and subreddit embeddings. • We conduct a systematic ablation study that disentangles the contributions of temporal context (year embeddings), exogenous trends (cross-attention), and their combination. Virality score. We define a per-subreddit engagement score that combines upvotes and comments: v=s+\beta\cdot c,\quad\beta=0.3, (1) where s is the post score (upvotes minus downvotes) and c is the comment count. The comment weight \beta=0.3 reflects that comments represent greater engagement effort than passive voting; a sensitivity analysis (Appendix B ) confirms that varying \beta\in[0.1,0.5] flips fewer than 0.2% of labels. Binary labelling. We adopt a subreddit-stratified top-10% labelling strategy : within each subreddit, a post is labelled viral ( y=1 ) if v exceeds the 90th percentile of that subreddit’s score distribution and the raw score s\geq 100 . All posts falling below this threshold are strictly retained in the dataset as negative samples ( y=0 ). The absolute-score floor prevents posts from tiny subreddits from being labelled viral despite negligible real-world engagement, while retaining the overwhelming majority of posts as negative examples provides the necessary class imbalance to train a robust discriminative model. Per-subreddit stratification is important because engagement scales differ by orders of magnitude across communities. 2 Related Work Virality and popularity prediction. Predicting which content will achieve outsized engagement has been studied across platforms and modalities. Szabo and Huberman szabo2010predicting demonstrated a log-linear relationship between early view counts and eventual popularity, showing that the first hours of engagement strongly predict final reach. Cheng et al. cheng2014can extended this to information cascades, finding that cascade structure carries predictive signal beyond simple popularity counts. Tatar et al. tatar2014predicting found that news article engagement follows a heavy-tailed distribution, with the top 10% of articles on major European platforms attracting roughly half of all comments. This skewed pattern both justifies treating virality as a binary threshold and demonstrates that even simple feature sets can yield meaningful predictions. On Reddit specifically, Lakkaraju et al. lakkaraju2013whats studied the role of title phrasing and submission timing, while Hessel et al. hessel2017cats used multimodal content together with community and temporal features to predict post popularity. Most of these approaches treat posts in isolation from broader societal context; our work differs by conditioning predictions on exogenous signals of real-world public attention. Exogenous signals and external context. A growing body of work recognises that engagement spikes are often driven by events external to the platform. Iamnitchi et al. iamnitchi2023modeling showed that while in-platform signals capture baseline activity, they fail to anticipate the sudden engagement surges tied to external events, which is a gap that exogenous features can fill. El-Amrany et al. elamrany2025viralreddit introduced the Reddit-V benchmark of 25,000 posts across five subreddit categories and evaluated several large language models for virality prediction, concluding that task-specific supervised learning outperforms zero-shot prompting. Their labelling strategy (top 20% within each subreddit) is related to ours, though we adopt a stricter 90th-percentile threshold with an absolute-score floor and additionally incorporate temporal trend signals that prior Reddit-focused studies leave unexplored. Wikipedia pageviews as an attention proxy. Wikipedia pageview data has been shown to be a reliable real-time proxy for public attention. Yasseri and Bright yasseri2016wikipedia showed that Wikipedia pageview time series can serve as a proxy for collective public attention, demonstrating that surges in views of politically relevant articles track real-world information-seeking behaviour around elections. Wikipedia pageview counts are particularly attractive as an attention signal because they are publicly available at hourly granularity, continuously recorded since 2007, and free of the algorithmic ranking and personalisation effects that confound platform-internal trend signals. Mestyán et al. mestyan2013early exploited Wikipedia activity to predict movie box-office success, establishing a precedent for using encyclopedia traffic as a leading indicator of real-world interest. We extend this line of work by encoding Wikipedia spike terms as dense embeddings and integrating them via cross-attention, allowing the model to learn which trending topics are semantically relevant to each post rather than relying on hand-crafted alignment features. Cross-attention and multi-stream fusion. Cross-attention mechanisms have become a standard tool for fusing heterogeneous information streams. In vision–language models, cross-attention conditions one modality on another to capture fine-grained alignment lu2019vilbert . Our architecture applies the same principle to a novel pairing: post-level representations query a temporal trends matrix, enabling the model to selectively attend to the subset of recent public-interest topics most aligned with the post’s content. 3 Methodology 3.1 Model Overview PostEncoder CrossAttnBlock Classifier Title all-mpnet \cdot 768d Body all-mpnet \cdot 768d chunk mean-pool +