The research, published in the open-access journal Discover Artificial Intelligence, arrives at a moment when the commercial stakes of short-video popularity could hardly be higher. TikTok’s monthly active user base has surpassed 1.6 billion people, average user sessions stretch beyond 97 minutes, and platform advertising revenue has reached roughly 23 billion dollars. Whether a video rapidly accumulates plays, likes, comments and shares determines not only a creator’s income but also how platforms allocate advertising resources and how recommender systems rank content in an endless stream of competing clips. Against this backdrop, short-video popularity prediction has matured into a core research problem in information retrieval and computational social science, with prior approaches leaning heavily on visual content analysis, textual semantics and social propagation dynamics.

What the research team, led by Di Wu, noticed was a persistent blind spot. User activity on social platforms follows pronounced circadian rhythms: engagement during peak afternoon hours can exceed late-night engagement by more than sixfold, and consumption patterns shift fundamentally between weekdays and weekends. Yet despite this strong time dependence, virtually no existing prediction framework treats the posting moment as an independent, first-class feature. The few studies that do incorporate temporal information rely on raw timestamps or linearly normalized values, which destroy the circular structure of time. Under a linear encoding, 23:00 and 00:00 appear numerically distant—separated by 23 units—when in behavioral terms they are nearly identical points on the daily cycle. This mismatch, the researchers argue, has prevented models from learning the periodic regularities embedded in hours, weekdays and months.

TimePop’s central innovation is a component the authors call the Temporal Embedding Module, or TEM. The module takes the Unix timestamp of a video’s upload and decomposes it into multiple temporal granularities: the hour of day, the day of the week and the month. Hour and month are encoded using paired sine and cosine functions, a mathematical trick that maps each cyclic variable onto points of a circle so that adjacent times of day land close together in the model’s representation space. This design borrows its mathematical foundation from the sinusoidal positional encoding introduced in the Transformer architecture, but adapts it specifically to the posting-time scenario. The weekday is represented as a seven-dimensional one-hot vector, supplemented by binary indicators for weekends and for the peak window between 15:00 and 18:00 UTC, along with a normalized variable capturing the video’s position within the dataset’s roughly three-year span. Together these components form a compact 14-dimensional temporal vector, which two fully connected layers project into a richer 64-dimensional embedding.

The temporal embedding does not work alone. TimePop fuses three streams of information: the temporal vector, a social feature vector capturing the creator’s follower count, following count, cumulative profile likes and total video output, and a lightweight text representation built by hashing the hashtags in the video description. A feature-wise gating mechanism—an additive attention-style network that assigns a normalized importance weight to each dimension of the concatenated representation—dynamically reweights the combined signal before a prediction head outputs four estimates simultaneously: expected play count, like count, comment count and share count. The authors deliberately kept the text layer simple, finding that swapping their hashtag hashing scheme for a frozen BERT encoder improved accuracy only marginally while multiplying inference latency roughly 47-fold. The entire model contains just 31,200 trainable parameters and occupies less than half a megabyte on disk.

To test the framework, the team assembled a dataset of 2,200 TikTok videos from 15 creators, spanning June 2019 to February 2022, with interaction counts recorded at the time of metadata retrieval. The popularity distribution was heavily right-skewed—the median play count stood at 224,900 while the mean exceeded 1.29 million—so the researchers applied a logarithmic transformation to the targets and trained the model with a joint Huber loss, which dampens the influence of extreme outliers. Benchmarked against seven baselines, including linear regression, random forests, XGBoost, LightGBM, a multilayer perceptron, an LSTM and a Transformer, TimePop swept all five evaluation metrics. Compared with the strongest baseline, the Transformer, it cut the mean absolute percentage error from 40.94 percent to 33.47 percent and raised the coefficient of determination from 0.6401 to 0.7512, all while training faster—198.4 seconds versus 312.8 seconds.

Ablation experiments pinpointed the temporal module as the engine of the improvement. Removing TEM caused the largest performance drop of any component, knocking 0.0789 off the coefficient of determination, while merely replacing the cyclic sine-cosine encoding with a linear timestamp cost 0.0531—direct evidence that preserving the circular geometry of time, not just adding parameters, drives the gains. The benefits were most dramatic for the videos that matter most commercially: in the tier of videos exceeding 10 million plays, TimePop reduced prediction error by 12.3 percentage points relative to the Transformer baseline, and by 6.3 points in the 1-to-10-million tier. The dataset itself showed why timing carries such signal: 34.2 percent of videos were posted during the 15:00 to 18:00 UTC window, and videos uploaded in that interval had a median play count more than six times higher than those posted between midnight and 05:00 UTC.

The authors also confronted two methodological concerns head-on. A leave-one-creator-out cross-validation, in which the model was retrained from scratch while all videos from one creator were held out, confirmed that TimePop’s advantage persists on unseen producers, with a mean coefficient of determination of 0.681 versus 0.587 for the Transformer. A separate confound-control experiment examined whether the model was simply learning video age or platform growth rather than genuine timing effects: a degenerate model using only the days-since-start variable reached a coefficient of determination of just 0.619, and removing that variable from the full model cost only 0.024, indicating that cyclic intra-day and intra-week patterns—not exposure duration—are the primary source of the gain. The team candidly notes that the dataset’s small creator pool and wide follower variance, ranging from zero to 22.3 million, inflate absolute accuracy figures, and that timestamps were decoded in UTC without creator-local timezone correction, so the observed peak window reflects a platform-level pattern rather than a universal local-time optimum.

The practical implications ripple outward in three directions. For creators, particularly smaller accounts that depend on algorithmic recommendation rather than subscription feeds, the model offers a data-driven reference for scheduling uploads, though the researchers caution that the association shows temporal sensitivity, not proof that changing posting time alone causes greater exposure. For platforms processing tens of millions of daily uploads, TimePop’s metadata-only design—requiring no video frames or audio extraction—means it can slot into real-time ranking pipelines as a lightweight popularity prior; its single-sample CPU latency of 1.83 milliseconds sits comfortably within typical online serving budgets, and its GPU inference is more than 14 times faster than the Transformer baseline. For advertisers, accurate forecasts open the door to predictive placement, deploying budgets on videos and creators before virality materializes rather than reacting after the fact.

Limitations remain, and the authors map them out with unusual candor. The model predicts cumulative interaction counts rather than the evolving popularity trajectory over days one, three and seven; it has been validated only on TikTok, leaving cross-platform transfer to Instagram Reels or YouTube Shorts unverified; and it ignores content quality signals from visuals and audio entirely. Future work, the team says, will scale the dataset through official research APIs, fuse the temporal module with multimodal content encoders, and apply causal inference methods to disentangle the true effect of posting time from confounders such as trending topics and platform traffic injections. Even so, the study makes a compelling case that in the attention economy, one of the cheapest pieces of metadata a platform records—the moment of upload—may be among the most informative predictors of whether a video soars or sinks.

Subject of Research: A posting-time-aware deep learning model for predicting the popularity of short videos on TikTok

Article Title: Timepop a posting time aware deep model for short video popularity prediction

Article References: Wu, D., Zhou, Y., Li, H., & Li, D. (2026). Timepop a posting time aware deep model for short video popularity prediction. Discover Artificial Intelligence, 6(1), Article 1286. https://doi.org/10.1007/s44163-026-02322-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02322-9

Keywords: short-video popularity prediction, TikTok, deep learning, temporal embedding, posting time, cyclic encoding, feature-wise gating, social media analytics, recommender systems, viral content, multimodal fusion, predictive modeling