Skip to content
Open access

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

Sep 2026 · Discover Artificial Intelligence · Vol 6 · 0 citations · 38 references
Computer Science

TL;DR

TimePop is a posting-time-aware deep learning framework for short-video popularity prediction that combines multi-granularity temporal attributes of the posting moment into a 14-dimensional continuous vector via sinusoidal-cosinusoidal cyclic encoding, effectively preserving the circular continuity of temporal signals.

Abstract

SShort-video platforms have experienced explosive growth, making automated popularity prediction one of the core challenges in social media analytics and recommender systems. Existing methods predominantly focus on visual content or social network features, while systematically neglecting the temporal signal embedded in the video posting time—a signal with strong prior predictive value. To address this gap, we propose TimePop, a posting-time-aware deep learning framework for short-video popularity prediction. The core innovation of TimePop lies in a dedicated Temporal Embedding Module (TEM): multi-granularity temporal attributes of the posting moment (hour, weekday, month) are encoded into a 14-dimensional continuous vector via sinusoidal-cosinusoidal cyclic encoding, effectively preserving the circular continuity of temporal signals. Building upon this, a feature-wise gating fusion mechanism is designed to dynamically weight and fuse temporal embeddings with social features and textual features, constructing an end-to-end four-dimensional joint popularity prediction model (play count, like count, comment count, share count). On a self-collected TikTok dataset of 2200 videos from 15 creators (June 2019–February 2022), where interaction counts were collected at the time of metadata retrieval, TimePop achieves leading performance across all five evaluation metrics against seven baseline methods: compared with the best-performing baseline Transformer, MAPE decreases from 40.94 to 33.47% (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\downarrow $$\end{document}7.47 percentage points), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document} improves from 0.6401 to 0.7512 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow $$\end{document}0.1111), and training time (198.4 s) is lower than that of Transformer (312.8 s). Ablation experiments confirm that TEM is the most critical performance source (removing TEM causes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\downarrow $$\end{document}0.0789), and the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document} gain of sinusoidal-cosinusoidal cyclic encoding over linear encoding reaches 0.0531. Stratified analysis reveals that TimePop reduces MAPE by 12.3 percentage points relative to the baseline in the high-popularity segment (play count >10 M), demonstrating the particularly strong discriminative value of temporal embedding features for predicting highly viral videos. Leave-one-creator-out cross-validation further confirms that TimePop’s advantage over the Transformer baseline is maintained across unseen creators (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}: 0.681 vs. 0.587), demonstrating cross-creator generalization beyond a single random split. Using only lightweight metadata as input, this study provides a practically meaningful time-aware prediction paradigm for content creator posting-time analysis, platform metadata-based recommendation support, and predictive advertising placement.

Read PDF

Similar papers

Streaming Video Moment Retrieval via Temporal Forecast Diffusion Model

With the advancement of online media, the demand for real-time interaction with streaming video has rapidly increased, highlighting the importance of Streaming Video Moment Retrieval (SVMR). This task aims to evaluate the relevance between streaming video and textual queries in real-time. Unlike offline videos, streami...

Yan-Hua Yang, Jia-Hua Li, Zhe Xu et al. · 0 citations
Conference Sep 2026

A lightweight graph convolutional network-based session recommendation method based on multihop attention mechanism

A novel Multi-Hop Lightweight Graph Convolutional Network (MH-LGCN) conversational recommendation model based on a multi-hop attention mechanism that abandons redundant operations in traditional graph convolution and designs a lightweight information propagation mechanism that retains only neighbor feature aggregation.

Xua Li, Ru-Jun Wang · 0 citations
Preprint Sep 2026

Beyond a Scalar: Distributional Serving Interfaces for Watch-Time Prediction

Watch time is the primary engagement signal in short video feeds, and its prediction directly affects ranking and exposure. Existing methods improve watch time prediction by correcting duration bias or modeling richer distributions, but most expose only an expected or debiased watch time at serving time. Even when vide...

Xuan Liu, Jing Qian, Zhan-Yu Liu et al. · 0 citations
Preprint Oct 2026

Rethinking User Encoders: Routing Diffusion States for Scenario-Adaptive User Representation

This work proposes FINDERS (FactorIzatioN of Diffusion statEs for Routing across Scenarios), a router-equipped dLLM framework comprising four operations: denoise, factorize, route, and compress that effectively pushes the boundaries of dLLMs beyond language generation, demonstrating their immense potential as powerful...

Jun-Ru Zhang, Jin-Yong Wen, Bin Dou et al. · 0 citations
Book Open access Sep 2026

DP-Rec: Towards Dynamic Patching for Efficient Long-Sequence Recommendation

Transformers have redefined sequential recommendation by effectively modeling dynamic user behaviors and long-range dependencies. However, they remain inherently inefficient: standard architectures operate at a fixed rate, allocating comparable computation to every item in a user’s history regardless of its information...

Dwipam Katariya, T. Caputo, Akshat Shreemali et al. · 1 citation
Open access Oct 2026

Personalized video recommendation with attentive multi-view learning and long-term and short-term interest modeling

Personalized video recommendation plays a crucial role in alleviating information overload on large-scale video platforms. However, existing approaches largely rely on identification (ID)-based features, which are insufficient for capturing the rich semantic information embedded in video content and user behaviors. To...

You-Wei Yuan, Lan-Jun Luo, Kun Wang et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.