Research map / Video / Video segmentation and tracking

SMTPD: A New Benchmark for Temporal Prediction of Social Media Popularity

CVPR 2025 · Video › Video segmentation and tracking · approach: Physics-inspired motion prediction · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces SMTPD, a 282K-sample multilingual YouTube benchmark with daily popularity for 30 days, and an LSTM-based multimodal baseline showing early popularity is key for temporal prediction.

Problem

Existing social media popularity datasets lack aligned temporal observations, so predictions are made at irregular times and rarely cover multimodal, multilingual content over time.

Approach

Collect new YouTube posts and observe their popularity every 24 hours for 30 days to create time-aligned popularity sequences, then regress the full 30-day sequence with a multimodal feature fusion model with LSTM decoder trained with a Composite Gradient Loss.

Key results

• Proposed baseline reaches AMAE 0.717 and ASRC 0.959 on SMTPD, surpassing the second-best method by 0.798 MAE and 0.109 SRC on day 30 (per paper's stated gap).
• Removing early popularity input raises day-30 MAE of the baseline to about 1.53-1.56 range, showing early popularity is critical.
• All compared methods achieve higher SRC on SMTPD than on SMPD, while MAE on SMPD increases by over 0.16 when moved to temporal setting.

Measured effects reported in the paper

Surprising findings

• Adding early popularity (day-1 score) as input dramatically reduces error for existing methods, suggesting predicting early popularity is the key bottleneck.
• BERT-Base instead of BERT-Multilingual raises MAE by about 0.19, indicating multilingual text encoding matters.

Limitations

• Does not yet model multi-frame video information
• Language diversity and deeper multimodal exploration left for future work
• Baseline relies on true early popularity; predicted early popularity leaves a large gap to true values

Experimental setup

Datasets: SMTPD SMPD

Benchmarks: SMTPD SMPD

Metrics: MAE AMAE SRC ASRC

Baselines: Ding et al. [13] Lai et al. [26] Xu et al. [52] SMPD

Builds on: SMPD Khosla popularity score transformation BERT-Multilingual ResNet-101

Closest papers by meaning

• Improving Multimodal Social Media Popularity Prediction via Selective Retrieval Knowledge Augmentation (AAAI 2025 · similarity 0.88)
• THGFormer: Time-Aware Hypergraph Learning for Multimodal Social Media Popularity Prediction (Student Abstract) (AAAI 2024 · similarity 0.87)
• CasFT: Future Trend Modeling for Information Popularity Prediction with Dynamic Cues-Driven Diffusion Models (AAAI 2025 · similarity 0.85)
• Public Opinion Field Effect and Hawkes Process Join Hands for Information Popularity Prediction (AAAI 2025 · similarity 0.84)
• In-context Prompt-augmented Micro-video Popularity Prediction (AAAI 2025 · similarity 0.83)
• SNS-Bench: Defining, Building, and Assessing Capabilities of Large Language Models in Social Networking Services (ICML 2025 · similarity 0.83)