Source-linked AI summary
Sequential Prediction of Social Media Popularity with Deep Temporal Context Networks
Bo Wu, Wen-Huang Cheng, Yongdong Zhang, Qiushi Huang, Jintao Li, Tao Mei
TL;DR
Social-media popularity prediction has largely neglected the sequentiality of social data. DTCN models temporal contexts and multi-scale temporal attention, and outperforms state-of-the-art deep prediction algorithms on sequential popularity prediction.
Problem
Existing social-media popularity predictors often neglect temporal sequentiality, limiting their success in sequential data scenarios such as news feeds, tweet timelines, and photo streams.
Method
DTCN jointly embeds multi-modal user-photo data, learns neighboring and periodic temporal contexts, and applies multiple time-scale temporal attention to predict future-photo popularity.
Results
5.79%–44.86% relative improvements: DTCN outperforms state-of-the-art deep prediction algorithms on sequential social-popularity prediction.
Takeaways & Limitations
Modeling both temporal and sequential coherence with temporal contexts and attention improves sequential popularity prediction over state-of-the-art deep prediction algorithms.
Takeaways & Limitations
Future work should incorporate social network structure and the influence of influential users into popularity prediction.
Abstract
from arXiv · showhide
Prediction of popularity has profound impact for social media, since it offers opportunities to reveal individual preference and public attention from evolutionary social systems. Previous research, although achieves promising results, neglects one distinctive characteristic of social data, i.e., sequentiality. For example, the popularity of online content is generated over time with sequential post streams of social media. To investigate the sequential prediction of popularity, we propose a novel prediction framework called Deep Temporal Context Networks (DTCN) by incorporating both temporal context and temporal attention into account. Our DTCN contains three main components, from embedding, learning to predicting. With a joint embedding network, we obtain a unified deep representation of multi-modal user-post data in a common embedding space. Then, based on the embedded data sequence over time, temporal context learning attempts to recurrently learn two adaptive temporal contexts for sequential popularity. Finally, a novel temporal attention is designed to predict new popularity (the popularity of a new user-post pair) with temporal coherence across multiple time-scales. Experiments on our released image dataset with about 600K Flickr photos demonstrate that DTCN outperforms state-of-the-art deep prediction algorithms, with an average of 21.51% relative performance improvement in the popularity prediction (Spearman Ranking Correlation).
1 Introduction
The paper frames social-media popularity prediction as a sequential task and introduces DTCN to model temporal context and attention across multiple time-scales. Its architecture combines multi-modal embedding, recurrent temporal-context learning, and temporal attention.
- Existing popularity predictors often ignore temporal order, limiting their success on sequential streams such as news feeds, tweet timelines, and photo streams.
- DTCN jointly models temporal and sequential coherence rather than treating time only as a latent factor or non-sequential feature.
- Neighboring Temporal Context captures short-term trends, while Periodic Temporal Context captures discontinuous long-term patterns associated with periodic events or activities.
- The framework maps user and visual features into a shared embedding space, then recurrently learns adaptive temporal contexts before prediction.
- DTCN integrates embedding, temporal-context learning, and multiple time-scale temporal attention in an end-to-end sequential prediction architecture.
- The study identifies consideration of both temporal and sequential characteristics as its central contribution to sequential social-media popularity prediction.
2 Related Work
Related work largely models popularity growth from early observations or uses temporal features and dynamic signals without modeling popularity sequentiality. The paper instead predicts popularity before publication using sequential temporal context.
- Existing popularity-growth methods require early-stage popularity patterns from an already published post.
- The paper predicts a corresponding post’s popularity before it is published, extending beyond methods that depend on observed post-popularity trajectories.
- Prior temporal approaches use variables, point processes, graphs, or contextual factors, but the reviewed models neglect sequentiality during prediction.
- The proposed perspective explores temporal context from sequential data and considers both sequential and temporal characteristics during prediction.
3 Social Media Popularity Prediction
The paper formulates Flickr photo-popularity prediction from ordered user-photo activity and defines temporal context and multiple time-scales for sequential modeling. Popularity is tied to expected attention, especially photo views, with normalization for large count variation.
- The task estimates how many attentions a new user-photo pair will receive after publication, including views, likes, or clicks.
- For Flickr, viewing count is used as the indicator of photo popularity because users can inspect photo details through thumbnails.
- Popularity values are log-normalized to suppress variation in photo view counts ranging from zero to millions.
- A user-photo sequence is an ordered set of user-photo pairs arranged by their sharing times.
- Temporal context is a time series built from a target’s previous user-post sequence and represented as ordered user, visual, and time triples.
- The paper defines four time-scale levels: minute of an hour, period of a day, day of a week, and week of a month.
4 Deep Temporal Context Network
DTCN predicts future popularity from time-ordered user-photo sequences by combining multimodal embedding, adaptive temporal-context learning, and multi-timescale temporal attention.
- DTCN takes a time-ordered user-photo sequence as input and predicts the popularity of a future photo.The framework is trained end-to-end with Multi-modal Joint Embedding, Temporal Context Learning, and Multiple Time-scale Temporal Attention.
- Multi-modal Joint Embedding maps user and photo information into a unified deep representation using separate user and photo analysis pipelines.The two-stream feedforward network uses nonlinear mappings and dropout, producing a 64-dimensional embedding representation when jointly trained with the full framework.
- A two-stream LSTM learns sequential and temporal coherence from the NTC and PTC sequences with Mean Squared Error as the optimization objective.The learning process uses temporal information alongside contextual information to model both short-term and long-term coherence.
- Temporal Context Learning constructs adaptive Neighboring Temporal Context and Periodic Temporal Context to represent short-term fluctuations and long-term periodic patterns.NTC uses neighboring items, whereas PTC uses discontinuous items aligned across recurring time blocks.
- Multiple Time-scale Temporal Attention weights contextual items using relative hidden states and multi-scale time vectors to incorporate temporal consistency into prediction.The attention score compares the target and prior context through cosine distance between multi-scale time vectors rather than relying only on hidden-state attention.
5 Experiments
Experiments evaluate DTCN and baselines on TPIC17 image sequences using SRC and MAE across multiple dataset sizes. Results compare sequential models, temporal-context variants, and single-context settings.
- Experimental Setup: TPIC17 contains 680K Flickr photos spanning over three years, with 100K, 200K, and 400K subsets evaluated using five-round moving-partition validation.Data are organized in time order and evaluated with recurrently moving training and testing windows.
- Experimental Setup: The evaluation uses Spearman Ranking Correlation (SRC) for ranking agreement and Mean Absolute Error (MAE) for averaged prediction error.SRC ranges from 0 to 1 and measures correlation between ground-truth and predicted popularity sets.
- Compared Methods: Compared methods include CNN-AlexNet, CNN-VGG, SVR, SVR(T), MLP, MLP(T), LSTM, and Contextual LSTM.CNN baselines use only original image files, whereas SVR and MLP variants incorporate user, visual, or temporal features as specified.
- Prediction Performance: DTCN achieves the best performance across all three dataset sizes, reaching SRC 0.6692 and minimum MAE 1.2341.It provides an average 21.51% relative SRC improvement over MLP, MLP(T), LSTM, and CLSTM.
- Prediction Performance: Sequential models outperform non-sequential CNN, SVR, and MLP approaches, while adding temporal information without sequential coherence provides limited predictive power.LSTM and CLSTM reach SRC values of 0.4629 and 0.6072, respectively, below DTCN.
- Temporal Context Analysis: Combining NTC and PTC improves SRC over either context alone, with single-context results falling from 0.5990 to 0.5875 and 0.5745.NTC settings outperform PTC settings on the 100K Flickr dataset, indicating stronger short-term patterns there.
6 Conclusions and Future Work
The paper presents DTCN as a framework for sequential social-media popularity prediction and identifies social-network structure and influential users as future directions. The released paper and dataset are intended to support further research.
- Conclusions: DTCN combines joint multi-modal embedding, short- and long-term temporal-context alignment, and multi-timescale temporal attention for sequential popularity prediction.The framework learns temporal coherence in short-term and long-term popularity fluctuations.
- Conclusions: On TPIC17, the prediction network outperforms state-of-the-art deep prediction algorithms by 5.79%–44.86% relative improvements.The conclusion characterizes the model as a deep prediction framework for sequential popularity prediction.
- Future Work: Future work includes incorporating social-network structure and analyzing the impact of influential users on popularity prediction.These are identified as open directions for social popularity prediction.