OmniTrend: Content-Context Modeling for Scalable Social Popularity Prediction

1. Problem Statement

Social media popularity prediction aims to estimate user engagement (e.g., likes, shares, views) for user-generated content. The paper identifies a fundamental limitation in existing methods: the entanglement of content attractiveness and contextual exposure.

• The Core Challenge: Observed engagement ($y$) is a joint outcome of two distinct factors:
  1. Intrinsic Attractiveness ($A$): The quality and appeal of the content itself (visual, audio, text).
  2. Contextual Exposure ($E$): The external factors determining how widely the content is shown (posting time, author activity, platform algorithms, trending topics).
• Current Limitations: Most existing multimodal models treat content and context as a single signal. This causes the learned representations to absorb platform-specific visibility biases, leading to:
  • Poor Interpretability: It is unclear why a post is popular (is it the content or the timing?).
  • Weak Transferability: Models trained on one platform often fail on another because they have learned the specific exposure patterns of the source platform rather than universal content appeal.
  • Unstable Supervision: The same content can receive different engagement labels depending on when or where it is posted.

2. Methodology: OmniTrend Framework

OmniTrend proposes a Content–Context Modeling framework that explicitly separates the prediction task into two additive components in log-space:
$$\log(1 + y_i) = \alpha_i + \phi_i$$
Where $\alpha_i$ represents content attractiveness and $\phi_i$ represents contextual exposure.

The architecture consists of three main components:

A. Cross-Platform Content Module (Learning $\alpha$)

• Goal: Learn a platform-agnostic representation of intrinsic appeal.
• Inputs: Multimodal signals (Images, Videos, Text).
• Architecture:
  • Encoders: Uses pretrained encoders (CLIP for images, X-CLIP for video, BGE-M3 for text) to map inputs to a shared 512-dimensional embedding space.
  • Cross-Modal Attention: Employs bidirectional cross-attention to align visual, textual, and video tokens, allowing the model to refine features based on interactions between modalities.
  • Gated Fusion: A gating mechanism controls the contribution of each modality, making the model robust to missing modalities (e.g., a post with only an image and no text).
  • Training: Trained jointly across multiple platforms to learn universal appeal. It uses a combination of Weighted Huber Loss, Pairwise Ranking Loss, and Cross-Modal Alignment Loss.

B. Platform-Specific Context Module (Learning $\phi$)

• Goal: Estimate exposure dynamics based on exogenous signals.
• Inputs: Metadata and external factors (posting time, author follower count, activity frequency, topic categories, geographic location).
• Constraint: Strictly excludes raw content embeddings to prevent leakage of attractiveness signals into the exposure model.
• Architecture: Uses a lightweight model (e.g., CatBoost or MLP) to predict the exposure component based on the context vector.
• Training: Trained on the residual target ($\tilde{y} - \hat{\alpha}$), effectively learning what the content module cannot explain.

C. Retrieval-Augmented Neighborhood Module

• Goal: Capture temporal diffusion and social contagion effects without mixing raw content into the exposure estimator.
• Mechanism:
  • Retrieves the top $K$ historical posts similar to the current post based on the Content Module's embeddings.
  • Computes aggregated statistics (weighted mean and variance) of the neighbors' popularity and author activity.
  • These statistics (not the raw content) are fed into the Context Module as "neighborhood features."
• Benefit: This allows the model to learn how popularity propagates over time (momentum) while maintaining a clean separation between what the content is and how it is exposed.

3. Key Contributions

1. Decoupled Framework: The first unified framework to explicitly separate social popularity into attractiveness (content) and exposure (context) components, providing a principled basis for interpretability and transfer.
2. Unified Cross-Platform Content Modeling: A shared module that learns intrinsic appeal across different media types (images vs. videos) and platforms, enabling zero-shot or few-shot transfer.
3. Retrieval-Enhanced Context Modeling: A novel strategy that uses retrieval to aggregate temporal and neighborhood statistics for exposure prediction, avoiding the need for explicit graph construction while capturing social diffusion.
4. Scalability and Interpretability: The modular design allows the content model to be updated independently of the platform-specific context model, making the system adaptable to platform evolution.

4. Experimental Results

The authors evaluated OmniTrend on four diverse benchmarks:

• Datasets: MicroLens (short video), ICIP (Twitter/Instagram images), SMPD-Image (large-scale images), and SMPD-Video (large-scale videos).
• Baselines: Compared against 12 state-of-the-art methods, including feature-based (SVR), sequential (CLSTM), multimodal (HMMVED, BLIP), and retrieval-based (MMRA) models.

Key Findings:

• Performance: OmniTrend achieved State-of-the-Art (SOTA) results across all datasets.
  • On SMPD-Image, it reduced MSE by ~13% and increased Spearman Rank Correlation (SRC) from 0.64 to 0.71 compared to the best baseline (MMRA).
  • On MicroLens, it achieved the lowest MAE (0.86) and highest SRC (0.61).
• Transferability: The content module trained on video data successfully transferred to image prediction tasks with minimal performance drop, demonstrating the robustness of the learned attractiveness signal.
• Ablation Studies:
  • Removing the Context module caused a significant drop in accuracy, proving exposure is critical.
  • Removing the Content module degraded ranking consistency, proving intrinsic appeal is essential.
  • The joint distribution of predicted $\alpha$ and $\phi$ was statistically independent, confirming the model successfully disentangled the two factors.

5. Significance

• Theoretical Insight: The paper challenges the monolithic view of popularity prediction, arguing that "popularity" is a product of two distinct mechanisms. This separation explains why similar content performs differently across platforms.
• Practical Application:
  • For Creators: The model can distinguish between "bad content" (low $\alpha$) and "bad timing/algorithm" (low $\phi$), offering actionable feedback.
  • For Platforms: The framework supports scalable deployment where the core content understanding remains stable while the exposure model adapts to specific platform algorithm changes.
• Future Direction: The modular architecture sets a precedent for building adaptive AI systems that can evolve with platform dynamics without requiring complete retraining of the core semantic understanding.