[Tsinghua University] HENN: Bridging the Heterogeneity Gap in Personality Prediction via Entropy

TL;DR

Predicting personality from social media is notoriously difficult due to the "Heterogeneity Gap"—the fact that a user's blog post, their profile picture, and their "like" frequency follow completely different mathematical distributions. In this paper, researchers from Tsinghua University propose HENN (Heterogeneity Entropy Neural Network). By introducing an unsupervised loss function based on Heterogeneity Entropy, the model learns a unified representation of a user’s "inner self" across diverse media, outperforming traditional multi-modal methods like CCA and Corr-AE by a wide margin.

The Core Challenge: Semantic and Heterogeneity Gaps

In psychological modeling, we face two walls:

1. The Semantic Gap: How do you go from a raw SIFT feature in an image to the abstract concept of "Extraversion"?
2. The Heterogeneity Gap: How do you mathematically combine a 1000-dimensional text vector with a time-series of "share" counts when they share no common physical units?

While prior works used Canonical Correlation Analysis (CCA) to find a shared subspace, they often struggled with the non-linear complexities of social media data.

Methodology: The Architecture of HENN

The authors propose a hierarchical approach that systematically breaks down these gaps.

1. HE-DBN (The Bottom Layer)

Each modality (Text, Image, Behavior) first passes through its own Deep Belief Network (DBN). This layer's job is to reduce the "Semantic Gap" by converting low-level noise into middle-level features. Interestingly, the authors use the Heterogeneity Entropy to decide when to stop adding layers—if entropy stops decreasing, the abstraction is sufficient.

2. HE-AE (The Middle Layer - The "Brain")

This is where the magic happens. An Auto-encoder takes the outputs from all DBNs. The loss function is modified to include the Heterogeneity Entropy (HE): $J=extReconstructionLoss+\lambda (extSparsity)+\mu \mathbf{HE}(\mathbf{F})$ Minimizing HE is mathematically equivalent to maximizing the Mutual Information between modalities. It forces the network to find features that are present across all media types, effectively bridging the "Heterogeneity Gap."

3. Common-DBN (The Top Layer)

A final DBN stack refines these merged features into a highly abstract "Personality Vector," which is then fed into a Support Vector Regressor (SVR) for final prediction.

Results: Why Entropy Matters

The researchers tested HENN on a dataset from Renren and Sina Microblog. The results were striking:

• Superior Accuracy: For the "Openness" trait, HENN achieved a PCC of 0.8810, compared to 0.8283 for CCA-based methods.
• Modality Synergy: Using Text, Image, and Behavior together resulted in a 41-60% reduction in MAE compared to using any single modality alone.
• Ablation Logic: Interestingly, the study found that Behavior Statistics (like sharing frequency) were the "noisiest" indicators, while the HE-DBN layers were the most critical for performance stability.

Critical Insights & Future Outlook

The standard approach to multi-modal learning is often "Concatenate and Hope." HENN proves that principled information theory—specifically entropy-based alignment—is a much more robust way to handle heterogeneous data.

Limitations: The ground truth for this study was based on "Observer Impressions" (volunteers watching users) rather than self-report questionnaires. While this is practical for real-world apps (like AI recruiters), it captures "perceived personality" rather than "true internal personality," which may differ.

Future Work: The logical next step is applying HENN to transformer-based architectures or exploring how these entropy-aligned features handle "missing modalities" (e.g., predicting personality when a user has no images).

Takeaway for Practitioners: When dealing with cross-media data, don't just stack layers. Focus on the Mutual Information between your inputs; it is the most reliable signal for high-level semantic extraction.