Introduction
In user-centric design, a "persona" represents a profile of a specific user segment sharing similar behaviors and needs. While vital for personalization, traditional persona creation is a manual and unscalable process. Modern machine learning attempts to automate this, yet faces two significant hurdles: the contextual gap between textual data and behavioral metadata, and the "black box" problem of deep learning models. Designers and clinicians require not just a classification, but an explanation of why a user belongs to a specific persona.
To address these challenges, we present ExBigBang (Explainable BigBang), a hybrid transformer-based framework that fuses textual analysis with contextual metadata while leveraging Explainable AI (XAI) for transparent decision-making.
The ExBigBang Framework
The ExBigBang architecture is built upon a hybrid transformer model designed to process disparate data types simultaneously.
As shown in the architecture diagram, the model employs a pre-trained BERT encoder for textual data and a dense network for tabular contextual features. These streams are integrated using a custom attention-based multimodal combiner, allowing the model to dynamically weight the importance of text versus metadata for each specific classification.
Contextual Feature Engineering
The framework incorporates three primary types of contextual features:
- Temporal Features: Capturing behavioral patterns over time (e.g., late-night activity ratios).
- Profiling Features: Aggregated metrics such as average sentiment over a user's history.
- Knowledge-based Features: Domain-specific lexicons, particularly useful for specialized fields like mental health.
Dynamic Feature Evolution and Cyclical Profiling
A core innovation of ExBigBang is its dynamic, cyclical approach to feature enrichment, termed the "BigBang" process.
Unlike static models, ExBigBang utilizes a loop where initial user data informs a profile, which then guides classification. These resulting persona insights are fed back into the profile, reflecting evolving user behaviors across timestamps ($T_1$ to $T_m$). This process increases the dimensionality and richness of the features, moving from raw data to sophisticated persona-based representations.
Experimental Evaluation
The model was evaluated using the Twitter Sentiment140 dataset, filtered to include users with at least 10 tweets to ensure sufficient context. The final dataset comprised approximately 470,000 tweets from 27,000 users, with the task of inferring depressive states as a proxy for complex persona classification.
Key Results
ExBigBang significantly outperformed traditional machine learning and standard deep learning baselines, achieving an F1-score of 0.84.
| Model | Accuracy | Precision | Recall | F1-Score |
|---|---|---|---|---|
| ExBigBang (Hybrid Transformer) | 0.91 | 0.80 | 0.87 | 0.84 |
| BiLSTM-Att + Learned Emb | 0.80 | 0.80 | 0.79 | 0.79 |
| XGBoost (Text-Tabular) | 0.76 | 0.76 | 0.79 | 0.78 |
| SVM (Text-Tabular) | 0.73 | 0.72 | 0.80 | 0.75 |
The integration of SHAP (SHapley Additive exPlanations) allowed for the identification of key influential features, such as avg_top_sent (average sentiment of core vocabulary) and night_ratio (longitudinal temporal behavior), proving that aggregated context is often more reliable than individual data points.
Conclusion
ExBigBang demonstrates that a hybrid, attention-based approach to fusing text and metadata provides superior accuracy and transparency in persona classification. While computationally intensive, the framework's ability to provide explainable insights makes it a powerful tool for designers and researchers. Future work will explore multimodal data integration and the use of Knowledge Graphs to further deepen contextual understanding.
Summary updated