Abstract
Data sparsity is a key challenge limiting the power of AI tools across various domains. The problem is especially pronounced in domains that require active user input rather than measurements derived from automated sensors. It is a critical barrier to harnessing the full potential of AI in domains requiring active user engagement, such as self-reported mood check-ins, where capturing a continuous picture of emotional states is essential. In this context, sparse data can hinder efforts to capture the nuances of individual emotional experiences such as causes, triggers, and contributing factors. Existing methods for addressing data scarcity often rely on heuristics or large established datasets, favoring deep learning models that lack adaptability to new domains. This paper proposes a novel probabilistic framework that integrates user-centric feedback-based learning, allowing for personalized predictions despite limited data. Achieving 60% accuracy in predicting user states among 64 options (chance of 1/64), this framework effectively mitigates data sparsity. It is versatile across various applications, bridging the gap between theoretical AI research and practical deployment.
Introduction
The "holy grail" of Empathetic AI (EAI) is the ability to understand and respond to a user's emotional state in real-time. However, most AI models are data-hungry, requiring massive datasets to find patterns. In the world of mental health and mood tracking, data is notoriously sparse. Users rarely check in multiple times a day, and even when they do, the "why" behind a mood shift is often hidden.
Traditional machine learning models, such as Transformers or XGBoost, struggle with this sparsity. They often fail to account for individual differences—for instance, a rainy day might make one person feel cozy and another feel depressed. This research introduces the Mood Shift Predictor with Sparse Check-ins (MSPSC), a personalized emotion predictor designed to work with both "active" (user check-ins) and "passive" (environmental) data, even when that data is sporadic.
Methodology
The MSPSC framework is built on several key pillars designed to handle the nuances of human emotion and the limitations of sparse data.
The Emotion Grid
Instead of simple "happy/sad" labels, the model uses a 64-point grid based on the psychological concept of Core Affect. This maps emotions across two axes: Attitude (displeasure to pleasure) and Energy (low to high). This granular approach allows for a more sophisticated representation of emotional states than binary or categorical labels.
Probabilistic Modeling and Personalization
The model assumes a user's emotion is a function of their history, current environment, and inherent tendencies. It uses a similarity function to find historical moments where environmental conditions (weather, calendar, fitness data) were similar to the present.
A critical innovation is the use of Personalization Coefficients ($w^u$). The model learns which environmental factors actually affect a specific user. If weather consistently correlates with a specific user's mood shifts but their calendar does not, the model increases the weight of weather data for that user's future predictions.
Prediction Pipeline
To avoid overwhelming the system or the user (addressing Hick’s Law and the Paradox of Choice), the model uses clustering to narrow down potential predictions to the most likely candidates. The system then updates its internal weights based on a utility-based reward system, similar to Reinforcement Learning from Human Feedback (RLHF), based on how well its predictions align with actual reported emotions.
Results
The MSPSC model was tested using data from the UpBeing App, which collects active mood reports and passive side information like weather (35.4% of dataset), calendar events (10.5%), and fitness data (1.3%).
- Superior Accuracy: MSPSC achieved an average of 46.47% to 48.35% accuracy across all users. Given that the chance of picking the correct emotion out of 64 options is only ~1.5%, this represents a significant predictive capability.
- Outperforming Deep Learning: In sparse data scenarios, MSPSC significantly outperformed complex models like Transformers (which achieved ~7% accuracy) and XGBoost (~27-30%).
- Performance in Consistency: In a case study of a "heavy user" with consistent data, MSPSC reached 64.15% accuracy. Notably, the model remained robust even for inconsistent users, performing slightly better on them (48.35%) than on the general consistent population.
Discussion and Future Work
The MSPSC framework demonstrates that personalized, probabilistic models can effectively mitigate the challenges of data sparsity in emotional tracking. By treating environmental influence as a learnable parameter, the model moves beyond general-purpose predictors to a truly personalized experience.
Future work will focus on refining heuristic thresholds for "correct" predictions and exploring how real-world deployment affects user behavior. A key question remains whether seeing a predicted mood shift helps a user regulate their emotions or inadvertently creates a self-fulfilling prophecy. Additionally, the model's reliance on passive data suggests that increasing access to diverse data streams (while maintaining privacy) will further enhance predictive power.
Summary updated