SensorPersona: An LLM-Empowered System for Continual Persona Extraction from Longitudinal Mobile Sensor Streams
Summary: arXiv:2604.06204v1 Announce Type: cross
In an age where personalization is critical for enhancing user experience in technology, a new frontier has been opened with the introduction of SensorPersona. This innovative system leverages Large Language Model (LLM) capabilities to continuously extract user personas from extensive sensor data collected from mobile devices.
Introduction
Personalization plays an essential role in the functionality of LLM-based agents, enabling them to adapt to individual user preferences and significantly improve the quality of their responses and task performance. However, traditional methods of persona inference have relied heavily on chat histories. This approach captures only the self-disclosed information from users, which limits the understanding of their comprehensive personas as it overlooks their everyday behaviors in the physical world.
The SensorPersona System
To address these limitations, the SensorPersona system has been developed. It operates on the premise that continuous, unobtrusive data from mobile devices can provide a richer, more nuanced understanding of user personas. The system utilizes a three-step approach:
- Context Encoding: SensorPersona begins with person-oriented context encoding. This process enriches the semantics of sensor contexts by interpreting continuous sensor streams.
- Hierarchical Persona Reasoning: The system employs hierarchical reasoning that integrates both intra- and inter-episode reasoning. This allows for the extraction of personas that encompass physical patterns, psychosocial traits, and life experiences.
- Incremental Verification and Updating: Finally, SensorPersona uses clustering-aware incremental verification alongside temporal evidence-aware updating, which helps in adapting to evolving user personas over time.
Evaluation and Results
To validate the effectiveness of SensorPersona, the researchers conducted extensive evaluations on a self-collected dataset. This dataset comprised 1,580 hours of sensor data gathered from 20 participants over a span of up to three months, across 17 cities on three continents. The results were compelling:
- SensorPersona achieved up to a 31.4% increase in recall for persona extraction compared to existing methods.
- The system demonstrated an 85.7% win rate in persona-aware agent responses.
- User satisfaction levels showed notable improvements when compared to state-of-the-art baselines.
Conclusion
SensorPersona represents a significant advancement in the field of personalized AI interactions. By utilizing longitudinal sensor streams to derive a deeper understanding of user personas, it sets a new standard for LLM-based agents. As technology continues to evolve, approaches like SensorPersona will be pivotal in enhancing user experience and satisfaction in AI-driven applications.