Hierarchical Multi-Persona Induction from User Behavioral Logs
A study introduces a structured approach to derive various evidence-based personas from user behavior logs utilizing LLMs. This technique compiles user activities into intent memories, subsequently categorizing and tagging them to create personas. The process of persona generation is framed as an optimization challenge, focusing on metrics such as cluster cohesion, alignment with evidence, and accuracy. Training of the model employs a groupwise variant of Direct Preference Optimization (DPO). Results from experiments conducted on an extensive service log and two public datasets indicate enhancements in persona coherence, grounding in evidence, and reliability, along with improved predictions for future interactions.
Key facts
- Framework aggregates user actions into intent memories
- Induces multiple evidence-grounded personas via clustering and labeling
- Formulates persona induction as optimization over quality metrics
- Uses groupwise DPO for training
- Evaluated on large-scale service log and two public datasets
- Improves coherence, evidence grounding, and trustworthiness
- Also improves future interaction prediction
- Published on arXiv under cs.AI
Entities
Institutions
- arXiv