Large language model as user daily behavior data generator: balancing population diversity and individual personality
Haoxin Li, Jingtao Ding, Jiahui Gong, Yong Li
TL;DR
BehaviorGen introduces an LLM-based framework to generate synthetic user behavior data that balances population diversity with individual personalization. Formally, user sequences are defined as $x_i=(d_i,t_i,l_i,b_i)$ and generation as $[\,\hat{x_1},...,\hat{x_O}] = G([x_1,...,x_I])$, enabling a privacy-preserving data augmentation/replacement pipeline. The paper evaluates three usage scenarios—pretraining augmentation, finetuning replacement, and finetuning augmentation—on Tencent and smartphone-usage datasets, reporting up to $18.9\%$ gains in target tasks and competitive performance when replacing real data. Ablation and case studies demonstrate the importance of prompts, role/format controls, and segmentation for producing high-fidelity yet diverse synthetic data that aligns with individual intent distributions. Overall, BehaviorGen offers a scalable path to improve daily-behavior modeling while mitigating privacy risks.
Abstract
Predicting human daily behavior is challenging due to the complexity of routine patterns and short-term fluctuations. While data-driven models have improved behavior prediction by leveraging empirical data from various platforms and devices, the reliance on sensitive, large-scale user data raises privacy concerns and limits data availability. Synthetic data generation has emerged as a promising solution, though existing methods are often limited to specific applications. In this work, we introduce BehaviorGen, a framework that uses large language models (LLMs) to generate high-quality synthetic behavior data. By simulating user behavior based on profiles and real events, BehaviorGen supports data augmentation and replacement in behavior prediction models. We evaluate its performance in scenarios such as pertaining augmentation, fine-tuning replacement, and fine-tuning augmentation, achieving significant improvements in human mobility and smartphone usage predictions, with gains of up to 18.9%. Our results demonstrate the potential of BehaviorGen to enhance user behavior modeling through flexible and privacy-preserving synthetic data generation.