MIDiff Generates Realistic Mobile Usage Data Despite Sparsity

Mobile usage data is invaluable for tasks like user behavior prediction and app recommendations, but its utility is often hampered by privacy concerns and the high cost of data collection. Furthermore, generative models struggle with mobile usage traces due to inherent sparsity from limited user activity, the complexity of jointly modeling heterogeneous variable types, and significant usage imbalance across different applications. To overcome these obstacles, this paper introduces Multivariate-Imaging Diffusion (MIDiff). This diffusion-based framework operates in an innovative "imaging space" created by the Cross-Gramian Angular Sum Field (C-GASF), which transforms sparse multivariate sequences into correlation images. MIDiff then employs a U-Net architecture with Triple Attention to maintain temporal consistency and preserve variable dependencies within these images. Experimental results demonstrate that MIDiff achieves state-of-the-art performance across various fidelity metrics. Notably, it significantly outperforms strong baselines like ZITS-VAE in Discriminative Accuracy, indicating its superior ability to generate diverse and realistic mobile usage traces. The availability of the code further supports its practical application and reproducibility.