Zero Inflated Time Series Generation

This repository contains the code and configurations for a study submitted as a scientific research article including two new proposed approach for zero inflated time series generation.

📄 Overview

Zero-Inflated Time Series (ZITS), characterized by excessive zeros relative to standard distributions, pose fundamental challenges for synthetic data generation due to the need to simultaneously preserve the temporal dependencies, the excess zero structure, and the heavy-tailed non-zero distributions. Current generative models, including variational autoencoders (VAE) and generative adversarial networks (GAN), lack the mechanisms for handling zero-inflated data and fail to generate data with similar characteristics.

This work proposes two architectures, ZITS-GAN and ZITS-VAE, designed for zero-inflated time series generation through a two-head decoder structure that separately models the zero pattern and the non-zero magnitude. The proposed architectures use dilated convolutional layers to capture long sequence dependencies and stabilize training through $\beta$-weighted ELBO for the ZITS-VAE and Wasserstein GAN with gradient penalty for ZITS-GAN. The proposed models were evaluated on two real world time series datasets with zero ratios of 59% and 74% and have demonstrated superior performance compared to state-of-the-art models including

• TimeGAN,
• TimeVAE,
• ChronoGAN,
• TransFusion,
• FIDE.

---

📊 Datasets

1. M5 forecasting Dataset

• Name: M5
• Description: provided on Kaggle, which contains daily unit sales collected by Walmart and distributed as several CSV files. The raw release comprises a wide table of bottom-level series, each row corresponds to a single stock keeping unit (SKU) at a single store, together with a separate calendar mapping that associates each recorded day index with its calendar date and auxiliary fields such as weekday, month, and special events. This publicly available data comprises 30,490 items over 1,941 days. The M5 data explicitly focuses on realistic retail demand and displays intermittency (lots of zero daily sales).
• Access: Available at 🔗 M5 Dataset Page

2. Household IoT Devices Dataset

• Name: Household IoT Devices
• Description: consists of time series data capturing the daily operational duration of household IoT devices characterized by discrete operating cycles over one year. To ensure statistical relevance and data quality, we retained from the raw dataset - comprising individual running cycle durations in seconds per device - only those devices exhibiting a sufficient number of running cycles.
• Access: unavailable.

---

🧠 Models Evaluated

• TimeGAN
• TimeVAE
• ChronoGAN
• TransFusion
• FIDE
• Proposed models: ZITS-GAN and ZITS-VAE

📈 Results Summary

We show here a summary of the evaluations made showing the difference in 0 ratios (Δ 0 ratio), the Long Sequence Predictive Score (LPS) and the Long Sequence Discriminative Score (LDS) obtained.

Model	Dataset	Δ 0 ratio	LPS	LDS
TimeVAE	M5	0.144	1.344	0.389
ZITS-GAN	M5	0.012	1.448	0.213
ZITS-VAE	M5	0.014	1.408	0.126
TimeVAE	Household IoT Devices	0.259	2847	0.5
ZITS-GAN	Household IoT Devices	0.038	2153	0
ZITS-VAE	Household IoT Devices	0.033	2153	0.248

---

📬 Contact

For questions, please contact: 📧 ardeleaneugenrichard@gmail.com

---