Thursday 27 February 2025
Generative models have come a long way in recent years, able to create realistic images and videos that can fool even the most discerning human eye. But what about time series data? That’s where things get tricky. Unlike still images or video frames, time series data is all about patterns and trends over time – and generating new data that follows those patterns is a much harder problem.
Enter PaD-TS, a new model from researchers that tackles this challenge head-on. By incorporating a population-aware training process and a novel dual-channel encoder architecture, PaD-TS is able to generate realistic time series data that not only looks good but also preserves the underlying statistical properties of the original dataset.
The problem with traditional generative models is that they often focus too much on individual samples, rather than the bigger picture. They might be able to create a single fake image or video that looks convincing, but when it comes to generating large datasets, things start to fall apart. PaD-TS takes a different approach by incorporating a population-aware training process that encourages the model to learn about the overall distribution of the data – not just individual samples.
This is achieved through a clever combination of two losses: one that focuses on the individual samples and another that looks at the entire dataset as a whole. By balancing these two losses, PaD-TS is able to generate data that not only looks realistic but also preserves the underlying statistical properties of the original dataset.
But how does it work in practice? The researchers tested PaD-TS on three different datasets: Sines, Stocks, and Energy. Each dataset has its own unique characteristics – for example, the Energy dataset has a strong seasonal component that’s easy to spot. But when generated using PaD-TS, the resulting data looks surprisingly realistic.
One of the key challenges in generating time series data is preserving the cross-correlation (CC) between different dimensions. In other words, if two variables are strongly correlated in the original data, the generated data should also reflect that correlation. PaD-TS does an excellent job of this, with results showing a strong alignment between the CC values of the original and synthetic data.
The researchers also tested PaD-TS on longer sequences, using datasets with generation lengths of up to 256 time steps. This is a significant challenge for any generative model, but PaD-TS rose to the task admirably.
Cite this article: “PaD-TS: A Novel Generative Model for Realistic Time Series Data”, The Science Archive, 2025.
Generative Models, Time Series Data, Pad-Ts, Population-Aware Training Process, Dual-Channel Encoder Architecture, Statistical Properties, Individual Samples, Dataset As A Whole, Cross-Correlation, Sequence Generation
