[1B3-GS-2] Machine learning

[1B3-GS-2-03] Zero-Inflated Poisson Transformer model for Count Time-Series Data

〇Daichi Kimura¹, Tomonori Izumitani¹ (1. NTT Communications)

Keywords:Time series, Count data, Transformer, Zero-Inflated

Long sequence time-series forecasting for counting quantities such as demand, sales, and transactions in stock market is important for various business areas. These kinds of real-world data have properties: such as time dependency, non-linearity, non-Gaussian distribution, zero-inflated and integer values. In this study, we propose a time-series forecasting model for zero-inflated count data. To consider time dependency and obtain long-term outputs, we utilize the Informer which is a long sequence time-series forecasting method based on the Transformer. In addition, we suppose a Poisson distribution and a Bernoulli distribution for the outputs of Informer models to deal with zero-inflated count data properties. We evaluated the method using two artificial and two real-world datasets. The results show that the proposed method can make precise forecasts with long-term adaptation to various trend lines. In particular, the proposed method showed highest prediction accuracy in five of the six experimental conditions using real datasets.