EKC2023

Price prediction of agricultural products using Spectral Temporal Graph Neural Network

Aug 16, 2023, 4:05 PM

15m

Orion 2

Basic Science [BS2] Recent developments of mathematics in solving industrial problems and its applications

Speaker

Youngho Min (Hankuk University of Foreign Studies)

Description

Agricultural products are an important part of human life and account for a significant proportion of the market economy. Agricultural prices have a direct impact on the livelihoods of suppliers and the household economies of consumers. Therefore, governments around the world are making great efforts to stabilize agricultural prices. The price of agricultural products is determined by the supply and demand of the respective year [1]. Many researchers have conducted various studies on supply and demand to predict agricultural product prices [2,3]. However, agricultural production is heavily dependent on various variables and has strong nonlinear characteristics. In particular, the uncertainty of the weather makes agricultural suppliers uneasy and has a negative impact on the formation of stable agricultural prices. Previous studies have used deep learning methods of the LSTM series, which show excellent performance in predicting the time series data, to predict agricultural prices [4]. We propose using the Spectral Temporal Graph Neural Network (StemGNN) [5], a multivariate time series data prediction method that reflects the correlation between weather and different agricultural products in the price prediction process, for predicting the prices of agricultural products. One of the characteristics of StemGNN is that it automatically learns the correlation between time-series data using a Self-Attention mechanism, without using predefined correlations between data. This allows for better incorporation of weather characteristics (including 15 features such as temperature and precipitation) into the price prediction of each agricultural product compared to previous methods. In addition, StemGNN is composed of a combination of Graph Fourier Transform (GFT), which models the correlation between time-series data, and Discrete Fourier Transform (DFT), which models temporal dependency, making it appropriate for incorporating seasonality of each agricultural product into price prediction. The proposed StemGNN model is applied to four agricultural products (potatoes, lettuce, onions, and cucumbers), and its performance is compared with benchmark models using LSTM and attention-based LSTM. We also find the optimal input period for each agricultural product when using StemGNN.

References

[1] Rao, J. M. (1989). Agricultural supply response: A survey. Agricultural economics, 3(1), 1-22.
[2] Weisong, M., Xiaoshuan, Z., Lingxian, Z., & Zettan, F. (2007). A structural model for analysis of fruit supply and demand applied to grapes in China. New Zealand Journal of Agricultural Research, 50(5), 1359-1365.
[3] Czyżewski, A., Bieniek-Majka, M., & Czakowski, D. (2018). Factors shaping supplydemand relations on the fruit and vegetable market in the light of the behavior of groups and producer organizations. Management, 22(1), 265-277.
[4] PARK, T. S., KEUM, J., KIM, H., KIM, Y. R., & MIN, Y. (2022). PREDICTING KOREAN FRUIT PRICES USING LSTM ALGORITHM. Journal of the Korean Society for Industrial and Applied Mathematics, 26(1), 23-48.
[5] Cao, D., Wang, Y., Duan, J., Zhang, C., Zhu, X., Huang, C., ... & Zhang, Q. (2020). Spectral temporal graph neural network for multivariate time-series forecasting. Advances in neural information processing systems, 33, 17766-17778.