Interpretable trend-seasonality pattern of transformer in time series forecasting

MPhil Thesis Defence


Title: "Interpretable trend-seasonality pattern of transformer in time 
series forecasting"

By

Mr. Yaowei HUANG


Abstract

Positional and temporal information is important in time series 
forecasting problems. Based on the Transformer architecture, the 
positional and temporal information is encoded and then concatenated as 
one vector input in the self-attention module. In this work, we explore 
the problems in the previous structure and propose a trend-seasonality 
pattern transformer that is interpretable and visualization. Different 
from previous works, our model uses positional embedding as the trend 
pattern and temporal embedding as the seasonality patterns, which two are 
computed separately with different self-attention modules and added 
afterward. This design helps remove the addition of heterogeneous vectors 
over different information, which may bring noise. Moreover, our model 
could also visualize the trend and seasonality pattern which is very 
practical in real-world applications. Our experiment on the real-world 
datasets shows that it outperforms the state-of-the-art and provides 
outputs that are interpretable.


Date:  			Thursday, 12 August 2021

Time:			9:00am - 11:00am

Zoom meeting:		https://hkust.zoom.com.cn/j/9125182610

Committee Members:	Prof. Tong Zhang (Supervisor)
 			Prof. Xiaofang Zhou (Chairperson)
 			Prof. James Kwok


**** ALL are Welcome ****