Predictive performance comparison of moving average kernel and trainable 1D convolutional kernel across four datasets

[bubblexl]

In this paper,4.4. Learnable Decomposition Generalization Analysis.How does the experiment design when testing learnable 1D conv on DLinear,do u have original code of this experiment

[Levi-Ackman]

"The original experimental code seems to be hard to find. However, I recommend that you can find a good implementation of DLinear at ‘https://github.com/thuml/Time-Series-Library/blob/main/models/DLinear.py’. Then, you just need to replace the moving average kernel used in ‘self.decomposition = series_decomp(configs.moving_avg)’ with our LD." To give a EXAMPLE: replace : self.decomposition = series_decomp(configs.moving_avg) uses: self.LD=LD(kernel_size=configs.moving_avg),

and in forward part, replace: seasonal_init, trend_init = self.decompsition(x) with: trend_init =self.LD(x) seasonal_init=x-trend_init

[bubblexl]

With an experimental setup with a lookback window of 96 and a prediction window of 720, the model did not perform as described in the paper on the traffic dataset.MAE:0.4->0.39