Inquiry about the TimeSeries WindowGenerator Class

General Discussion
,
#1

Hi everyone,

I am studying the code for time series forecasting using TensorFlow, and I have encountered an issue I can’t quite understand. I’d appreciate any help with this.

When setting input_width=6, label_width=1, and shift=1, everything works as expected, as shown in Figure 1.
01

Similarly, with input_width=6, label_width=1, and shift=3, everything is fine, as shown in Figure 2.

02
02741×373 28.9 KB

However, when I set input_width=6, label_width=3, and shift=1, the labeling of the target values corresponding to the inputs seems incorrect, leading to data leakage. In this case, we are explicitly giving the information of indices 4 and 5 to the model, and the target values have already been seen by the model. Specifically, we provide the model with 6 timesteps of information (indices 0, 1, 2, 3, 4, 5) and expect it to forecast the next 3 values (indices 6, 7, 8).

03
031232×377 33.5 KB

In a worse scenario, if we set input_width=6, label_width=10, and shift=1, aiming to forecast the next 10 values, only indices 4, 5, 6 are given to the model as target labels. This is because the labeling starts from the index -3 up to the end. The correct labeling should assign indices 6, 7, …, 15 as target values.

04
041611×393 38.3 KB

Could anyone explain why this happens and how to correctly configure the WindowGenerator class to avoid this issue?

Thanks in advance!

#3

The TimeSeries WindowGenerator class is a tool used in time series analysis to create sliding window datasets for training machine learning models. It helps in preparing data by splitting it into input-output pairs, allowing models to learn patterns over time.

#4

Thanks Mark!
I know about its use cases. I think it doesn’t work correctly in all possible scenarios, especially in cases that we have assigned bigger than 1 values of label_width, as I described above.

#5

The WindowGenerator class returns expected output when provided with “reasonable” input arguments. It is not made to cover “all possible scenarios”.

#6

Exactly!
However, it is completely reasonable to expect your time-series model to be able to forecast 3 upcoming values, especially when it is being considered in the class. If it is supposed to work correctly only for a label_width of 1, why is this being considered as an attribute of the class?