cwmeijer
commented
1 year ago With some hacking, I got the following image out:
The line is the raw data which is a single channel, max daily temp, from the weather dataset. The time series is classified as 'summer' as opposed to 'winter' (I excluded all autumn and spring data). The colors show the values towards the class summer. The values seem opposite of what we would expect. So, not bad. Not good either, but at least a clear pattern.
Here is the one for winter:
Again, good pattern, but opposite of what we expect.
Note that the binary model that I trained did not yield any sensible results and no useful patterns in the saliency maps. This is, I think, because the model, with 28 inputs as time series length, was trained on only 45 instances. In my experiments, the trained model didn't change predictions so easily by masking parts. I can try again later with other masking stategies or with a model that was trained on more data.
The current masking strategy was inputing the mean of the train data. Masking with the mean of the current instance wasn't successful as masking this was hardly changed any predictions by the model.
TL;DR Conclusion is that in the most simple case RISE for time works well, but with opposite values. I'm looking for a bug now.
Run rise and see if the results make any sense.