In the following, we describe how to use the scripts in this folder. We assume you have downloaded our dataset from UCI Machine Learning repository (link should be provided soon).

In Figure 7 we showed two examples of recordings of wine and banana stimuli, one of each. Script Plot_Figure7.py reproduces this plot using matplotlib. The inductions used in the paper were recorded on July 23rd, and are indexed in the dataset as 17 and 19. It will generate a PNG file file in the same folder. Just run:

python Plot_Figure7.py

It will read the dataset (which should be in the same folder as this script), get presentations with id 17 and 19 and plot them.

Below is the result:

Because the main topic of the paper is online analyses, one possible methodology is to split the dataset in chunks of equal size. In the paper, we divided each induction in time series of 10min of length. The length of the chunks is determined by variable deltaT (in hours).

The script to perform this operation is called Features_split.py, and should be located in this directory. Assuming that the dataset is also in this folder, simply run the following

python Features_split.py

It takes about 20 seconds. Once it is finished, it will save two files: "Dataset_Split10min.npy" with the data reformated for supervised training, and "Dataset_SplitHash.npy" with the indices for each induction. The hash will be used during the cross-validation to ensure fair training (see description in our paper). This routine was optimized by only using numpy functions.

There are a few important considerations to be taken into account. Many inductions have issues with missing data points. This is, actually, part of what makes this dataset interesting. However, it also requires special treatments. In this example script, we have performed linear interpolation to fill in missing points, except when the missing points happened at the beggining of the induction. In this case we simply removed the first few windows from the final dataset.

In the original papel we have used the Inhibitory Support Vector Machine (click here for an implementation), we will use in this example a regular Support Vector Machine despite not being Fisher consistent for multiclass problems (see this paper). The only extra dependence needed for this section is scikit-learn.