3.4.1 Pre-Processing & Feature Extraction
We shall discuss how we used S-Golay filter to even out the signals and then DWT based wave-let analysis to extract features from Neuro-signals.
3.4.1.1 Savitzky–Golay Filter
In layman words [24] if we try to understand this filter, it is a polynomial based filter in which least square polynomial method find out the filtered signals with combinedly evaluating the neighbor signals. It can be computed for a signal such as Sj = f(tj), (j = 1, 2, …, n) by following Equation (3.3).
Here, ‘m’ is frame span, ci is no. of convolution coeffs. and Q is the resultant signal. ‘m’ is used to calculate instances of ci with a polynomial. In our case, this was used to smoothen Neuro-signals by a frame size of 5 with a polynomial of degree 4.
3.4.1.2 Discrete Wavelet Transform
In layman words, it is used to convert incoming signal into sequences of smaller waves using multi-stage decomposition. This enables us to analyze multiple oscillatory signals in an approximation and detail coefficient form. Figure 3.3 has shown the decomposing structural of brain neuron signals when it is preprocessed using low & high filtering methods. Low filter pass method (L) removes high voltage fluctuations and saves slow trends. These in turn provide approximation (A) of the signal. High pass filter method (H) eliminates the slow trends and saves the high voltage fluctuations. The resultant of (H) provides us with detail coefficient (D) which is also known as Wavelet coefficient. The Wavelet function is shown in Equations (3.4) and (3.5).
Here ‘a’ and ‘b’ are scaling parameter and translation parameter containing discrete values. ‘m’ is frequency and ‘n’ is time belonging to Z. The computation of (A) and (D) is shown in scaling function (3.6) and wavelet function (3.7).
Figure 3.3 DWT schematic.
Here, φj,k(n) is the scaling function belonging to (L) and ωj,k(n) is the wavelet function belonging to (H), M is length of signal, n is the discrete variable lies between 0 and M − 1, J = log2(M), with k and j taking values from {0 – J − 1}. The values of Ai and Di are computed below by Equations (3.8) and (3.9).
In previous works we have seen that theta (4–8 Hz) is preferably explored for finding judgement tasks, studying the cortical activity in left side of brain. We used 4-levels of signal decomposition by Daubechies 4 wavelet technique which results into a group of 5 wavelets coeffs where one group represent one oscillatory signal and presents Neuro-signal pattern through D1–D4 and A4. They have “5 frequency bands—(1–4 Hz), (4–8 Hz), (8–13 Hz), (13–22 Hz) and (32–100 Hz)”.
3.4.2 Dataset Description
We recorded Neuro-signals from 25 participants through 14 channels who all were aged between 18 and 38 years of age belonging to IIT Roorkee. They were shown 39 (13 product types × 3 samples of each product) images. A total of 325 (13 × 25) Neuro-signals were recorded wherein each image was displayed for 4 s. After the collection we asked users to label their choice for each item in terms of Like/Dislike. All participants were instructed to truly evaluate and label correct choices. The following Figures 3.4 and 3.5 shall provide an overview of the image and dataset.
3.5 Result
3.5.1 Individual Result Analysis
Here, we display the output of the 5 classification algorithms we used on the given dataset for deliberating on Neuromarketing using Machine Learning Algorithms. The output was user-independent as we assume that for predicting the choice of a user his/her training data is not required. In Figure 3.6 we depict the overall accuracy received while running it on the mentioned columns, i.e., Decision Tree, Bagging Decision Tree, Gaussian Naïve Bayes Classifier, Kernel SVM and Random Forest Classifier.
Figure 3.4 Images used for visual evaluation.
Figure 3.5 Sample of EEG signal for a product with corresponding Brain map and choice label.
We compiled all participant’s EEG data into a single file called as “Master file” with appropriate like=1/dislike=0 labelling for all rows. We observe here that Kernel SVM has the highest achieved accuracy followed by Decision Tree whereas all other 3 produce near about close results.
In Figure 3.7, we shall be showing the individual ROC AUC curve results
