Forecasting Unplanned Purchase Behavior under Buy-One Get-One-Free Promotions Using Functional Near-Infrared Spectroscopy

Figures

• 

  Figure 1. Noninvasive mapping of brain function using neuroimaging technologies. EEG records voltage fluctuations caused by electrical currents flowing through the brain because of active neurons using an array of electrodes placed on the scalp. fMRI measures hemodynamic responses associated with brain activity by relying primarily on the local blood-oxygen-level-dependent (BOLD) signal, which detects changes in blood oxygenation caused by neural activity. fNIRS measures the changes in oxygenation and deoxygenation hemoglobin concentrations in the brain using near-infrared light.

• 

  Figure 2. Typical scheme of a BCI system consisting of five main stages. The process is as follows: (1) cognitive task; (2) fNIRS acquisition; (3) feature extraction methods; (4) pattern classification; (5) response monitoring. We went through this whole process to detect UPB.

• 

  Figure 3. Overall procedures of performing Tasks 1 and 2. Participants are free to choose what clothes they want to buy from online shopping malls, which are divided into Task 1 with BOGOF conditions and Task 2 without BOGOF conditions. The whole experiment consists of trial number mark, experimental task, and rest. Each task is conducted for 25 s followed by a rest for 30 s. A total of five trials were conducted. The tasks are randomized and counterbalanced in sequence. Before and after the experiment, a self-reported measurement related to UPB was conducted using the popular survey platform, Qualtrics.

• 

  Figure 4. fNIRS channel configuration (a) and source-detector pair (b). In the left panel, the grey circles indicate the sources, whereas the orange circles represent the detectors, resulting in a total of 15 fNIRS channels in the prefrontal cortex (PFC). According to the 10–20 international system, the probe sets are symmetrically placed at FPz between Chs. 7 and 10. On the right panel, the source-detector pair measures lights from a diffuse volume of tissue beneath the pair as shown in the model of light propagation. These lights can reach approximately 8 mm into the brain cortex at a source-detector spacing of 3 cm. Lights at two wavelengths (780 nm and 850 nm) are used to reconstruct changes in oxy- and deoxy-hemoglobin concentrations from the modified Beer–Lambert law. A detector captures the lights resulting from the interaction with HbO and HbR, following a crescent-shaped path back to the surface of the skin. The crescent-shaped paths depict the traveling area of the near-infrared light (NIR) photons, while the blue dotted arrows indicate the light scattering. The red-colored arrows show the extra distance traveled by photons, which is corrected by the differential path length…

• 

  Figure 5. Concept of support vector machine (SVM). The green squares and the pink circles are support vectors. SVM should find the optimal hyperplane (solid black line) divided into Class 1 and Class 2 with a maximum margin (2l). Hence, SVM estimates the hyperplane in the two-dimensional space and classifies two classes.

• 

  Figure 6. Typical ROC curves and their AUCs. The ROC curves show the relationship between true-positive rate (TPR) and false-positive rate (FPR) to validate the reliability of results calculated by the SVM classifier. TPR is the ratio of correctly judging the UPB elicited by BOGOF as the UPB. In contrast, FPR is the ratio of incorrectly judging the non-UPB as UPB. Through these TPRs and FPRs, the AUC can be calculated by quantifying the entire 2-D region under the ROC curve. The red dotted diagonal line (AUC = 0.5) depicts a baseline. The AUC of the purple line (AUC = 1.0) is considered to be the best reliability in the tested classifier results.

• 

  Figure 7. Topographical maps of averaged HbO activities under (a) Task 1 and (b) Task 2. Most areas show little activation except for the left VLPFC in (a). On the other hand, extensive brain activations appear in mPFC, VLPFC, and OFC in (b). These regions have the function of inhibiting UPB as an important predictor of impulsiveness. Thus, these findings provide empirical evidence that the BOGOF condition sufficiently encourages UPB to differentiate it from non-UPB.

• 

  Figure 8. Classification results between UPB and non-UPB using SVM for each subject during cognitive tasks in accordance with the BOGOF condition. The red error bars refer to the standard error of mean (SEM). “A” in the x-axis indicates the average classification accuracy across 33 subjects with average accuracy and SEM (94.23%±0.73). The accuracies of all subjects are ranging from 86.42 to 99.90%, suggesting that the fNIRS data could be used as a biomarker to differentiate between UPB and non-UPB.

• 

  Figure 9. ROC curves and their AUCs of two representative subjects with (a) the lowest accuracy and (b) the highest accuracy. The red dotted diagonals represent the baseline. The reliability of the classification results is not guaranteed if the SVM curves (blue lines) locate below the baseline, but the classification results are reliable if the SVM curves locate above the baseline. Therefore, all SVM results are trustworthy, and (b) is more reliable than (a).

• 

  Figure 10. Screenshots of the detection results for UPB patterns. If the classification accuracy is above 80%, a message indicates that you are ready to detect unplanned purchase patterns (a); otherwise, indicates that you are not able to detect unplanned purchase patterns (b). In this experiment, all subjects received a message stating that this system can detect UPB patterns because each subject achieved a classification accuracy of higher than 86%.