Detection of Short-Form Video Addiction with Wearable Sensors Via Temporally-Coherent Domain Adaptation

Abstract

Short-form Video Addiction (SVA), a novel digital addiction of the modern world, proliferates among young adults and is not formally diagnosable. SVA detection from resulting bio-signals is crucial to prevent its adverse impacts. Existing formal methods involve large and expensive neuro-imaging devices in laboratory setups that are intrusive and not feasible to use in daily life. A possible non-intrusive solution can be using wearable sensors which is challenging due to the resulting noisy and faint signals. To address this problem, we investigated multi-modal wearable sensing technology to detect SVA in a non-intrusive fashion. However, fusing multi-modal sensors effectively presented different challenges due to the presence of signal heterogeneity. In this study, we proposed a novel multi-modal temporally coherent domain adaptation method to effectively detect SVA using Electroencephalogram (EEG) and Electrodermal Activity (EDA) sensors. We also investigated the nature and properties of SVA with the help of different components of EEG and EDA signals. We evaluated our proposed method for SVA detection and fatigue assessment tasks. Experimental evaluation shown the proposed model's superior performance (10% accuracy) over state-of-art domain adaptation models.

Figure 1:

Overall system architecture of our proposed framework. Firstly, we selected the concurrent moving frames simultaneously from the source and the target domain. Then the LSTM encoder module encoded the time domain segment into domain invariant feature space using MMD loss. Finally, the classifier modules predicted the activity labels.

Figure 2:

LSTM Encoder topology. The input time series sequence is fed to the two layers of the LSTM network. Finally, the last LSTM layer gets fully connected to the encoded feature layer.

Figure 3:

Pre-training the source and the target encoders. We pretrained both of the encoders as a part of the corresponding autoencoders. The objective function of the autoencoders was to regenerate the same sensor data. MMD loss tried to minimize the distance between the bottleneck feature spaces during the pretraining.

Figure 4:

The sensor setup used for data collection. a. Muse S sensor, b. Empatica E4 sensor

Figure 5:

Comparison of AUC score with the baseline methods.

Figure 6:

Effect of different EDA components on the performance of our detection framework on different subjects. Error bars show the one standard deviation from the average accuracy score of 10 runs.

Figure 7:

Effect of different EEG frequency bands on the performance of our detection framework.

Figure 8:

Parameter Sensitivity Analysis on Beta Parameter