Figure 2: Proposed model architecture Table 3: Comparison of Sherlock and the proposed model on Twitter dataset Table 3 provides a comprehensive comparison between the per- formance of Sherlock and the proposed model in terms of the F1 score and accuracy metrics. The evaluation is conducted for Table 2: Time spent on preprocessing steps Figure 4 demonstrates the conversion of a JSON document into relational tables, where each key-value pair is represented as a sep arate column. Subsequently, a feature extraction process is applied to these values to obtain the feature vectors. This step enables us to utilize the Sherlock model for classification, allowing us to establish 1 baseline performance measure. We then proceed to process the Jata into graphs, following the procedures outlined in Figure 2. The classification task is then performed using a graph neural network model. We summarize the time spent on each preprocessing step in Table 2, providing an overview of the time requirements for these tasks on our two different datasets. The key-value pair extraction step takes 1,205s and 3,525s for the Twitter and Meetup datasets respectively, which is the most time-consuming preprocessing step The feature extraction and graph processing steps take less time than the key-value pair extraction. The feature extraction step is the only preprocessing step required to train the Sherlock model. Table 4: Comparison of Sherlock and the proposed model on the Meetup dataset Table 5: Training time and model size comparison Figure 4: transformation of JSON document to relational table
