Fig 1.
Three key issues to be addressed.
Fig 2.
Overview of the proposed NeighborSense framework.
Fig 3.
The NeighborSense layer with the adaptive gating module.
The dashed box illustrates how the gate zi,r is computed from neighborhood statistics and used to control relation-wise aggregation.
Table 1.
Dataset statistics.
Table 2.
Relations between users extracted from Twibot-22.
Table 3.
Parameter settings.
Table 4.
Benchmarking results of bot detection approaches.
Fig 4.
Model Scale vs. Detection accuracy.
Fig 5.
Model performance with training data proportion ranging from 10% to 70% of the whole dataset.
Fig 6.
Distribution of zi,r for different relations.
Fig 7.
Visualization of human and bot representations on Twibot-20 via t-SNE 2-D projections.