Figure 1 Discretization of the NYC study area into grid cells More about this image found in Discretization of the NYC study area into grid cells

Figure 2 Schematic workflow of the candidate stop selection process using convolutional sliding windows More about this image found in Schematic workflow of the candidate stop selection process using convolutio...

Figure 3 Illustration of the GNN action execution process More about this image found in Illustration of the GNN action execution process

Figure 4 Q-Network architecture More about this image found in Q-Network architecture

Figure 5 Sensitivity analysis of the number of stops based on PDRs and ECA More about this image found in Sensitivity analysis of the number of stops based on PDRs and ECA

Figure 6 Sensitivity analysis of convolution kernel size k and minimum inter-stop distance d_m More about this image found in Sensitivity analysis of convolution kernel size k and mini...

Figure 7 Performance comparison of the proposed CWMSS method against baselines across three key metrics More about this image found in Performance comparison of the proposed CWMSS method against baselines acros...

Figure 8 Training convergence curves of the DQN-GNN model across three optimization scenarios More about this image found in Training convergence curves of the DQN-GNN model across three optimization ...

Figure 9 Urban snapshots of the NYC transit environment More about this image found in Urban snapshots of the NYC transit environment

Figure 10 Impact of sampling size N on cost function across optimization scenarios More about this image found in Impact of sampling size N on cost function across optimiza...

Figure 11 Performance comparison of Best-of-N and Greedy strategies across three optimization scenarios More about this image found in Performance comparison of Best-of-N and Greedy strategies across three opti...

Figure 1 Overview of the experimental workflow More about this image found in Overview of the experimental workflow

Figure 2 Token-level contribution distributions of LIME, SHAP and BreakDown for negative (left) and non-negative (right) sentiment predictions. Top-5 influential tokens are shown for each method More about this image found in Token-level contribution distributions of LIME, SHAP and BreakDown for nega...

Figure 3 Comparison of mean absolute token contributions produced by LIME, SHAP and BreakDown across sentiment categories. Error bars represent 95% confidence intervals. (a) Mean absolute contribution (±95% CI) across explainers More about this image found in Comparison of mean absolute token contributions produced by LIME, SHAP and ...

Figure 4 Comparison of contribution magnitude distributions produced by LIME, SHAP and BreakDown for RQ1 More about this image found in Comparison of contribution magnitude distributions produced by LIME, SHAP a...

Figure 5 Consistency of Top-10 influential tokens across explanation methods measured by ranked biased overlap (RBO) More about this image found in Consistency of Top-10 influential tokens across explanation methods measure...

Figure 6 Direction conflicts among LIME, SHAP and BreakDown explanations for overlapping Top-10 tokens More about this image found in Direction conflicts among LIME, SHAP and BreakDown explanations for overlap...