3.1.2. TCDF

With regard to TCDF (Figure 9), only two mean attention scores were significant, and solely in the Far–Superficial dipole condition and for two different targets (X2 and X3), such that a target-wise comparison cannot be made. More concretely, a connection X3 → X2 was found (in accordance with the ground truth), as well as a connection X2 → X3 (unlike the ground truth).

**Figure 9.** TCDF, Mean attention score rankings. Colors: "1", green, denoting the highest attention score in one TCN—one column). The columns represent the targets (T X1, T X2, T X3), the rows the predictors (X1, X2, X3). Self-connectivity is excluded.

#### 3.1.3. LSTM-NUE

Next, when focusing on LSTM-NUE, as can be seen from the colors from Figure 10, for target time series X3, the GC scores (in both dipole conditions) were higher than expected according to the ground truth. Unexpected GC scores are surrounded by black rectangles in the top panel. Column-wise strength rankings (rankings for one particular target) are correct for two out of three targets (X1, X2) in both conditions, as can be seen by comparing with column-wise Ground Truth 1 (Figure 10, bottom right panel). The overall ranking in the Far–Deep condition was more in accordance with the overall ranking in Ground Truth 1 (Figure 10, bottom left panel) than the ranking found in the Far–Superficial condition because connectivity strength was observed to be the weakest for the corresponding false positives (as shown in yellow in the top panel).

**Figure 10.** LSTM-NUE, Neural-Network Granger Scores rankings (Top) versus Truth 1 (Bottom), excluding self-connectivity. Color coding: dark green > light green > yellow > orange > red. The columns represent the targets, the rows the time series used for prediction.
