KANB-Enhanced neural networks for mobility forecasting
Νευρωνικά δίκτυα ενισχυμένα με KANB για πρόβλεψη κινητικότητας

Bachelor Dissertation
Author
Kontoulis, Markos
Κοντούλης, Μάρκος
Date
2026-09View/ Open
Keywords
Vessel traffic forecasting ; AIS ; Transformer ; Adaptive activation functions ; KANBAbstract
This thesis examines whether trainable KANB activation functions improve short-term mobility
forecasting when they replace the fixed activation in a Transformer feed-forward block. Ten
Aegean and Brest grid-cell series are forecast at six horizons from 5 to 30 minutes, as well as
the PEMS04 road-traffic benchmark. Hyperparameters were tuned sequentially on the
validation set with KANB21 and five development series; the test split was not used during
tuning.
The selected configuration for the maritime data uses Adam with a learning rate of 3 × 10− 4 and
ε = 10− 5, batch size 256, 150 epochs, two Transformer blocks and four attention heads, without
early stopping. The final comparison covers KANB17, KANB18, KANB21, ReLU, GELU and
ParametricGELU over five random seeds, giving 300 successful runs. GELU has the lowest
mean validation relative MAE (0.8037). On the test set, KANB17 has the lowest macro-average
relative MAE (0.8498), while ReLU ranks first on five of the ten individual series. The KANB21–
GELU test difference is only −0.0012 and is not statistically supported (exact p = 0.912). All nine
KANB-versus-baseline confidence intervals include zero. The neural model beats persistence
on eight series, but Brest 74768 and 75108 remain persistence-dominated. KANB is therefore a
competitive low-parameter alternative in this setting, not a consistent improvement over
standard activations. A validation-only follow-up with KANB17 tested batch size 512,
ReduceLROnPlateau and a one-time learning-rate restart. None improved on the selected
batch-256 constant-learning-rate protocol.
For the PEMS04 dataset, 71 features were used after being derived from observations,
neighboring sensors, daily and weekly lags and calendar variables. The selection of the five
sensors was done using training and validation data only. The test set wasn’t used until the final
comparison. In the final test, which contained 150 runs with early stopping and best-weight
restoration, KANB18 had the lowest relative MAE (0.747688) against ReLU (0.749232) and
GELU (0.750365). Against GELU, KANB18 performed better on all five sensors, 16 of 25 paired
sensor-seed runs and five of the six forecast horizons. The aggregate improvement was
0.357%. The exact two-sided sensor-level test gave p=0.0625, and the confidence interval
included zero.


