Window-length sweep¶
The window-length sweep is the driver behind paper Figure 4 and the HLE dimension of Figure 2.
Run¶
Two-phase ordering¶
The sweep runs ISOLATED methods first, then GRAPHLET methods. The ordering is load-bearing — graphlet methods read the neighbour predictions dumped by the isolated phase.
| Phase | Methods |
|---|---|
| ISOLATED | lstm, transformer × {learnable, sinusoidal, rope} |
| GRAPHLET | graphlet, transformer_graphlet |
Outputs¶
swissrivernetwork/benchmark/dump/predictions/<path_extra_keys>-evalwl{W}/— onewt_hat_*.csvper station and method.swissrivernetwork/benchmark/visualize_results/outputs/win_lens/<graph>_<method>_win_lens_resu.csv— aggregated RMSE / MAE / NSE perwindow_len.
Rendering Figure 4¶
The notebook reads the CSVs above and emits the final PDF under
visualize_results/figures/.
Eval-window leakage
If you are reproducing numbers from a pre-4daeff3 run, discard
everything at W ≠ 90 for graphlet methods and re-run — see
the explainer.