Skip to content

Clean-up rel-bench example #1326

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
rusty1s merged 4 commits into from
Nov 2, 2026
Merged

Clean-up rel-bench example #1326

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3ffec20
update
rusty1s Nov 2, 2026
63bda1f
Merge branch 'main' into cleanup_relbench_example
rusty1s Nov 2, 2026
6fd72d5
update
rusty1s Nov 2, 2026
2d2c7ca
Merge branch 'cleanup_relbench_example' of github.com:NVIDIA/structur…
rusty1s Nov 2, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
62 changes: 31 additions & 31 deletions examples/relational/relbench/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,41 +12,41 @@ python main.py --dataset=rel-f1 --task=driver-dnf

- `--num_estimators`: Number of estimators (default: `1`).
- `--context_size`: Maximum number of context rows (default: `10_000`).
- `--context_mode`: Context selection strategy, either `latest` or `random` (default: `latest`).
- `--context_mode`: Context selection strategy, either `last` or `rnd` (default: `last`).
- `--num_neighbors`: The number of neighbors to sample for each hop (default: `[16, 16]`).
- `--text`: Include text embeddings from the task entity table.
- `--text_dim`: Final text embedding dimension after PCA (default: `64`).
- `--num_lags`: The number of lag targets to use as additional input (default: `20`).
- `--text_dim`: Text embedding dimension after PCA for the entity table (default: `0`).

## Results

Optimal hyperparameters found via the [RelArena-α](../../../benchmark/relational/relarena) evaluation protocol, using 8 estimators:

| Dataset | Task | Val AUROC | Test AUROC | Size | Mode | Neighbors | Text Model | Text Dim |
| :----------- | :---------------- | --------: | ---------: | ----: | :------- | ---------: | :---------- | -------: |
| `rel-amazon` | `user-churn` | 0.6935 | 0.6947 | `20K` | `latest` | `[8, 8]` | | |
| `rel-amazon` | `item-churn` | 0.8215 | 0.8254 | `20K` | `latest` | `[32, 32]` | | |
| `rel-avito` | `user-visits` | 0.6956 | 0.6711 | `20K` | `latest` | `[32, 32]` | | |
| `rel-avito` | `user-clicks` | 0.6369 | 0.6824 | `20K` | `latest` | `[32, 32]` | | |
| `rel-event` | `user-repeat` | 0.7330 | 0.8090 | `1K` | `random` | `[4, 4]` | | |
| `rel-event` | `user-ignore` | 0.8467 | 0.8380 | `2K` | `latest` | `[16, 16]` | | |
| `rel-f1` | `driver-dnf` | 0.7937 | 0.8418 | `2K` | `latest` | `[4, 4]` | | |
| `rel-f1` | `driver-top3` | 0.8816 | 0.9077 | `1K` | `random` | `[4, 4]` | | |
| `rel-hm` | `user-churn` | 0.7036 | 0.6992 | `20K` | `latest` | `[1, 1]` | | |
| `rel-stack` | `user-engagement` | 0.9001 | 0.9019 | `20K` | `random` | `[8, 8]` | `mini`[^1] | `32` |
| `rel-stack` | `user-badge` | 0.8932 | 0.8809 | `20K` | `random` | `[8, 8]` | `mpnet`[^2] | `64` |
| `rel-trial` | `study-outcome` | 0.6636 | 0.7393 | `20K` | `latest` | `[8, 8]` | `mini`[^1] | `32` |

| Dataset | Task | Val nMAE | Test nMAE | Size | Mode | Neighbors | Text Model | Text Dim |
| :----------- | :---------------- | -------: | --------: | ----: | :------- | ---------: | :--------- | -------: |
| `rel-amazon` | `user-ltv` | 21.60 | 25.01 | `20K` | `random` | `[64, 64]` | | |
| `rel-amazon` | `item-ltv` | 7.38 | 7.56 | `20K` | `random` | `[]` | | |
| `rel-avito` | `ad-ctr` | 32.40 | 32.95 | `2K` | `latest` | `[64, 64]` | | |
| `rel-event` | `user-attendance` | 31.25 | 31.70 | `20K` | `random` | `[64, 64]` | | |
| `rel-f1` | `driver-position` | 44.63 | 43.55 | `20K` | `latest` | `[8, 8]` | | |
| `rel-hm` | `item-sales` | 9.81 | 8.00 | `20K` | `random` | `[]` | | |
| `rel-stack` | `post-votes` | 11.36 | 12.49 | `20K` | `random` | `[]` | | |
| `rel-trial` | `study-adverse` | 10.21 | 9.42 | `20K` | `latest` | `[]` | `mini`[^1] | `32` |
| `rel-trial` | `site-success` | 87.90 | 84.31 | `10K` | `latest` | `[8, 8]` | `mini`[^1] | `32` |
Optimal hyperparameters found via the [RelArena-α](../../../benchmark/relational/relarena) evaluation protocol, using 8 estimators and 20 lag targets:

| Dataset | Task | Val AUROC | Test AUROC | Size | Mode | Neighbors | Text Dim | Text Model |
| :----------- | :---------------- | --------: | ---------: | ----: | :----- | ---------: | -------: | :---------- |
| `rel-amazon` | `user-churn` | 69.35 | 69.47 | `20K` | `last` | `[8, 8]` | `0` | |
| `rel-amazon` | `item-churn` | 82.15 | 82.54 | `20K` | `last` | `[32, 32]` | `0` | |
| `rel-avito` | `user-visits` | 69.56 | 67.11 | `20K` | `last` | `[32, 32]` | `0` | |
| `rel-avito` | `user-clicks` | 63.69 | 68.24 | `20K` | `last` | `[32, 32]` | `0` | |
| `rel-event` | `user-repeat` | 73.30 | 80.90 | `1K` | `rnd` | `[4, 4]` | `0` | |
| `rel-event` | `user-ignore` | 84.67 | 83.80 | `2K` | `last` | `[16, 16]` | `0` | |
| `rel-f1` | `driver-dnf` | 79.37 | 84.18 | `2K` | `last` | `[4, 4]` | `0` | |
| `rel-f1` | `driver-top3` | 88.16 | 90.77 | `1K` | `rnd` | `[4, 4]` | `0` | |
| `rel-hm` | `user-churn` | 70.36 | 69.92 | `20K` | `last` | `[1, 1]` | `0` | |
| `rel-stack` | `user-engagement` | 90.01 | 90.19 | `20K` | `rnd` | `[8, 8]` | `32` | `mini`[^1] |
| `rel-stack` | `user-badge` | 89.32 | 88.09 | `20K` | `rnd` | `[8, 8]` | `64` | `mpnet`[^2] |
| `rel-trial` | `study-outcome` | 66.36 | 73.93 | `20K` | `last` | `[8, 8]` | `32` | `mini`[^1] |

| Dataset | Task | Val nMAE | Test nMAE | Size | Mode | Neighbors | Text Dim | Text Model |
| :----------- | :---------------- | -------: | --------: | ----: | :----- | ---------: | -------: | :--------- |
| `rel-amazon` | `user-ltv` | 21.60 | 25.01 | `20K` | `rnd` | `[64, 64]` | `0` | |
| `rel-amazon` | `item-ltv` | 7.38 | 7.56 | `20K` | `rnd` | `[]` | `0` | |
| `rel-avito` | `ad-ctr` | 32.40 | 32.95 | `2K` | `last` | `[64, 64]` | `0` | |
| `rel-event` | `user-attendance` | 31.25 | 31.70 | `20K` | `rnd` | `[64, 64]` | `0` | |
| `rel-f1` | `driver-position` | 44.63 | 43.55 | `20K` | `last` | `[8, 8]` | `0` | |
| `rel-hm` | `item-sales` | 9.81 | 8.00 | `20K` | `rnd` | `[]` | `0` | |
| `rel-stack` | `post-votes` | 11.36 | 12.49 | `20K` | `rnd` | `[]` | `0` | |
| `rel-trial` | `study-adverse` | 10.21 | 9.42 | `20K` | `last` | `[]` | `32` | `mini`[^1] |
| `rel-trial` | `site-success` | 87.90 | 84.31 | `10K` | `last` | `[8, 8]` | `32` | `mini`[^1] |

[^1]: `sentence-transformers/all-MiniLM-L6-v2`

Expand Down
47 changes: 47 additions & 0 deletions examples/relational/relbench/lag_target.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,47 @@
# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

import pandas as pd
import relbench


def add_lag_target(
df: pd.DataFrame,
task: relbench.base.EntityTask,
num_lags: int,
) -> pd.DataFrame:
r"""Add lag target features to the task table."""
df = df.copy()
right = df.sort_values(task.time_col)
left = pd.DataFrame(
{
task.entity_col: df[task.entity_col],
"__lookup_time__": df[task.time_col],
"__row__": range(len(df)),
}
)
for lag in range(1, num_lags + 1):
column = f"{task.target_col}_lag_{lag}"
if task.task_type == relbench.base.TaskType.REGRESSION:
df[column] = float("NaN")
else:
df[column] = pd.Series(pd.NA, index=df.index, dtype="object")

merged = pd.merge_asof(
left.sort_values("__lookup_time__"),
right,
by=task.entity_col,
left_on="__lookup_time__",
right_on=task.time_col,
direction="backward",
allow_exact_matches=False,
)
mask = merged[task.time_col].notna()
rows = merged.loc[mask, "__row__"].to_numpy()
df.loc[rows, column] = merged.loc[mask, task.target_col].to_numpy()

# Find the next lag strictly before the previous observation.
left = merged.loc[mask, [task.entity_col, "__row__", task.time_col]]
left = left.rename(columns={task.time_col: "__lookup_time__"})

return df
Loading
Loading