Create multitab.py

multitab.py

@@ -0,0 +1,168 @@
+import h5py
+import datasets
+from typing import List
+
+_CITATION = ""  # paste your paper bibtex here if you want
+_DESCRIPTION = """
+MultiTab benchmark datasets (HIGGS, ACS Income, AliExpress) preprocessed
+into HDF5 with a multi-task tabular structure: categorical + numerical
+features and multiple labels per sample.
+"""
+
+_HOMEPAGE = ""
+_LICENSE = ""
+
+
+class MultitabConfig(datasets.BuilderConfig):
+    """Config for one of the MultiTab datasets (e.g. higgs, acs_income, aliexpress)."""
+
+    def __init__(
+        self,
+        h5_path: str,
+        task_names: List[str],
+        **kwargs,
+    ):
+        """
+        Args:
+            h5_path: Name of the HDF5 file in the repo (e.g. 'higgs.h5').
+            task_names: List of label keys in the HDF5 groups for this dataset
+                        (e.g. ['Target', 'm_bb', ...] or ['click', 'conversion']).
+        """
+        super().__init__(**kwargs)
+        self.h5_path = h5_path
+        self.task_names = task_names
+
+
+class Multitab(datasets.GeneratorBasedBuilder):
+    """MultiTab benchmark datasets hosted under one repo."""
+
+    BUILDER_CONFIG_CLASS = MultitabConfig
+    VERSION = datasets.Version("1.0.0")
+
+    BUILDER_CONFIGS = [
+        MultitabConfig(
+            name="higgs",
+            description="HIGGS dataset preprocessed for MultiTab.",
+            h5_path="higgs.h5",
+            task_names=[
+                "Target",
+                "m_bb",
+                "m_jj",
+                "m_jjj",
+                "m_jlv",
+                "m_lv",
+                "m_wbb",
+                "m_wwbb",
+            ],
+        ),
+        MultitabConfig(
+            name="acs_income",
+            description="ACS Income dataset preprocessed for MultiTab.",
+            h5_path="acs_incom.h5",   # matches the upload name
+            task_names=[
+                "MAR",
+                "PINCP",
+            ],
+        ),
+        MultitabConfig(
+            name="aliexpress",
+            description="AliExpress recommendation dataset preprocessed for MultiTab.",
+            h5_path="aliexpress.h5",
+            task_names=[
+                "click",
+                "conversion",
+            ],
+        ),
+    ]
+
+    def _info(self) -> datasets.DatasetInfo:
+        """
+        Schema for each example.
+
+        We keep features as list fields; HF Dataset Viewer will show them as arrays per cell.
+        Both 'features_categorical' and 'features_numerical' exist in the schema; if a given
+        dataset/split doesn't have one of them, we return an empty list.
+        """
+        features_dict = {
+            "features_categorical": datasets.Sequence(datasets.Value("int64")),
+            "features_numerical": datasets.Sequence(datasets.Value("float32")),
+        }
+
+        # One column per task label (float32 for generality; ints are cast)
+        for t in self.config.task_names:
+            features_dict[t] = datasets.Value("float32")
+
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=datasets.Features(features_dict),
+            supervised_keys=None,
+            homepage=_HOMEPAGE,
+            license=_LICENSE,
+            citation=_CITATION,
+        )
+
+    def _split_generators(self, dl_manager: datasets.DownloadManager):
+        """
+        Map HF splits to the 'train', 'val', 'test' groups in your HDF5 file.
+        """
+        # Download/copy the HDF5 from the repo to the local cache
+        h5_path = dl_manager.download(self.config.h5_path)
+
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.TRAIN,
+                gen_kwargs={"filepath": h5_path, "split_name": "train"},
+            ),
+            datasets.SplitGenerator(
+                name=datasets.Split.VALIDATION,
+                gen_kwargs={"filepath": h5_path, "split_name": "val"},
+            ),
+            datasets.SplitGenerator(
+                name=datasets.Split.TEST,
+                gen_kwargs={"filepath": h5_path, "split_name": "test"},
+            ),
+        ]
+
+    def _generate_examples(self, filepath: str, split_name: str):
+        """
+        Iterate over one split in the HDF5 file and yield examples.
+
+        Expected structure per HDF5 group (train/val/test):
+          - features_categorical: (N, C)  [optional]
+          - features_numerical:  (N, D)   [optional]
+          - one array per task in self.config.task_names, shape (N,)
+        """
+
+        cfg = self.config
+
+        with h5py.File(filepath, "r") as f:
+            grp = f[split_name]
+
+            has_cat = "features_categorical" in grp
+            has_num = "features_numerical" in grp
+
+            first_task = cfg.task_names[0]
+            n_samples = grp[first_task].shape[0]
+
+            for idx in range(n_samples):
+                example = {}
+
+                # Categorical features
+                if has_cat:
+                    cat_row = grp["features_categorical"][idx]
+                    example["features_categorical"] = cat_row.astype("int64").tolist()
+                else:
+                    example["features_categorical"] = []
+
+                # Numerical features
+                if has_num:
+                    num_row = grp["features_numerical"][idx]
+                    example["features_numerical"] = num_row.astype("float32").tolist()
+                else:
+                    example["features_numerical"] = []
+
+                # Task labels
+                for t in cfg.task_names:
+                    example[t] = float(grp[t][idx])
+
+                yield idx, example