scripts/build_release.py

"""
Build the SAGE open-source dataset release from the local working copies.

Reads the four real-world networks used in
  "Discovering hierarchical governing equations for community-structured
   complex network dynamics"  (SAGE framework)

and re-emits them in one unified, paper-release layout:

    data/<dataset>/
        x.csv [y.csv] [z.csv]   states, shape (T, N), headerless
        adj.csv                 adjacency, shape (N, N), integer 0/1
        community.csv           node,community  (N rows + header)
        metadata.json           machine-readable description
        README.md               human-readable description

State values are copied token-for-token from the sources so that no
floating-point precision is lost or introduced.

Usage:
    python build_release.py
"""

from __future__ import annotations

import csv
import json
import os
import shutil
import sys
from collections import Counter

# --------------------------------------------------------------------------
# Paths
# --------------------------------------------------------------------------

DESKTOP = os.path.join(os.path.expanduser("~"), "Desktop")
PIPELINE = os.path.join(DESKTOP, "agent动力学发现", "实验", "动力学方程发现流程")
SRC_DATA = os.path.join(PIPELINE, "数据集")
SRC_TRAFFIC = os.path.join(os.path.expanduser("~"), "Downloads", "真实交通数据集.zip")
SRC_TRAFFIC_EXTRACTED = os.path.join(
    os.path.dirname(os.path.abspath(__file__)), "_traffic_src"
)

OUT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
OUT_DATA = os.path.join(OUT, "data")


# --------------------------------------------------------------------------
# Low-level helpers
# --------------------------------------------------------------------------

def read_tokens(path: str, delimiter: str = ",") -> list[list[str]]:
    """Read a headerless CSV as raw string tokens."""
    with open(path, newline="", encoding="utf-8-sig") as fh:
        return [row for row in csv.reader(fh, delimiter=delimiter) if row]


def write_tokens(path: str, rows: list[list[str]]) -> None:
    os.makedirs(os.path.dirname(path), exist_ok=True)
    with open(path, "w", newline="", encoding="utf-8") as fh:
        csv.writer(fh, lineterminator="\n").writerows(rows)


def write_states(path: str, rows: list[list[str]]) -> tuple[int, int]:
    """Write a (T, N) state matrix, normalising numeric tokens."""
    out = [[_norm_token(tok) for tok in row] for row in rows]
    write_tokens(path, out)
    return len(out), len(out[0])


def transpose(rows: list[list[str]]) -> list[list[str]]:
    return [list(col) for col in zip(*rows)]


def _norm_token(tok: str) -> str:
    """Keep the source value but drop redundant trailing zeros / -0.0."""
    tok = tok.strip()
    if tok in ("", "nan", "NaN", "None"):
        return "nan"
    try:
        val = float(tok)
    except ValueError:
        return tok
    if val == int(val) and abs(val) < 1e15:
        return str(int(val))
    return repr(val)


def write_adjacency(path: str, rows: list[list[str]]) -> tuple[int, int]:
    n = len(rows)
    if any(len(r) != n for r in rows):
        raise ValueError(f"{path}: adjacency is not square ({n} rows)")
    out = [[str(int(float(v))) for v in row] for row in rows]
    write_tokens(path, out)
    return n, n


def normalise_community(raw: list[list[str]]) -> list[int]:
    """Accept either (node, community) rows or a K-column one-hot block."""
    if raw and len(raw[0]) == 2:
        first = raw[0][0].strip().lower()
        body = raw[1:] if first in ("node", "node_id", "id") else raw
        labels = [int(float(r[1])) for r in body]
    else:
        labels = []
        for row in raw:
            vals = [float(v) for v in row]
            labels.append(max(range(len(vals)), key=vals.__getitem__))
    # Sources already use dense 0..K-1 labels; keep them untouched so that users
    # can compare against the original files. Only remap if the labels are not
    # already dense, and then preserve order of first appearance.
    distinct = sorted(set(labels))
    if distinct == list(range(len(distinct))):
        return labels
    seen: dict[int, int] = {}
    for lab in labels:
        if lab not in seen:
            seen[lab] = len(seen)
    print(f"  note: remapped community labels {distinct} -> 0..{len(seen) - 1}")
    return [seen[lab] for lab in labels]


def write_community(path: str, labels: list[int]) -> Counter:
    body = [["node", "community"]] + [[str(i), str(l)] for i, l in enumerate(labels)]
    write_tokens(path, body)
    return Counter(labels)


def find_source_dir(root: str, *required: str) -> str | None:
    """Return the first directory under `root` that holds all `required` files."""
    for dirpath, _dirnames, filenames in os.walk(root):
        if all(name in filenames for name in required):
            return dirpath
    return None


def write_metadata(path: str, meta: dict) -> None:
    os.makedirs(os.path.dirname(path), exist_ok=True)
    with open(path, "w", encoding="utf-8") as fh:
        json.dump(meta, fh, indent=2, ensure_ascii=False)
        fh.write("\n")


README_TEMPLATE = """# {title}

`{slug}` -- one of the four real-world community-structured networks in
**Discovering hierarchical governing equations for community-structured complex
network dynamics** (SAGE framework).

## Files

| File | Shape | Description |
|------|-------|-------------|
{file_rows}
`T` = time steps, `N` = nodes. State files are headerless CSV; row `t` is the
state of all `N` nodes at time step `t`, and column `i` is the full trajectory of
node `i`.

## At a glance

| Property | Value |
|----------|-------|
| Nodes `N` | {N} |
| Time steps `T` | {T} |
| State variables `D` | {D} |
| Time step `dt` | {dt} |
| Communities `K` | {K} |
| Community sizes | {community_sizes} |
| Adjacency | {adjacency} |
{derivative_row}
## Provenance

{source}

{citation}
## Reading the data

```python
import sys; sys.path.insert(0, "scripts")
from load_dataset import load

ds = load("{slug}")
print(ds)                      # <Dataset '{slug}' T={T} N={N} D={D} K={K}>
print(ds.x.shape, ds.adj.shape, ds.community.shape)
```

The release ships raw observational states only. To reproduce the Savitzky-Golay
smoothing, derivative estimation and 60/20/20 blocked time split used in the
paper, run:

```bash
python scripts/preprocess.py --dataset {slug}
```
"""


# --------------------------------------------------------------------------
# Dataset builders
# --------------------------------------------------------------------------

def _reflow(text: str) -> str:
    """Collapse source-code indentation in a prose block, keeping blank lines."""
    return "\n\n".join(
        para for para in (" ".join(line.strip() for line in chunk.splitlines()).strip()
                          for chunk in text.strip().split("\n\n"))
        if para
    ) + "\n"


def write_readme(path: str, info: dict) -> None:
    file_rows = "".join(
        f"| `{key}.csv` | `(T, N)` | {desc} |\n"
        for key, desc in info["variables"].items()
        if f"{key}.csv" in info["state_files"]
    )
    file_rows += (
        "| `adj.csv` | `(N, N)` | Adjacency matrix, 0/1 integers, zero diagonal |\n"
        "| `community.csv` | `(N, 2)` | `node,community` -- community label per node, as in the source data |\n"
        "| `metadata.json` | -- | Machine-readable description of this dataset |\n"
    )
    fields = dict(info)
    fields["community_sizes"] = ", ".join(str(s) for s in info["community_sizes"])
    fields["file_rows"] = file_rows
    fields["derivative_row"] = (
        f"| Derivative note | {info['derivative']} |\n" if "derivative" in info else ""
    )
    # Re-flow the multi-line source/citation blocks so they render as clean
    # paragraphs without the indentation used in the Python source.
    fields["source"] = _reflow(info["source"])
    fields["citation"] = _reflow(info["citation"])
    text = README_TEMPLATE.format(**fields)
    os.makedirs(os.path.dirname(path), exist_ok=True)
    with open(path, "w", encoding="utf-8") as fh:
        fh.write(text)


def build_mobile_comms() -> dict:
    src = os.path.join(SRC_DATA, "real_mobile_comms")
    dst = os.path.join(OUT_DATA, "urban_mobile_communication")
    os.makedirs(dst, exist_ok=True)

    shapes = {}
    for var in ("x", "y", "z"):
        shapes[var] = write_states(
            os.path.join(dst, f"{var}.csv"), read_tokens(os.path.join(src, f"{var}.csv"))
        )

    n, _ = write_adjacency(os.path.join(dst, "adj.csv"),
                           read_tokens(os.path.join(src, "adj.csv")))
    sizes = write_community(os.path.join(dst, "community.csv"),
                            normalise_community(read_tokens(os.path.join(src, "community.csv"))))

    t = shapes["x"][0]
    return dict(
        slug="urban_mobile_communication",
        title="Urban mobile communication network",
        shape=f"{t} x {n} x 3",
        T=t, N=n, D=3, K=len(sizes),
        community_sizes=[sizes[k] for k in sorted(sizes)],
        variables={"x": "Internet activity", "y": "Phone-call activity", "z": "SMS activity"},
        dt="10 min (native); aggregated to 60 min (6-step mean) in the paper",
        adjacency="Four-neighbour spatial grid on the Milan 100x100 raster",
        source="Telecom Italia Big Data Challenge, Milan, 1-7 November 2013. The "
               "three state variables are the Internet, call and SMS activity of "
               "each grid cell, already normalised per cell by the data provider. "
               "The paper aggregated the native 10-minute samples to 60-minute "
               "resolution (mean over 6 steps), giving 168 time steps.",
        citation="Barlacchi, G. et al. A multi-source dataset of urban life in the "
                 "city of Milan and the Province of Trentino. Sci. Data 2, 150055 "
                 "(2015). https://doi.org/10.1038/sdata.2015.55\n                "
                 "Source portal: https://dataverse.harvard.edu/dataverse/BigDataChallenge",
    )


def build_eeg() -> dict:
    src = os.path.join(SRC_DATA, "real_brain_eeg")
    dst = os.path.join(OUT_DATA, "brain_eeg")
    os.makedirs(dst, exist_ok=True)

    # The '2' variants are the 64-channel / 2440-step / 5-community release used
    # in the paper (Supplementary Section 3.2); the un-suffixed files are an
    # earlier 610-step / 8-community draft and are not part of the release.
    shapes = {}
    for var, srcfile in (("x", "x2.csv"), ("y", "y2.csv")):
        shapes[var] = write_states(
            os.path.join(dst, f"{var}.csv"), read_tokens(os.path.join(src, srcfile))
        )

    n, _ = write_adjacency(os.path.join(dst, "adj.csv"),
                           read_tokens(os.path.join(src, "adj2.csv")))
    sizes = write_community(os.path.join(dst, "community.csv"),
                            normalise_community(read_tokens(os.path.join(src, "community2.csv"))))

    t = shapes["x"][0]
    return dict(
        slug="brain_eeg",
        title="Brain EEG network",
        shape=f"{t} x {n} x 2",
        T=t, N=n, D=2, K=len(sizes),
        community_sizes=[sizes[k] for k in sorted(sizes)],
        variables={"x": "Hilbert analytic amplitude", "y": "Hilbert quadrature component"},
        dt="0.00625 s (160 Hz sampling rate)",
        adjacency="Channel co-activation graph, 806 directed entries",
        source="PhysioNet EEG Motor Movement/Imagery Dataset (EEGMMIDB), record "
               "S001R01, 64 channels. Signals were band-pass filtered at 1-20 Hz "
               "and re-referenced with common average referencing (CAR). Each "
               "channel was then converted to an analytic signal by the Hilbert "
               "transform, giving the two state variables released here: the "
               "amplitude component (x) and the quadrature component (y). The 64 "
               "channels are grouped into five dynamical communities.",
        citation="Schalk, G., McFarland, D. J., Hinterberger, T., Birbaumer, N. & "
                 "Wolpaw, J. R. BCI2000: a general-purpose brain-computer "
                 "interface (BCI) system. IEEE Trans. Biomed. Eng. 51, 1034-1043 "
                 "(2004).\n                PhysioNet: https://physionet.org/content/eegmmidb/1.0.0/",
    )


def build_power_grid() -> dict:
    src = os.path.join(SRC_DATA, "power_grid_rerun")
    dst = os.path.join(OUT_DATA, "power_grid")
    os.makedirs(dst, exist_ok=True)

    # Supplementary Section 3.3 describes a single state variable per bus. The
    # working copy additionally holds y.csv, but y is not an independent
    # observation: it equals the central-difference derivative of x to float32
    # precision (max relative residual 4.7e-8). It is therefore not released as a
    # state variable; the derivative recipe is recorded in metadata instead.
    shapes = {}
    for var in ("x",):
        shapes[var] = write_states(
            os.path.join(dst, f"{var}.csv"), read_tokens(os.path.join(src, f"{var}.csv"))
        )

    n, _ = write_adjacency(os.path.join(dst, "adj.csv"),
                           read_tokens(os.path.join(src, "adj.csv")))
    sizes = write_community(os.path.join(dst, "community.csv"),
                            normalise_community(read_tokens(os.path.join(src, "community.csv"))))

    t = shapes["x"][0]
    return dict(
        slug="power_grid",
        title="Power grid network (IEEE 39-bus New England)",
        shape=f"{t} x {n} x 1",
        T=t, N=n, D=1, K=len(sizes),
        community_sizes=[sizes[k] for k in sorted(sizes)],
        variables={"x": "Bus phase angle (rotor-angle state)"},
        derivative="np.gradient(x, axis=0) / 0.01 reproduces the working copy's "
                   "y.csv exactly (verified to float32 precision). It is a derived "
                   "quantity, not a second state variable, so it is not shipped.",
        dt="0.01 s",
        adjacency="Physical transmission-line topology, 0/1",
        source="IEEE 39-bus New England test system, 39 buses treated as nodes. "
               "Buses are connected when directly linked by a transmission line. "
               "Each bus carries a single state variable, the phase angle (x); "
               "this matches the single-state-variable description in "
               "Supplementary Section 3.3. The buses fall into three dynamical "
               "communities of 13 each.",
        citation="Athay, T., Podmore, R. & Virmani, S. A practical method for the "
                 "direct analysis of transient stability. IEEE Trans. Power "
                 "Appar. Syst. PAS-98, 573-584 (1979).",
    )


def build_rail_transit() -> dict:
    dst = os.path.join(OUT_DATA, "urban_rail_transit")
    os.makedirs(dst, exist_ok=True)

    # The archive was authored on macOS and stores Chinese entry names without
    # the UTF-8 flag, so the extracted folder name differs by platform. Locate
    # the source files by walking the tree rather than by assuming a name.
    src = find_source_dir(SRC_TRAFFIC_EXTRACTED, "in_15min.csv", "out_15min.csv")
    if src is None:
        raise FileNotFoundError(
            f"Traffic sources not found under {SRC_TRAFFIC_EXTRACTED}. "
            f"Extract {SRC_TRAFFIC} into that directory first."
        )

    # Source layout is (N, T); the release layout is (T, N).
    shapes = {}
    for var, srcfile in (("x", "in_15min.csv"), ("y", "out_15min.csv")):
        rows = transpose(read_tokens(os.path.join(src, srcfile)))
        shapes[var] = write_states(os.path.join(dst, f"{var}.csv"), rows)

    n, _ = write_adjacency(os.path.join(dst, "adj.csv"),
                           read_tokens(os.path.join(src, "adj.csv")))
    sizes = write_community(os.path.join(dst, "community.csv"),
                            normalise_community(read_tokens(os.path.join(src, "group.csv"))))

    t = shapes["x"][0]
    return dict(
        slug="urban_rail_transit",
        title="Urban rail transit network (Beijing)",
        shape=f"{t} x {n} x 2",
        T=t, N=n, D=2, K=len(sizes),
        community_sizes=[sizes[k] for k in sorted(sizes)],
        variables={"x": "Station passenger inflow (boarding)",
                   "y": "Station passenger outflow (alighting)"},
        dt="15 min",
        adjacency="Physical rail-network connectivity between stations, 0/1",
        source="Beijing urban rail transit passenger-flow dataset, collected from "
               "the Automatic Fare Collection (AFC) system between 05:00 and 23:00 "
               "on the 25 weekdays within 29 February - 3 April 2016 (five "
               "consecutive weeks; 25 x 18 h x 4 = 1800 steps), covering 17 lines "
               "and 276 stations. Weekends are not included. The two state "
               "variables are the passenger inflow (boarding, x) and outflow "
               "(alighting, y) of each station. The released file was aggregated "
               "to 15-minute resolution; values are raw passenger counts. The 276 "
               "stations are grouped into eight communities.",
        citation="Zhang, J. et al. Deep learning architecture for short-term "
                 "passenger flow forecasting in urban rail transit. IEEE Trans. "
                 "Intell. Transp. Syst. 22, 7004-7014 (2021).",
    )


BUILDERS = {
    "urban_mobile_communication": build_mobile_comms,
    "brain_eeg": build_eeg,
    "power_grid": build_power_grid,
    "urban_rail_transit": build_rail_transit,
}


# --------------------------------------------------------------------------
# Entry point
# --------------------------------------------------------------------------

def main() -> int:
    if not os.path.isdir(SRC_TRAFFIC_EXTRACTED):
        if os.path.exists(SRC_TRAFFIC):
            import zipfile
            zf = zipfile.ZipFile(SRC_TRAFFIC)
            zf.extractall(SRC_TRAFFIC_EXTRACTED)
            zf.close()
        else:
            print(f"WARNING: traffic archive not found at {SRC_TRAFFIC}", file=sys.stderr)

    summary = {}
    for slug, builder in BUILDERS.items():
        # Rebuild from scratch so that files dropped from a dataset (e.g. a
        # derived array no longer classified as a state) do not linger.
        shutil.rmtree(os.path.join(OUT_DATA, slug), ignore_errors=True)
        info = builder()
        info["state_files"] = [
            f"{key}.csv" for key in ("x", "y", "z")
            if os.path.exists(os.path.join(OUT_DATA, slug, f"{key}.csv"))
        ]
        write_metadata(os.path.join(OUT_DATA, slug, "metadata.json"), info)
        write_readme(os.path.join(OUT_DATA, slug, "README.md"), info)
        summary[slug] = info
        print(f"[ok] {slug:28s} {info['shape']:>18s}  K={info['K']}")

    manifest = os.path.join(OUT_DATA, "manifest.json")
    write_metadata(manifest, {
        "release": "SAGE real-world network dataset release",
        "datasets": list(summary.values()),
    })
    print(f"\nWrote {len(summary)} datasets to {OUT_DATA}")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())