rank_through_time.py

import matplotlib
matplotlib.use("Agg")
matplotlib.rcParams["figure.dpi"] = 150
import pathlib
import pandas as pd
import matplotlib.pyplot as plt
import matplotlib.ticker as mticker


def _add_ranks(df):
    df = df.copy()
    df["cutoff"] = pd.to_datetime(df["cutoff"])
    df["rank"] = df.groupby(["metric", "subdataset", "frequency", "cutoff"])[
        "value"
    ].rank(method="min")
    return df


def _style_rank_ax(ax, n_models):
    ax.set_ylabel("Rank")
    ax.set_ylim(n_models + 0.5, 0.5)
    ax.yaxis.set_major_locator(mticker.MultipleLocator(1))
    ax.tick_params(axis="x", rotation=45)
    ax.grid(True, alpha=0.3)


def _style_value_ax(ax, metric):
    ax.set_ylabel(metric)
    ax.tick_params(axis="x", rotation=45)
    ax.grid(True, alpha=0.3)


def _finish_fig(fig):
    """Add a single shared legend at the bottom and adjust layout."""
    handles, labels = fig.axes[0].get_legend_handles_labels()
    fig.legend(
        handles, labels,
        loc="lower center",
        ncol=min(len(labels), 4),
        fontsize="small",
        bbox_to_anchor=(0.5, 0),
    )
    fig.subplots_adjust(bottom=0.18)
    fig.tight_layout(rect=[0, 0.08, 1, 1])


# ── Public figure builders ───────────────────────────────────────────────────


def plot_rank_per_category(df, metric):
    """Grid of rank-over-time subplots, one per (subdataset, frequency)."""
    df = _add_ranks(df)
    models = sorted(df["model"].unique())
    n_models = len(models)
    categories = sorted(
        df[["subdataset", "frequency"]]
        .drop_duplicates()
        .itertuples(index=False, name=None)
    )

    fig, axes = plt.subplots(
        nrows=len(categories), ncols=1,
        figsize=(10, 4 * len(categories)),
        sharex=False, sharey=True,
    )
    if len(categories) == 1:
        axes = [axes]

    for ax, (subdataset, frequency) in zip(axes, categories):
        sub = df[
            (df["metric"] == metric)
            & (df["subdataset"] == subdataset)
            & (df["frequency"] == frequency)
        ]
        pivot = sub.pivot_table(index="cutoff", columns="model", values="rank").sort_index()
        for model in models:
            if model in pivot.columns:
                ax.plot(pivot.index, pivot[model], marker="o", label=model)
        ax.set_title(f"{subdataset} / {frequency}")
        _style_rank_ax(ax, n_models)

    fig.suptitle(f"Rank through time — {metric.upper()}", fontsize=14)
    _finish_fig(fig)
    return fig


def plot_avg_rank(df, metric):
    """Average rank across all categories over time."""
    df = _add_ranks(df)
    models = sorted(df["model"].unique())
    n_models = len(models)
    sub = df[df["metric"] == metric]
    avg_rank = (
        sub.groupby(["model", "cutoff"])["rank"]
        .mean()
        .reset_index()
        .rename(columns={"rank": "avg_rank"})
    )
    pivot = avg_rank.pivot_table(index="cutoff", columns="model", values="avg_rank").sort_index()

    fig, ax = plt.subplots(figsize=(10, 5))
    for model in models:
        if model in pivot.columns:
            ax.plot(pivot.index, pivot[model], marker="o", label=model)
    ax.set_title(f"Average rank across all categories — {metric}", fontsize=14)
    ax.set_xlabel("Cutoff date")
    _style_rank_ax(ax, n_models)
    _finish_fig(fig)
    return fig


def plot_value_per_category(df, metric):
    """Grid of raw-metric-over-time subplots, one per (subdataset, frequency)."""
    df = df.copy()
    df["cutoff"] = pd.to_datetime(df["cutoff"])
    models = sorted(df["model"].unique())
    categories = sorted(
        df[["subdataset", "frequency"]]
        .drop_duplicates()
        .itertuples(index=False, name=None)
    )

    fig, axes = plt.subplots(
        nrows=len(categories), ncols=1,
        figsize=(10, 4 * len(categories)),
        sharex=False,
    )
    if len(categories) == 1:
        axes = [axes]

    for ax, (subdataset, frequency) in zip(axes, categories):
        sub = df[
            (df["metric"] == metric)
            & (df["subdataset"] == subdataset)
            & (df["frequency"] == frequency)
        ]
        pivot = sub.pivot_table(index="cutoff", columns="model", values="value").sort_index()
        for model in models:
            if model in pivot.columns:
                ax.plot(pivot.index, pivot[model], marker="o", label=model)
        ax.set_title(f"{subdataset} / {frequency}")
        _style_value_ax(ax, metric)

    fig.suptitle(f"Model {metric.upper()} through time", fontsize=14)
    _finish_fig(fig)
    return fig


def plot_avg_value(df, metric):
    """Average raw metric across all categories over time."""
    df = df.copy()
    df["cutoff"] = pd.to_datetime(df["cutoff"])
    models = sorted(df["model"].unique())
    sub = df[df["metric"] == metric]
    avg_val = (
        sub.groupby(["model", "cutoff"])["value"]
        .mean()
        .reset_index()
        .rename(columns={"value": "avg_value"})
    )
    pivot = avg_val.pivot_table(index="cutoff", columns="model", values="avg_value").sort_index()

    fig, ax = plt.subplots(figsize=(10, 5))
    for model in models:
        if model in pivot.columns:
            ax.plot(pivot.index, pivot[model], marker="o", label=model)
    ax.set_title(f"Average {metric} across all categories", fontsize=14)
    ax.set_xlabel("Cutoff date")
    _style_value_ax(ax, metric)
    _finish_fig(fig)
    return fig


def plot_rank_for_subdataset(df, metric, subdataset):
    """Rank over time for a single subdataset (all frequencies as subplots)."""
    df = _add_ranks(df)
    models = sorted(df["model"].unique())
    n_models = len(models)
    frequencies = sorted(
        df[df["subdataset"] == subdataset]["frequency"].unique()
    )

    fig, axes = plt.subplots(
        nrows=len(frequencies), ncols=1,
        figsize=(10, 4 * len(frequencies)),
        sharex=False, sharey=True,
        squeeze=False,
    )

    for ax_row, frequency in zip(axes, frequencies):
        ax = ax_row[0]
        sub = df[
            (df["metric"] == metric)
            & (df["subdataset"] == subdataset)
            & (df["frequency"] == frequency)
        ]
        pivot = sub.pivot_table(index="cutoff", columns="model", values="rank").sort_index()
        for model in models:
            if model in pivot.columns:
                ax.plot(pivot.index, pivot[model], marker="o", label=model)
        ax.set_title(f"{subdataset} / {frequency}")
        _style_rank_ax(ax, n_models)

    fig.suptitle(f"Rank through time — {metric.upper()}", fontsize=14)
    _finish_fig(fig)
    return fig


def plot_value_for_subdataset(df, metric, subdataset):
    """Raw metric over time for a single subdataset (all frequencies as subplots)."""
    df = df.copy()
    df["cutoff"] = pd.to_datetime(df["cutoff"])
    models = sorted(df["model"].unique())
    frequencies = sorted(
        df[df["subdataset"] == subdataset]["frequency"].unique()
    )

    fig, axes = plt.subplots(
        nrows=len(frequencies), ncols=1,
        figsize=(10, 4 * len(frequencies)),
        sharex=False,
        squeeze=False,
    )

    for ax_row, frequency in zip(axes, frequencies):
        ax = ax_row[0]
        sub = df[
            (df["metric"] == metric)
            & (df["subdataset"] == subdataset)
            & (df["frequency"] == frequency)
        ]
        pivot = sub.pivot_table(index="cutoff", columns="model", values="value").sort_index()
        for model in models:
            if model in pivot.columns:
                ax.plot(pivot.index, pivot[model], marker="o", label=model)
        ax.set_title(f"{subdataset} / {frequency}")
        _style_value_ax(ax, metric)

    fig.suptitle(f"Model {metric.upper()} through time", fontsize=14)
    _finish_fig(fig)
    return fig


# ── CLI: save all figures to disk ────────────────────────────────────────────

if __name__ == "__main__":
    OUT = pathlib.Path("figures/rank_through_time")
    OUT.mkdir(parents=True, exist_ok=True)

    raw = pd.read_csv("mock_evaluation_results.csv")
    raw = raw[raw["model"] != "zero_model"]
    metrics = sorted(raw["metric"].unique())

    for metric in metrics:
        for fn, prefix in [
            (plot_rank_per_category, "rank_per_category"),
            (plot_value_per_category, "value_per_category"),
            (plot_avg_rank, "avg_rank"),
            (plot_avg_value, "avg_value"),
        ]:
            fig = fn(raw, metric)
            path = OUT / f"{prefix}_{metric}.png"
            fig.savefig(path, dpi=150, bbox_inches="tight")
            plt.close(fig)
            print(f"Saved {path}")

    print("Done.")