FPB Multi-LLM .py

# FPB Multi-LLM .py
# Compute expert-signal features from multiple LLMs (FinBERT, RoBERT
import argparse
import os
import numpy as np
import pandas as pd
from scipy.special import rel_entr

EPS = 1e-12

def row_normalize(a: np.ndarray) -> np.ndarray:
    s = a.sum(axis=1, keepdims=True)
    s[s == 0.0] = 1.0
    return a / s

def kl_divergence(P: np.ndarray, Q: np.ndarray) -> np.ndarray:
    P = np.clip(P, EPS, 1.0)
    Q = np.clip(Q, EPS, 1.0)
    return np.sum(rel_entr(P, Q), axis=1)  # Σ p * log(p/q)

def l1_distance(P: np.ndarray, Q: np.ndarray) -> np.ndarray:
    return 0.5 * np.sum(np.abs(P - Q), axis=1)  # ∈ [0,1]

def load_probs(df: pd.DataFrame, cols: list, label: str) -> np.ndarray:
    missing = [c for c in cols if c not in df.columns]
    if missing:
        raise ValueError(f"[{label}] Missing columns: {missing}")
    arr = df[cols].to_numpy(dtype=float)
    return row_normalize(arr)

def main():
    ap = argparse.ArgumentParser(description="Compute Multi-LLM expert signals (Stage 3 only).")
    ap.add_argument("--input", required=True, help="CSV with per-model probabilities")
    ap.add_argument("--dataset", required=True, help="Dataset tag, e.g., 50Agree | AllAgree")
    ap.add_argument("--out_dir", default=".", help="Output directory")
    ap.add_argument("--out_subdir", default="MultiLLM", help="Subdirectory under out_dir to save multi-llm features")
    # Column names: default to common names; override if your headers differ
    ap.add_argument("--fin_cols", nargs=3, default=["fin_p_neg","fin_p_neu","fin_p_pos"],
                    help="FinBERT prob columns [neg neu pos]")
    ap.add_argument("--rob_cols", nargs=3, default=["rob_p_neg","rob_p_neu","rob_p_pos"],
                    help="RoBERTa prob columns [neg neu pos]")
    args = ap.parse_args()

    df = pd.read_csv(args.input)

    # Load & normalize probabilities
    P = load_probs(df, args.fin_cols, "FinBERT")
    Q = load_probs(df, args.rob_cols, "RoBERTa")

    # Expert-signal features (paper)
    df["MultiLLM_L1_distance"]   = l1_distance(P, Q)
    df["MultiLLM_L1_similarity"] = 1.0 - df["MultiLLM_L1_distance"]
    df["MultiLLM_KL_F_to_R"]     = kl_divergence(P, Q)
    df["MultiLLM_KL_R_to_F"]     = kl_divergence(Q, P)

    # Optional: simple agreement flag (same argmax class)
    df["MultiLLM_agree"] = (np.argmax(P, axis=1) == np.argmax(Q, axis=1)).astype(int)

    save_dir = args.out_dir
    if args.out_subdir:
        save_dir = os.path.join(args.out_dir, args.out_subdir)
    os.makedirs(save_dir, exist_ok=True)
    out_path = os.path.join(save_dir, f"FPB_MultiLLM_{args.dataset}.csv")
    df.to_csv(out_path, index=False)
    print(f"✅ Saved expert-signal features to: {out_path}")
    print("   Added columns: MultiLLM_L1_distance, MultiLLM_L1_similarity, "
          "MultiLLM_KL_F_to_R, MultiLLM_KL_R_to_F, MultiLLM_agree")

if __name__ == "__main__":
    main()