Candidate/Customer Obsession & Inclusion Questions

import numpy as np
import pandas as pd
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_auc_score, confusion_matrix

def compute_reweighting_weights(df, protected_col, label_col, alpha=1.0, eps=1e-8):
    """
    df: DataFrame containing protected_col and label_col
    alpha: strength of reweighting (0=no reweighting, 1=full)
    Returns: numpy array of sample weights
    """
    # Marginal P(Y=y)
    py = df[label_col].value_counts(normalize=True).to_dict()
    # Conditional P(Y=y | A=a)
    p_y_given_a = df.groupby(protected_col)[label_col].value_counts(normalize=True).unstack(fill_value=0)
    # Ensure access p(y|a)
    weights = []
    for _, row in df.iterrows():
        a = row[protected_col]
        y = row[label_col]
        p_y = py.get(y, eps)
        p_y_a = p_y_given_a.loc[a].get(y, eps) if a in p_y_given_a.index else eps
        # Inverse-ratio: desired weight ~ P(Y=y)/P(Y=y|A=a)
        inv_ratio = p_y / max(p_y_a, eps)
        # Blend with 1.0 (no-op) by alpha
        w = (1 - alpha) * 1.0 + alpha * inv_ratio
        weights.append(w)
    # Normalize to keep average weight = 1 (stabilizes training)
    weights = np.array(weights, dtype=float)
    weights = weights / weights.mean()
    return weights

# Example training with scikit-learn
def train_with_weights(X_train, y_train, sample_weights):
    clf = LogisticRegression(max_iter=200, solver='lbfgs')
    clf.fit(X_train, y_train, sample_weight=sample_weights)
    return clf

def candidate_nps(responses):
    """
    Compute Net Promoter Score from a list of integers 0-10.
    - responses: iterable of ints or None
    - Returns integer NPS in [-100, 100]
    - Raises ValueError for any non-int or out-of-range value (except None which is ignored)
    - Empty or all-None -> returns 0
    """
    if responses is None:
        return 0

    # Normalize to list to allow multiple passes
    cleaned = []
    for i, r in enumerate(responses):
        if r is None:
            continue
        if not isinstance(r, int):
            raise ValueError(f"Response at index {i} is not an int: {r!r}")
        if r < 0 or r > 10:
            raise ValueError(f"Response at index {i} out of range 0-10: {r}")
        cleaned.append(r)

    total = len(cleaned)
    if total == 0:
        return 0

    promoters = sum(1 for x in cleaned if x >= 9)
    detractors = sum(1 for x in cleaned if x <= 6)

    nps = round((promoters - detractors) / total * 100)
    # Ensure integer and bounds
    return int(max(-100, min(100, nps)))