衡量反馈的质量与相关性

来自人类反馈的强化学习（RLHF）

Mina Parham

AI Engineer

异常反馈的应用

例如：

评价（正面）：
- "我很喜欢这个产品！"
评价（负面）：
- "服务很差。"
评价（中性）：
- "做到了该做的。"
评价（离群）：
- "天空是蓝的。"

四星评分，手正加上第五颗星

检测异常反馈

import numpy as np

def least_confidence(prob_dist):
    simple_least_conf = np.nanmax(prob_dist) 
    num_labels = float(prob_dist.size)  # number of labels
    least_conf = (1 - simple_least_conf) * (num_labels / (num_labels - 1))
    return least_conf

def filter_low_confidence_predictions(prob_dists, threshold=0.5):
    filtered_indices = [i for i, prob_dist in enumerate(prob_dists) 
                        if least_confidence(prob_dist) > threshold]
    return filtered_indices

检测异常反馈

prob_distribution_array = np.array([
    [0.1, 0.1, 0.2],   # Low confidence (0.2)
    [0.6, 0.2, 0.1],   # High confidence (0.6)
    [0.3, 0.3, 0.4]   # Medium confidence (0.4)
])

# Filter function with 0.5 threshold
filtered_feedback_indices, filtered_confidences = 
filter_low_confidence_predictions(prob_distribution_array, threshold=0.5)

print(f"Filtered Confidence Scores: {filtered_confidences}")

Filtered Confidence Scores: [0.6]

K-means

适合检测异常，且实现快速
用领域知识或分析方法确定聚类数

k-means 算法示意图。

用 k-means 进行异常检测

import numpy as np
import pandas as pd
from sklearn.cluster import KMeans


def detect_anomalies(data, n_clusters=3):
    kmeans = KMeans(n_clusters=n_clusters, random_state=42)
    clusters = kmeans.fit_predict(data)
    centers = kmeans.cluster_centers_

    # Calculate distances from cluster centers
    distances = np.linalg.norm(data - centers[clusters], axis=1)
    return distances

用 k-means 进行异常检测

feedback_data = np.array([
    [4.0],  # Close to center of cluster
    [4.5],  # Close to center of cluster
    [1.0],  # Anomaly - far from main group
    [4.1],  # Close to center of cluster
    [3.9]  # Close to center of cluster
])

anomalies = detect_anomalies(confidences, n_clusters=1)

print(anomalies)

[0.5 1.  2.5   0.6 0.4]

Vamos praticar!

来自人类反馈的强化学习（RLHF）