如何使用 Claude Code 实现 A/B 测试

A/B 测试实现全景

A/B 测试是数据驱动决策不可或缺的手段。然而，要自己从零实现正确的统计处理、用户分流和结果追踪并非易事。借助 Claude Code，你可以高效构建统计学上严谨的 A/B 测试基础设施。

测试分流引擎

> 创建一个用户 A/B 分组分流引擎。
> 确保同一用户始终返回相同的变体。

import crypto from 'crypto';

interface Experiment {
  id: string;
  variants: { id: string; weight: number }[];
}

class ABTestEngine {
  private experiments: Map<string, Experiment> = new Map();

  register(experiment: Experiment) {
    const totalWeight = experiment.variants.reduce((sum, v) => sum + v.weight, 0);
    if (Math.abs(totalWeight - 100) > 0.01) {
      throw new Error('变体权重之和必须为 100');
    }
    this.experiments.set(experiment.id, experiment);
  }

  assign(experimentId: string, userId: string): string {
    const experiment = this.experiments.get(experimentId);
    if (!experiment) throw new Error(`未找到实验: ${experimentId}`);

    const hash = crypto
      .createHash('md5')
      .update(`${experimentId}:${userId}`)
      .digest('hex');
    const bucket = parseInt(hash.substring(0, 8), 16) % 100;

    let cumulative = 0;
    for (const variant of experiment.variants) {
      cumulative += variant.weight;
      if (bucket < cumulative) return variant.id;
    }
    return experiment.variants[0].id;
  }
}

// 使用示例
const engine = new ABTestEngine();
engine.register({
  id: 'checkout-flow',
  variants: [
    { id: 'control', weight: 50 },
    { id: 'new-design', weight: 50 },
  ],
});

在 React 组件中使用

import { createContext, useContext, useEffect } from 'react';

function useExperiment(experimentId: string): string {
  const engine = useContext(ABTestContext);
  const userId = useCurrentUserId();
  const variant = engine.assign(experimentId, userId);

  useEffect(() => {
    trackEvent('experiment_exposure', {
      experimentId,
      variant,
      userId,
    });
  }, [experimentId, variant, userId]);

  return variant;
}

// 在组件中使用
function CheckoutPage() {
  const variant = useExperiment('checkout-flow');

  return variant === 'new-design'
    ? <NewCheckoutFlow />
    : <CurrentCheckoutFlow />;
}

结果的统计分析

为了正确评估 A/B 测试的结果，需要进行统计显著性计算。

interface TestResult {
  sampleSize: number;
  conversions: number;
}

function calculateSignificance(control: TestResult, treatment: TestResult) {
  const p1 = control.conversions / control.sampleSize;
  const p2 = treatment.conversions / treatment.sampleSize;

  const pooledP = (control.conversions + treatment.conversions) /
    (control.sampleSize + treatment.sampleSize);

  const se = Math.sqrt(
    pooledP * (1 - pooledP) * (1 / control.sampleSize + 1 / treatment.sampleSize)
  );

  const zScore = (p2 - p1) / se;
  const pValue = 2 * (1 - normalCDF(Math.abs(zScore)));

  return {
    controlRate: (p1 * 100).toFixed(2) + '%',
    treatmentRate: (p2 * 100).toFixed(2) + '%',
    improvement: (((p2 - p1) / p1) * 100).toFixed(2) + '%',
    pValue: pValue.toFixed(4),
    significant: pValue < 0.05,
  };
}

function normalCDF(x: number): number {
  const a1 = 0.254829592, a2 = -0.284496736;
  const a3 = 1.421413741, a4 = -1.453152027;
  const a5 = 1.061405429, p = 0.3275911;
  const sign = x < 0 ? -1 : 1;
  x = Math.abs(x) / Math.sqrt(2);
  const t = 1.0 / (1.0 + p * x);
  const y = 1.0 - ((((a5 * t + a4) * t + a3) * t + a2) * t + a1) * t * Math.exp(-x * x);
  return 0.5 * (1.0 + sign * y);
}

总结

借助 Claude Code，你可以从用户分流到统计显著性计算，一站式构建 A/B 测试基础设施。关于与功能开关的联动，请参阅功能开关实现；关于数据分析集成，请参阅数据分析实现指南。

统计检验理论方面，推荐参考 Evan Miller - Sample Size Calculator。