cdp-edge 1.0.0

package/server-edge-tracker/modules/ml/logistic.ts

@@ -0,0 +1,226 @@
+/**
+ * CDP Edge — Logistic Regression (pure TS, sem deps externas)
+ * Treina modelo de predição de conversão com dados reais do D1.
+ *
+ * Features usadas (todas normalizadas 0-1):
+ *   utm_source, engagement_score, intention_level, recency,
+ *   has_email, has_phone, is_br, hour_normalized
+ */
+
+import { Env } from '../../types.js';
+import { D1Database, KVNamespace } from '@cloudflare/workers-types';
+
+// ── Tipos ───────────────────────────────────────────────────────────────────────
+export interface DatasetRow {
+  features: number[];
+  label: number;
+}
+
+export interface LogisticModel {
+  bias: number;
+  weights: number[];
+  accuracy: number;
+  positiveRate: number;
+  sampleSize: number;
+  threshold: number;
+  featureNames: string[];
+  trainedAt: string;
+}
+
+export interface ExtractedFeatures {
+  utm_score: number;
+  engagement: number;
+  intention: number;
+  recency: number;
+  has_email: number;
+  has_phone: number;
+  is_br: number;
+  hour: number;
+}
+
+// ── Feature Engineering ───────────────────────────────────────────────────────
+
+const UTM_SCORES: Record<string, number> = {
+  facebook: 0.90, instagram: 0.90, meta: 0.90,
+  google:   0.82, youtube:   0.82,
+  tiktok:   0.75,
+  email:    0.68, sms: 0.68,
+  organic:  0.30,
+  direct:   0.20,
+};
+
+const INTENTION_SCORES: Record<string, number> = {
+  comprador:   1.00, high_intent: 1.00,
+  interessado: 0.60,
+  nurture:     0.30,
+  curioso:     0.15,
+};
+
+export function extractFeatures(row: any): number[] {
+  const src       = (row.utm_source || '').toLowerCase().trim();
+  const intention = (row.intention_level || '').toLowerCase().trim();
+  const daysSince = row.days_since_lead || 0;
+
+  return [
+    UTM_SCORES[src] ?? (src ? 0.10 : 0.05),                    // utm_score
+    Math.min((row.engagement_score || 0) / 5, 1),               // engagement (0-5 → 0-1)
+    INTENTION_SCORES[intention] ?? 0,                           // intention
+    Math.max(0, 1 - daysSince / 90),                            // recency (0=90 dias, 1=hoje)
+    row.has_email    ? 1 : 0,                                   // has_email
+    row.has_phone    ? 1 : 0,                                   // has_phone
+    row.is_br        ? 1 : 0,                                   // is_br
+    ((row.hour || 12) / 23),                                    // hour normalized
+  ];
+}
+
+// ── Sigmoid ───────────────────────────────────────────────────────────────────
+
+function sigmoid(z: number): number {
+  if (z > 20)  return 1;
+  if (z < -20) return 0;
+  return 1 / (1 + Math.exp(-z));
+}
+
+function dot(weights: number[], features: number[]): number {
+  return features.reduce((sum, f, i) => sum + (weights[i] || 0) * f, 0);
+}
+
+// ── Treinamento ───────────────────────────────────────────────────────────────
+
+/**
+ * Treina regressão logística com gradiente descendente.
+ */
+export function trainLogisticRegression(dataset: DatasetRow[], opts: { iterations?: number; learningRate?: number; lambda?: number } = {}): LogisticModel | null {
+  if (!dataset || dataset.length < 50) {
+    return null; // dados insuficientes
+  }
+
+  const iterations   = opts.iterations   || 200;
+  const learningRate = opts.learningRate  || 0.1;
+  const lambda       = opts.lambda        || 0.01; // L2 regularization
+  const nFeatures    = dataset[0].features.length;
+
+  let bias    = 0;
+  let weights = new Array(nFeatures).fill(0);
+
+  const positives = dataset.filter(d => d.label === 1).length;
+  const positiveRate = positives / dataset.length;
+
+  // Se menos de 5% positivos, não treina (dados de compra insuficientes)
+  if (positiveRate < 0.03) return null;
+
+  for (let iter = 0; iter < iterations; iter++) {
+    let dBias    = 0;
+    const dWeights = new Array(nFeatures).fill(0);
+
+    for (const { features, label } of dataset) {
+      const z     = dot(weights, features) + bias;
+      const pred  = sigmoid(z);
+      const error = pred - label;
+
+      dBias += error;
+      for (let j = 0; j < nFeatures; j++) {
+        dWeights[j] += error * features[j];
+      }
+    }
+
+    const n = dataset.length;
+    bias -= learningRate * (dBias / n);
+    for (let j = 0; j < nFeatures; j++) {
+      // L2: penaliza pesos grandes para evitar overfitting
+      weights[j] -= learningRate * ((dWeights[j] / n) + lambda * weights[j]);
+    }
+  }
+
+  // Calcular acurácia no conjunto de treino
+  let correct = 0;
+  const threshold = positiveRate > 0.3 ? 0.5 : Math.max(0.3, positiveRate * 1.5);
+
+  for (const { features, label } of dataset) {
+    const z    = dot(weights, features) + bias;
+    const pred = sigmoid(z) >= threshold ? 1 : 0;
+    if (pred === label) correct++;
+  }
+
+  const accuracy = correct / dataset.length;
+
+  return {
+    bias,
+    weights,
+    accuracy,
+    positiveRate,
+    sampleSize: dataset.length,
+    threshold,
+    featureNames: ['utm_score', 'engagement', 'intention', 'recency', 'has_email', 'has_phone', 'is_br', 'hour'],
+    trainedAt: new Date().toISOString(),
+  };
+}
+
+// ── Inferência ────────────────────────────────────────────────────────────────
+
+/**
+ * Prediz score de conversão (0-100) usando pesos treinados.
+ */
+export function predictWithWeights(model: LogisticModel, features: number[]): number {
+  const z    = dot(model.weights, features) + model.bias;
+  const prob = sigmoid(z);
+  return Math.round(prob * 100);
+}
+
+// ── Helpers de persistência ───────────────────────────────────────────────────
+
+export const LTV_WEIGHTS_KV_KEY = 'ltv_weights_active';
+
+export async function loadActiveWeights(env: Env): Promise<LogisticModel | null> {
+  // 1. Tentar KV (cache ~7 dias)
+  if (env.GEO_CACHE) {
+    try {
+      const cached = await env.GEO_CACHE.get(LTV_WEIGHTS_KV_KEY, 'json') as LogisticModel | null;
+      if (cached?.weights?.length) return cached;
+    } catch (err: any) {
+      console.error('[Logistic] Error fetching LTV weights from KV cache:', {
+        key: LTV_WEIGHTS_KV_KEY,
+        error: err?.message || String(err),
+        stack: err?.stack,
+      });
+    }
+  }
+
+  // 2. Fallback: D1
+  if (!env.DB) return null;
+  try {
+    const row = await env.DB.prepare(
+      `SELECT weights_json FROM ltv_model_weights WHERE is_active = 1 ORDER BY trained_at DESC LIMIT 1`
+    ).first();
+    if (!row || !(row as any).weights_json) return null;
+    const model = JSON.parse((row as any).weights_json) as LogisticModel;
+
+    // Popular KV para próximas requests
+    if (env.GEO_CACHE && model?.weights?.length) {
+      env.GEO_CACHE.put(LTV_WEIGHTS_KV_KEY, JSON.stringify(model), { expirationTtl: 604800 }).catch(() => {});
+    }
+    return model;
+  } catch {
+    return null;
+  }
+}
+
+export async function saveWeights(DB: D1Database, model: LogisticModel): Promise<void> {
+  if (!DB || !model) return;
+  const now = new Date().toISOString();
+
+  // Desativar modelo anterior
+  await DB.prepare(`UPDATE ltv_model_weights SET is_active = 0 WHERE is_active = 1`).run();
+
+  // Inserir novo como ativo
+  await DB.prepare(`
+    INSERT INTO ltv_model_weights (trained_at, is_active, sample_size, positive_rate, accuracy, weights_json)
+    VALUES (?, 1, ?, ?, ?, ?)
+  `).bind(
+    now,
+    model.sampleSize,
+    model.positiveRate,
+    model.accuracy,
+    JSON.stringify(model),
+  ).run();
+}