algo-ecom-ranking
結合相關性、轉換和商業指標，設計電商商品排序
中文翻譯
---
name: "algo-ecom-ranking"
description: "設計多目標的電商商品排序，結合相關性、轉換和商業指標。當使用者需要建立超越文字相關性的商品排序系統、在相關性與商業目標之間取得平衡，或實作 learning-to-rank 時使用，即使他們說的是『product sorting』『search result ranking』或『how to rank products』。"
metadata:
category: "WP-43 電商搜尋演算法"
tags: ["ecommerce", "ranking", "learning-to-rank", "multi-objective"]
---
# 電商商品排序
## 概述
電商排序把文字相關性（BM25）和商業訊號（CTR、轉換率、營收、毛利）合成一個統一的排序分數。用 learning-to-rank（LTR）模型，以點擊和轉換資料訓練，去最佳化對商業有意義的結果。
## 何時使用
**觸發條件：**
- 建立超越純文字相關性的商品搜尋 / 瀏覽排序
- 把商業指標（毛利、庫存）納入排序
- 實作 learning-to-rank 流程
**不該使用的情況：**
- 只需要純文字搜尋相關性時（用 BM25）
- 沒有點擊 / 轉換資料時（先從規則式排序開始）
## 演算法
```
IRON LAW: Relevance Is Necessary But NOT Sufficient for E-Commerce Ranking
A result that is textually relevant but has zero sales history, no
reviews, and is out of stock serves no one. E-commerce ranking must
balance: relevance (does it match the query?), quality (is it a good
product?), and commercial value (does it generate revenue?).
```
### 階段 1：輸入驗證
為每組商品與查詢配對收集特徵：文字相關性分數（BM25）、歷史 CTR、轉換率、平均評分、評論數、價格競爭力、庫存水位、毛利。
**關卡：** 具備最低限度的特徵，且有 30 天以上的點擊資料。
### 階段 2：核心演算法
**規則式基準：** Score = w₁×relevance + w₂×popularity + w₃×rating + w₄×recency。手動調權重。
**LTR 做法：**
1. 從點擊記錄產生訓練資料（點擊 = 正樣本，略過 = 負樣本，並去除版位偏誤）
2. 特徵：文字比對、行為（CTR、加入購物車率）、商品品質（評分、評論）、新鮮度、價格
3. 訓練：LambdaMART 或最佳化 NDCG 的梯度提升排序模型
4. 混合：final_score = α × LTR_score + (1-α) × business_boost
### 階段 3：驗證
離線評估：NDCG@10、MRR。線上 A/B 測試：每次搜尋的營收、點擊率、轉換率。
**關卡：** NDCG 比基準提升，A/B 測試在主要指標上為正面。
### 階段 4：輸出
回傳排好序的商品清單，附上分數拆解。
## 輸出格式
```json
{
"results": [{"product_id": "P123", "rank": 1, "final_score": 0.92, "components": {"relevance": 0.85, "popularity": 0.95, "quality": 0.90}}],
"metadata": {"query": "wireless earbuds", "model": "lambdamart", "ndcg_at_10": 0.72}
}
```
## 範例
### 輸入輸出範例
**輸入：** 查詢「laptop」，500 筆符合的商品
**預期：** 最前面的結果要兼顧文字比對 + 高轉換 + 好評分，不能只看關鍵字相關性。
### 邊界情況
| 輸入 | 預期 | 原因 |
|-------|----------|-----|
| 新商品，沒有歷史 | 依靠文字相關性 + 分類平均 | 冷啟動，沒有行為訊號 |
| 缺貨商品 | 降權或移除 | 顯示買不到的商品會讓使用者不爽 |
| 贊助商品 | 把廣告排名和自然排序混合 | 要清楚區分贊助與自然結果 |
## 注意事項
- **訓練資料的版位偏誤**：排名高的商品不管品質好壞都會有比較多點擊。用逆傾向加權（inverse propensity weighting）或隨機實驗來去除訓練資料的偏誤。
- **熱門偏誤**：沒有多樣性控制的話，熱門商品會壟斷排名，新的或小眾商品完全沒有曝光。要加探索加分。
- **營收最佳化 ≠ 使用者滿意**：依毛利排序會把貴的商品往上推。使用者覺得結果被商業操弄，就會失去信任。
- **特徵新鮮度**：點擊訊號每天都在變。要常常重新訓練或更新特徵。過期的特徵會讓排序品質下降。
- **分類專屬模型**：單一排序模型不一定適用所有分類。3C 的排序和服飾的排序不一樣。
## 參考資料
- LambdaMART 的實作，見 `references/lambdamart.md`
- 去除版位偏誤的技巧，見 `references/position-debiasing.md`英文原文
---
name: "algo-ecom-ranking"
description: "Design multi-objective e-commerce product ranking combining relevance, conversion, and business metrics. Use this skill when the user needs to build a product ranking system beyond text relevance, balance relevance with commercial objectives, or implement learning-to-rank — even if they say 'product sorting', 'search result ranking', or 'how to rank products'."
metadata:
category: "WP-43 電商搜尋演算法"
tags: ["ecommerce", "ranking", "learning-to-rank", "multi-objective"]
---
# E-Commerce Product Ranking
## Overview
E-commerce ranking combines text relevance (BM25) with commercial signals (CTR, conversion rate, revenue, margin) into a unified ranking score. Uses learning-to-rank (LTR) models trained on click and conversion data to optimize for business-relevant outcomes.
## When to Use
**Trigger conditions:**
- Building a product search/browse ranking beyond pure text relevance
- Incorporating business metrics (margin, inventory) into ranking
- Implementing a learning-to-rank pipeline
**When NOT to use:**
- For pure text search relevance only (use BM25)
- When no click/conversion data exists (start with rule-based ranking)
## Algorithm
```
IRON LAW: Relevance Is Necessary But NOT Sufficient for E-Commerce Ranking
A result that is textually relevant but has zero sales history, no
reviews, and is out of stock serves no one. E-commerce ranking must
balance: relevance (does it match the query?), quality (is it a good
product?), and commercial value (does it generate revenue?).
```
### Phase 1: Input Validation
Collect features per product-query pair: text relevance score (BM25), historical CTR, conversion rate, average rating, review count, price competitiveness, inventory level, margin.
**Gate:** Minimum features available, click data from 30+ days.
### Phase 2: Core Algorithm
**Rule-based baseline:** Score = w₁×relevance + w₂×popularity + w₃×rating + w₄×recency. Manually tune weights.
**LTR approach:**
1. Generate training data from click logs (clicked = positive, skipped = negative, with position debiasing)
2. Features: text match, behavioral (CTR, add-to-cart rate), product quality (rating, reviews), freshness, price
3. Train: LambdaMART or gradient-boosted ranking model optimizing NDCG
4. Blend: final_score = α × LTR_score + (1-α) × business_boost
### Phase 3: Verification
Evaluate offline: NDCG@10, MRR. A/B test online: revenue per search, click-through rate, conversion rate.
**Gate:** NDCG improves over baseline, A/B test positive on primary metric.
### Phase 4: Output
Return ranked product list with score decomposition.
## Output Format
```json
{
"results": [{"product_id": "P123", "rank": 1, "final_score": 0.92, "components": {"relevance": 0.85, "popularity": 0.95, "quality": 0.90}}],
"metadata": {"query": "wireless earbuds", "model": "lambdamart", "ndcg_at_10": 0.72}
}
```
## Examples
### Sample I/O
**Input:** Query "laptop", 500 matching products
**Expected:** Top results balance text match + high conversion + good ratings, not just keyword relevance.
### Edge Cases
| Input | Expected | Why |
|-------|----------|-----|
| New product, no history | Rely on text relevance + category avg | Cold start — no behavioral signal |
| Out of stock item | Demote or remove | Showing unavailable products frustrates users |
| Sponsored product | Blend ad rank with organic | Separate sponsored from organic clearly |
## Gotchas
- **Position bias in training data**: Higher-ranked items get more clicks regardless of quality. Debias training data using inverse propensity weighting or randomization experiments.
- **Popularity bias**: Without diversity controls, popular items dominate rankings. New or niche products get no exposure. Add exploration bonus.
- **Revenue optimization ≠ user satisfaction**: Ranking by margin pushes expensive products up. Users lose trust if results feel commercially manipulated.
- **Feature freshness**: Click signals change daily. Retrain or update features frequently. Stale features degrade ranking quality.
- **Category-specific models**: A single ranking model may not work across all categories. Electronics ranking differs from fashion ranking.
## References
- For LambdaMART implementation, see `references/lambdamart.md`
- For position debiasing techniques, see `references/position-debiasing.md`