🔀 第15课：重排策略

🎯 学习目标

🗂️ 课程定位

📖 核心理论

重排（Re-ranking）是推荐系统流水线的最后一环，在精排基础上进行全局优化。重排不追求单点最优，而是优化整个推荐列表的质量，包括多样性、新鲜度、业务规则等。

📐 数学公式

💻 代码实现

import numpy as np
np.random.seed(42)
print("="*60+"\n重排策略\n"+"="*60)
n_items=30;n_candidates=15
# 精排分数
scores=np.random.random(n_candidates)*5+1
categories=np.random.choice(["科技","娱乐","体育","财经","生活"],n_candidates)
freshness=np.random.random(n_candidates)  # 0=旧,1=新
print(f"候选物品数:{n_candidates} 类别分布:{dict(zip(*np.unique(categories,return_counts=True)))}")
# 原始排序
orig_order=np.argsort(-scores)
print(f"\n精排Top-10:")
for i in orig_order[:10]:
    print(f"  物品{i}:分数={scores[i]:.2f} 类别={categories[i]} 新鲜度={freshness[i]:.2f}")
# MMR多样性重排
def mmr_rerank(scores,sim_matrix,lam=0.5,top_k=10):
    selected=[];candidates=list(range(len(scores)))
    for _ in range(min(top_k,len(scores))):
        mmr_scores={}
        for i in candidates:
            relevance=scores[i]
            if selected:
                max_sim=max(sim_matrix[i][j]for j in selected)
            else:max_sim=0
            mmr_scores[i]=lam*relevance-(1-lam)*max_sim
        best=max(mmr_scores,key=mmr_scores.get)
        selected.append(best);candidates.remove(best)
    return selected
# 物品相似度矩阵
item_embs=np.random.randn(n_candidates,8)
norms=np.linalg.norm(item_embs,axis=1); sim_matrix=(item_embs@item_embs.T)/(norms[:,None]*norms[None,:]+1e-10)
mmr_order=mmr_rerank(scores,sim_matrix,lam=0.6)
print(f"\nMMR重排Top-10:")
for i in mmr_order[:10]:
    print(f"  物品{i}:分数={scores[i]:.2f} 类别={categories[i]}")
# 类别多样性
def diversity_score(order,categories,top_k=10):
    cats=[categories[i]for i in order[:top_k]]
    return len(set(cats))/len(set(categories))
print(f"\n类别多样性: 原始={diversity_score(orig_order,categories):.2f} MMR={diversity_score(mmr_order,categories):.2f}")
# 新鲜度加权
print("\n--- 新鲜度加权 ---")
fresh_weight=0.3
fresh_scores=scores*(1-fresh_weight)+freshness*fresh_weight*5
fresh_order=np.argsort(-fresh_scores)
print(f"新鲜度加权Top-5:")
for i in fresh_order[:5]:
    print(f"  物品{i}:原始分={scores[i]:.2f} 新鲜度={freshness[i]:.2f} 综合分={fresh_scores[i]:.2f}")
# 打散策略
print("\n--- 打散策略 ---")
def scatter_rerank(order,categories,min_gap=2):
    result=[];last_cat_count={}
    for i in order:
        cat=categories[i]
        last_pos=last_cat_count.get(cat,-999)
        if len(result)-last_pos>=min_gap or not result:
            result.append(i)
            last_cat_count[cat]=len(result)-1
        else:
            # 寻找后续可插入位置
            for j in range(len(result)+1,len(order)):
                if j<1:continue
            result.append(i)  # 简化：强行插入
            last_cat_count[cat]=len(result)-1
    return result
scatter_order=scatter_rerank(orig_order,categories)
print(f"打散后Top-10:")
for i in scatter_order[:10]:
    print(f"  物品{i}:分数={scores[i]:.2f} 类别={categories[i]}")
# 业务规则重排
print("\n--- 业务规则 ---")
# 置顶规则+去重+过滤
fixed_top=2  # 前2位固定
banned=set([3])  # 过滤物品3
final_order=[]
final_order.extend(orig_order[:fixed_top])
for i in orig_order[fixed_top:]:
    if i not in banned and i not in final_order:
        final_order.append(i)
print(f"业务规则后Top-10: {final_order[:10]}")
# 综合评估
print("\n--- 综合评估 ---")
def eval_rerank(name,order):
    avg_score=np.mean([scores[i]for i in order[:10]])
    div=diversity_score(order,categories)
    avg_fresh=np.mean([freshness[i]for i in order[:10]])
    print(f"  {name}: 平均分={avg_score:.2f} 多样性={div:.2f} 新鲜度={avg_fresh:.2f}")
eval_rerank("原始排序",orig_order)
eval_rerank("MMR重排",mmr_order)
eval_rerank("新鲜度加权",fresh_order)

🔴 运行结果

✅ 验证通过

🔍 代码解读

💡 关键洞察

📝 本章小结

🧪 练习

📚 延伸阅读

💼 面试考点

🏭 工业案例