📰 第27课：新闻推荐项目

🎯 学习目标

🗂️ 课程定位

📖 核心理论

新闻推荐是内容推荐的重要场景，核心挑战是时效性——新闻价值随时间快速衰减。同时新闻冷启动问题突出（每分钟都有新文章），需要内容特征快速匹配。

📐 数学公式

💻 代码实现

import numpy as np
from collections import defaultdict
np.random.seed(42)
print("="*60+"\n新闻推荐项目\n"+"="*60)
n_users=30;n_news=50;n_topics=6
topics=["国际","科技","财经","体育","娱乐","健康"]
# 新闻特征
news_topic=np.random.randint(n_topics,size=n_news)
news_time=np.sort(np.random.randint(1,25,size=n_news))[::-1]  # 时间(小时前)
news_quality=np.random.random(n_news)*0.5+0.5
news_emb=np.random.randn(n_news,8)  # 文章embedding
print(f"用户:{n_users} 新闻:{n_news} 话题:{n_topics}")
# 用户阅读历史
user_reads=defaultdict(list)
user_topic_pref=np.random.dirichlet(np.ones(n_topics),n_users)
for u in range(n_users):
    n_reads=np.random.randint(5,15)
    for _ in range(n_reads):
        topic=np.random.choice(n_topics,p=user_topic_pref[u])
        topic_news=np.where(news_topic==topic)[0]
        if len(topic_news)==0:continue
        nid=np.random.choice(topic_news)
        user_reads[u].append(nid)
print(f"总阅读数:{sum(len(r)for r in user_reads.values())}")
# 时间衰减
def time_decay(hours,half_life=6):
    return 0.5**(hours/half_life)
# 内容推荐
print("\n--- 内容推荐 ---")
target=0
read_set=set(user_reads[target])
user_profile=np.zeros(8)
for nid in user_reads[target]:
    weight=time_decay(np.random.randint(1,24))
    user_profile+=news_emb[nid]*weight
if len(user_reads[target])>0:user_profile/=len(user_reads[target])
content_scores=news_emb@user_profile
content_rec=sorted([(i,content_scores[i])for i in range(n_news)if i not in read_set],key=lambda x:-x[1])[:5]
print(f"内容推荐:{[(f'新闻{i}',f'{s:.3f}')for i,s in content_rec]}")
# 协同推荐
print("\n--- 协同推荐 ---")
read_sets={u:set(reads)for u,reads in user_reads.items()}
def jaccard_sim(a,b):return len(a&b)/len(a|b)if a|b else 0
sims={u:jaccard_sim(read_sets.get(target,set()),read_sets.get(u,set()))for u in range(n_users)if u!=target}
top_sim=sorted(sims.items(),key=lambda x:-x[1])[:5]
cf_candidates=defaultdict(float)
for u,sim in top_sim:
    for nid in read_sets.get(u,set()):
        if nid not in read_set:
            cf_candidates[nid]+=sim
cf_rec=sorted(cf_candidates.items(),key=lambda x:-x[1])[:5]
print(f"协同推荐:{[(f'新闻{i}',f'{s:.3f}')for i,s in cf_rec]}")
# 时效性加权
print("\n--- 时效性加权 ---")
for half_life in [2,6,12,24]:
    time_scores={nid:content_scores[nid]*time_decay(news_time[nid],half_life)for nid in range(n_news)if nid not in read_set}
    rec=sorted(time_scores.items(),key=lambda x:-x[1])[:5]
    avg_age=np.mean([news_time[nid]for nid,_ in rec])
    print(f"半衰期={half_life}h:推荐平均时效={avg_age:.1f}h前")
# 冷启动新闻
print("\n--- 新闻冷启动 ---")
new_news=np.random.randn(3,8)
for i in range(3):
    sims=[np.dot(new_news[i],news_emb[j])/(np.linalg.norm(new_news[i])*np.linalg.norm(news_emb[j])+1e-10)for j in range(n_news)]
    similar=np.argsort(sims)[-3:][::-1]
    print(f"  新新闻{i}:相似已有新闻={similar}(话题={[topics[news_topic[j]]for j in similar]})")
# 评估
print("\n--- 评估 ---")
hits=0;total=0
for u in range(n_users):
    if len(user_reads[u])<2:continue
    hidden=user_reads[u][-1];train=user_reads[u][:-1]
    up=np.zeros(8)
    for nid in train:up+=news_emb[nid]
    if len(train)>0:up/=len(train)
    scores=news_emb@up
    scores[hidden]=-999
    top5=np.argsort(-scores)[:5]
    if hidden in top5:hits+=1
    total+=1
print(f"HitRate@5:{hits/total:.4f}")

🔴 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