📱 第29课：短视频推荐项目

🎯 学习目标

🗂️ 课程定位

📖 核心理论

短视频推荐是当前最活跃的推荐场景。抖音/TikTok的成功证明了短视频推荐的巨大商业价值。核心挑战是多目标优化（完播率、互动率、时长）和内容冷启动。

📐 数学公式

💻 代码实现

import numpy as np
from collections import defaultdict
np.random.seed(42)
print("="*60+"\n短视频推荐项目\n"+"="*60)
n_users=40;n_videos=50
# 视频特征
video_duration=np.random.randint(5,60,n_videos)  # 秒
video_category=np.random.choice(["搞笑","知识","生活","游戏","音乐"],n_videos)
video_quality=np.random.random(n_videos)*0.5+0.5
video_freshness=np.random.randint(1,48,n_videos)  # 小时
video_emb=np.random.randn(n_videos,8)
print(f"用户:{n_users} 视频:{n_videos}")
# 用户行为数据
behaviors=defaultdict(list)
for u in range(n_users):
    n_watched=np.random.randint(10,30)
    for _ in range(n_watched):
        vid=np.random.randint(n_videos)
        duration=video_duration[vid]
        watch_time=min(np.random.exponential(duration*0.6),duration)
        completion_rate=watch_time/duration
        liked=np.random.random()<0.15
        shared=np.random.random()<0.05
        commented=np.random.random()<0.08
        behaviors[u].append((vid,completion_rate,liked,shared,commented,watch_time))
total_actions=sum(len(b)for b in behaviors.values())
print(f"总行为数:{total_actions}")
# 多目标建模
print("\n--- 多目标排序 ---")
def sigmoid(x):return 1/(1+np.exp(-np.clip(x,-500,500)))
# 完播率模型
completion_labels=np.array([b[1]>0.8for u in range(n_users)for b in behaviors[u]])
# 互动率(点赞+评论+分享)
interact_labels=np.array([b[2]or b[3]or b[4]for u in range(n_users)for b in behaviors[u]])
# 时长模型
duration_labels=np.array([b[5]for u in range(n_users)for b in behaviors[u]])
print(f"完播率:{completion_labels.mean():.4f} 互动率:{interact_labels.mean():.4f} 平均时长:{duration_labels.mean():.1f}s")
# 简单多目标融合
target=0
target_behaviors=behaviors[target]
watched=set(b[0]for b in target_behaviors)
# 计算每个视频的多目标得分
multi_scores={}
for v in range(n_videos):
    if v in watched:continue
    # 完播率预估(基于视频时长)
    completion_prob=1/(1+video_duration[v]/30)
    # 互动率预估(基于视频质量)
    interact_prob=video_quality[v]*0.3
    # 时长预估
    expected_duration=video_duration[v]*completion_prob*0.6
    # 时效性
    freshness=1/(1+video_freshness[v]/12)
    # 多目标融合
    score=0.4*completion_prob+0.25*interact_prob+0.15*expected_duration/60+0.2*freshness
    multi_scores[v]=score
ranked=sorted(multi_scores.items(),key=lambda x:-x[1])
print(f"多目标推荐Top-5:")
for v,s in ranked[:5]:
    print(f"  视频{v}:综合={s:.3f} 完播率预估={1/(1+video_duration[v]/30):.3f} 时长={video_duration[v]}s 分类={video_category[v]}")
# EE策略
print("\n--- EE策略 ---")
# 新视频用Thompson Sampling
new_videos=np.where(video_freshness<6)[0]
print(f"新视频数:{len(new_videos)}")
alpha_ts=np.ones(n_videos);beta_ts=np.ones(n_videos)
for u in range(n_users):
    for b in behaviors[u]:
        vid=b[0]
        if b[1]>0.5:alpha_ts[vid]+=1
        else:beta_ts[vid]+=1
ts_scores=np.array([np.random.beta(alpha_ts[v],beta_ts[v])for v in range(n_videos)])
ts_rec=np.argsort(-ts_scores)[:5]
print(f"Thompson推荐:{[f'视频{v}'for v in ts_rec]}")
# 完播率分析
print("\n--- 完播率分析 ---")
for cat in ["搞笑","知识","生活","游戏","音乐"]:
    cat_videos=np.where(video_category==cat)[0]
    if len(cat_videos)==0:continue
    cat_completions=[]
    for u in range(n_users):
        for vid,cr,_,_,_,_ in behaviors[u]:
            if video_category[vid]==cat:cat_completions.append(cr)
    if cat_completions:
        print(f"  {cat}:平均完播率={np.mean(cat_completions):.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