跳到主要内容

设计推荐系统

问题​

如何用 Python 设计一个推荐系统？协同过滤和基于内容的推荐有什么区别？

答案​

架构​

协同过滤​

collaborative_filtering.py
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

class UserBasedCF:
    """基于用户的协同过滤"""
    def __init__(self, ratings: np.ndarray):
        # ratings: user x item 评分矩阵
        self.ratings = ratings
        # 计算用户相似度矩阵
        self.user_sim = cosine_similarity(ratings)

    def recommend(self, user_id: int, top_k: int = 10) -> list[int]:
        # 找出最相似的用户
        sim_scores = self.user_sim[user_id]
        similar_users = np.argsort(sim_scores)[::-1][1:50]  # 排除自己

        # 加权预测评分
        user_rated = set(np.where(self.ratings[user_id] > 0)[0])
        scores = {}

        for item in range(self.ratings.shape[1]):
            if item in user_rated:
                continue  # 跳过已评分的
            weighted_sum = sum(
                sim_scores[u] * self.ratings[u, item]
                for u in similar_users if self.ratings[u, item] > 0
            )
            scores[item] = weighted_sum

        return sorted(scores, key=scores.get, reverse=True)[:top_k]


class ItemBasedCF:
    """基于物品的协同过滤"""
    def __init__(self, ratings: np.ndarray):
        self.ratings = ratings
        # 计算物品相似度（转置后按列计算）
        self.item_sim = cosine_similarity(ratings.T)

    def recommend(self, user_id: int, top_k: int = 10) -> list[int]:
        user_ratings = self.ratings[user_id]
        rated_items = np.where(user_ratings > 0)[0]

        scores = {}
        for item in range(self.ratings.shape[1]):
            if user_ratings[item] > 0:
                continue
            score = sum(
                self.item_sim[item, j] * user_ratings[j] for j in rated_items
            )
            scores[item] = score

        return sorted(scores, key=scores.get, reverse=True)[:top_k]

基于内容推荐​

content_based.py
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity

class ContentBasedRecommender:
    def __init__(self, items: list[dict]):
        self.items = items
        # TF-IDF 提取文本特征
        corpus = [f"{item['title']} {item['tags']}" for item in items]
        self.tfidf = TfidfVectorizer(max_features=5000)
        self.item_vectors = self.tfidf.fit_transform(corpus)

    def recommend(self, item_id: int, top_k: int = 10) -> list[int]:
        """根据物品找相似物品"""
        target = self.item_vectors[item_id]
        similarities = cosine_similarity(target, self.item_vectors).flatten()
        # 排除自身
        similar_indices = np.argsort(similarities)[::-1][1:top_k + 1]
        return similar_indices.tolist()

深度学习排序​

ranking_model.py
import torch
import torch.nn as nn

class DeepFM(nn.Module):
    """DeepFM 排序模型"""
    def __init__(self, field_dims: list[int], embed_dim: int = 8):
        super().__init__()
        self.embedding = nn.ModuleList([
            nn.Embedding(dim, embed_dim) for dim in field_dims
        ])
        # FM 一阶
        self.linear = nn.ModuleList([
            nn.Embedding(dim, 1) for dim in field_dims
        ])
        # DNN 部分
        total_dim = len(field_dims) * embed_dim
        self.dnn = nn.Sequential(
            nn.Linear(total_dim, 128), nn.ReLU(), nn.Dropout(0.3),
            nn.Linear(128, 64), nn.ReLU(), nn.Dropout(0.3),
            nn.Linear(64, 1),
        )

    def forward(self, x):
        # x: (batch, num_fields) 各字段索引
        embeds = [self.embedding[i](x[:, i]) for i in range(x.shape[1])]
        embed_concat = torch.cat(embeds, dim=-1)  # DNN 输入
        linear_out = sum(self.linear[i](x[:, i]).squeeze() for i in range(x.shape[1]))

        # FM 交叉
        embed_stack = torch.stack(embeds, dim=1)  # (batch, fields, dim)
        sum_square = embed_stack.sum(dim=1).pow(2).sum(dim=-1)
        square_sum = embed_stack.pow(2).sum(dim=1).sum(dim=-1)
        fm_out = 0.5 * (sum_square - square_sum)

        dnn_out = self.dnn(embed_concat).squeeze()
        return torch.sigmoid(linear_out + fm_out + dnn_out)

常见面试问题​

Q1: User-Based CF vs Item-Based CF？​

答案：

维度User-BasedItem-Based
思路相似用户喜欢的喜欢物品的相似物品
实时性差（用户行为多变）好（物品关系稳定）
适用场景用户少、物品多物品少、用户多（电商）
冷启动新用户无法推荐新物品无法推荐

Q2: 如何解决冷启动问题？​

答案：

  • 新用户：热门推荐 → 引导兴趣标签 → 基于内容推荐
  • 新物品：基于内容特征找相似已有物品
  • 系统冷启动：初期用基于规则/编辑推荐，积累数据后再切换

Q3: 召回和排序的区别？​

答案：

  • 召回：从全量候选（百万级）中快速筛出千级候选集，可以多路召回
  • 排序：用复杂模型对召回结果精排，输出最终推荐列表
  • 重排：业务规则调整（去重、多样性、运营干预）

相关链接​