阿里手淘猜你喜欢Swing算法介绍

Swing算法原理比较简单，是阿里早期使用到的一种召回算法，在阿里多个业务被验证过非常有效的一种召回方式，它认为 user-item-user 的结构比 itemCF 的单边结构更稳定。

Swing指的是秋千，例如用户 u uu 和用户 v vv ，都购买过同一件商品i ii，则三者之间会构成一个类似秋千的关系图。若用户 u 和用户 v 之间除了购买过 i 外，还购买过商品 j ，则认为两件商品是具有某种程度上的相似的。

也就是说，商品与商品之间的相似关系，是通过用户关系来传递的。为了衡量物品 i 和 j j的相似性，考察都购买了物品 i 和 j 的用户 u和用户 v ， 如果这两个用户共同购买的物品越少，则物品 i和 j 的相似性越高。

Swing算法的表达式如下：

swing python 实现：

# -*- coding: utf-8 -*-

"""
    Author  : Thinkgamer
    File    : Swing.py
    Software: PyCharm
    Desc    : 基于movie lens数据集实现Swing算法
"""

import pandas as pd
from itertools import combinations
import json
import os

alpha = 0.5
top_k = 20

def load_data(train_path, test_path):
    train_data = pd.read_csv(train_path, sep="\t", engine="python", names=["userid", "movieid", "rate", "event_timestamp"])
    test_data = pd.read_csv(test_path, sep="\t", engine="python", names=["userid", "movieid", "rate", "event_timestamp"])

    print(train_data.head(5))
    print(test_data.head(5))
    return train_data, test_data

def get_uitems_iusers(train):
    u_items = dict()
    i_users = dict()
    for index, row in train.iterrows():
        u_items.setdefault(row["userid"], set())
        i_users.setdefault(row["movieid"], set())

        u_items[row["userid"]].add(row["movieid"])
        i_users[row["movieid"]].add(row["userid"])
    print("使用的用户个数为：{}".format(len(u_items)))
    print("使用的item个数为：{}".format(len(i_users)))
    return u_items, i_users

def cal_similarity(u_items, i_users):
    item_pairs = list(combinations(i_users.keys(), 2))
    print("item pairs length：{}".format(len(item_pairs))) # 1410360
    item_sim_dict = dict()
    cnt = 0
    for (i, j) in item_pairs:
        cnt += 1
        print(cnt)
        user_pairs = list(combinations(i_users[i] & i_users[j], 2))
        result = 0.0
        for (u, v) in user_pairs:
            result += 1 / (alpha + list(u_items[u] & u_items[v]).__len__())

        item_sim_dict.setdefault(i, dict())
        item_sim_dict[i][j] = result
        # print(item_sim_dict[i][j])

    return item_sim_dict

def save_item_sims(item_sim_dict, path):
    new_item_sim_dict = dict()
    for item, sim_items in item_sim_dict.items():
        new_item_sim_dict.setdefault(item, dict())
        new_item_sim_dict[item] = dict(sorted(sim_items.items(), key = lambda k:k[1], reverse=True)[:top_k])
    json.dump(new_item_sim_dict, open(path, "w"))
    print("item 相似 item（{}）保存成功！".format(top_k))
    return new_item_sim_dict

def evaluate(item_sim_dict, test):
    # 可以参考《推荐系统开发实战》中的cf验证方式

    pass

if __name__ == "__main__":
    train_data_path = "../../data/ml-100k/ua.base"
    test_data_path = "../../data/ml-100k/ua.test"
    item_sim_save_path = "../../model/swing/item_sim_dict.json"

    train, test = load_data(train_data_path, test_data_path)
    if not os.path.exists(item_sim_save_path):
        u_items, i_users = get_uitems_iusers(train)
        item_sim_dict = cal_similarity(u_items, i_users)

        new_item_sim_dict = save_item_sims(item_sim_dict, item_sim_save_path)
    else:
        new_item_sim_dict = json.load(open(item_sim_save_path, "r"))

    evaluate(new_item_sim_dict, test)

————————————————

.Swing Spark实现
创建Swing类，其中的评估函数和predict函数这里并未提供，感兴趣的可以自己实现

/**
  * @ClassName: Swing
  * @Description: 实现Swing算法
  * @author: Thinkgamer
  **/

class SwingModel(spark: SparkSession) extends Serializable{
    var alpha: Option[Double] = Option(0.0)
    var items: Option[ArrayBuffer[String]] = Option(new ArrayBuffer[String]())
    var userIntersectionMap: Option[Map[String, Map[String, Int]]] = Option(Map[String, Map[String, Int]]())

    /*
     * @Description 给参数 alpha赋值
     * @Param double
     * @return cf.SwingModel
     **/
    def setAlpha(alpha: Double): SwingModel = {
        this.alpha = Option(alpha)
        this
    }

    /*
     * @Description 给所有的item进行赋值
     * @Param [array]
     * @return cf.SwingModel
     **/
    def setAllItems(array: Array[String]): SwingModel = {
        this.items = Option(array.toBuffer.asInstanceOf[ArrayBuffer[String]])
        this
    }

    /*
     * @Description 获取两两用户有行为的item交集个数
     * @Param [spark, data]
     * @return scala.collection.immutable.Map<java.lang.String,scala.collection.immutable.Map<java.lang.String,java.lang.Object>>
     **/
    def calUserRateItemIntersection(data: RDD[(String, String, Double)]): Map[String, Map[String, Int]] = {
        val rdd = data.map(l => (l._1, l._2)).groupByKey().map(l => (l._1, l._2.toSet))
        val map = (rdd cartesian rdd).map(l => (l._1._1, (l._2._1, (l._1._2 & l._2._2).toArray.length)))
            .groupByKey()
            .map(l => (l._1, l._2.toMap))
            .collectAsMap().toMap
        map.take(10).foreach(println)
        map
    }

    def fit(data: RDD[(String, String, Double)]): RDD[(String, String, Double)]= {
        this.userIntersectionMap = Option(this.calUserRateItemIntersection(data))
        println(this.userIntersectionMap.take(10))

        val rdd = data.map(l => (l._2, l._1)).groupByKey().map(l => (l._1, l._2.toSet))
        val result: RDD[(String, String, Double)] = (rdd cartesian rdd).map(l => {
            val item1 = l._1._1
            val item2 = l._2._1
            val intersectionUsers = l._1._2 & l._2._2
            var score = 0.0
            for(u1 <- intersectionUsers){
                for(u2 <- intersectionUsers){
                    score += 1.0 / (this.userIntersectionMap.get.get(u1).get(u2).toDouble + this.alpha.get)
                }
            }
            (item1, item2, score) // (item1, item2, swingsocre)
        })
        result
    }

    def evalute(test: RDD[(String, String, Double)]) = { }

    def predict(userid: String) = { }

    def predict(userids: Array[String]) = { }
    
}

main函数调用

object Swing {
    def main(args: Array[String]): Unit = {
        val spark = SparkSession.builder().master("local[10]").appName("Swing").enableHiveSupport().getOrCreate()
        Logger.getRootLogger.setLevel(Level.WARN)

        val trainDataPath = "data/ml-100k/ua.base"
        val testDataPath = "data/ml-100k/ua.test"

      	import spark.sqlContext.implicits._
        val train: RDD[(String, String, Double)] = spark.sparkContext.textFile(trainDataPath).map(_.split("\t")).map(l => (l(0), l(1), l(2).toDouble))
        val test: RDD[(String, String, Double)] = spark.sparkContext.textFile(testDataPath).map(_.split("\t")).map(l => (l(0), l(1), l(2).toDouble))

        val items: Array[String] = train.map(_._2).collect()

        val swing = new SwingModel(spark).setAlpha(1).setAllItems(items)
        val itemSims: RDD[(String, String, Double)] = swing.fit(train)

        swing.evalute(test)
        swing.predict("")
        swing.predict(Array("", ""))

        spark.close()
    }
}

转载
https://blog.csdn.net/Gamer_gyt/article/details/115678598