Skip to content

Instantly share code, notes, and snippets.

@yoshi0309
Last active February 22, 2020 08:04
Show Gist options
  • Save yoshi0309/33bd912d91c0bb5cdf30 to your computer and use it in GitHub Desktop.
Save yoshi0309/33bd912d91c0bb5cdf30 to your computer and use it in GitHub Desktop.
#!/usr/bin/env python
# -*- coding: utf-8 -*-
import sys
import itertools
import csv
import datetime
import time
from math import sqrt
from operator import add
from os.path import join, isfile, dirname
from pyspark import SparkConf, SparkContext
from pyspark.mllib.recommendation import ALS
def parseRating(line):
"""
rating.csv のレコード行からデータを抽出し、タプルにして返す
user_id, timestamp, restaurant_id, total(rating)
"""
fields = line.strip().split(",")
try:
date = datetime.datetime.strptime(fields[11],'"%Y-%m-%d %H:%M:%S"')
epoch = time.mktime(date.timetuple())%10
except ValueError:
epoch = 0
return fields[2], str(epoch)+","+fields[1]+","+fields[3]
def parseRestaurant(line):
"""
restaurants.csv のレコード行からデータを抽出し、タプルにして返す。
restaurant_id, name
"""
fields = line.strip().split(",")
return int(fields[0]), fields[1]
def parseUserids(line):
"""
"""
fields = line.split(",")
return fields[0],int(fields[1])
def getUserId(line):
"""
ratings.csv から userid を抽出して返す
"""
fields = line.strip().split(",")
for field in fields:
if field.startswith("\""):
field = field[1:]
if field.endswith("\""):
field = field[:-1]
return fields[2]
def reformat(x):
id = x[1] + 1 #zipWithUniqueId は 0 から始まるので、1ずつずらす。
return str(x[0]),int(id)
def parseJoinedData(x):
"""
return user_id, restaurant_id, total, timestamp
total (rating) の値は、0から始まるので +1 する。
"""
fields = x[0].split(",")
return x[1],int(fields[1]),float(fields[2])+1,float(fields[0])
def computeRmse(model, data, n):
"""
Compute RMSE (Root Mean Squared Error).
"""
# print "Start prediction.."
# print "Data count : %i " % data.count()
predictions = model.predictAll(data.map(lambda x: (x[0], x[1])))
# print "Prediction done. count: %i . And join data." % predictions.count()
predictionsAndRatings = predictions.map(lambda x: ((x[0], x[1]), x[2])) \
.join(data.map(lambda x: ((x[0], x[1]), x[2]))) \
.values()
# print "Start calculate RMSE : %i " % predictionsAndRatings.count()
return sqrt(predictionsAndRatings.map(lambda x: (x[0] - x[1]) ** 2).reduce(add) / float(n))
if __name__ == "__main__":
# set up environment
conf = SparkConf().setAppName("ldgourmentALS").set("spark.executor.memory", "1g") \
.set("spark.python.worker.memory","3g")
sc = SparkContext(conf=conf)
# load personal ratings
# userid, itemid, rating
myRatings = [
(0,169,5.0), #天一
(0,3127,5.0), #一蘭
(0,333,5.0), #二郎
# (0,142,2.0) # 香港ガーデン
]
myRatingsRDD = sc.parallelize(myRatings, 1)
# 1. ratings.csv を読み込む
# 2. 1行目(idから始まるレコード)をフィルター
# 3. ratings.csvの行から userid / itemid / raitng のみを抽出しタプルに変換(parseRating)
# 4. rate が 0 のものを除外。
homeDir = sys.argv[1]
# ratings = sc.textFile(join(homeDir, "ratings.csv")).filter(lambda x: not(x.startswith("id")) ).map(parseRating).filter(lambda x : not(x[1].endswith("0")))
ratings = sc.textFile(join(homeDir, "ratings.csv")).filter(lambda x: not(x.startswith("id")) ).map(parseRating)
# restaurantの名前を取り出す。
# 1. raestaurants.csv を読み込む
# 2. id から始まるレコードをフィルター(1行目)
# 3. restaurants.csv の各行から restaurantid / name のみを抽出しタプルに変換(parseRestaurant)
restaurantsRdd = sc.textFile(join(homeDir, "restaurants.csv")).filter(lambda x: not(x.startswith("id")) ).map(parseRestaurant)
restaurants = dict(restaurantsRdd.collect())
# ueridから連番に置き換える
# (元のuseridは半角英数のため,MLLibではそのまま扱えない。
# useridに対しintergerの連番をふる。
# userid / userid_int
userids = sc.textFile(join(homeDir, "ratings.csv")).filter(lambda x : not(x.startswith("id"))).map(getUserId).distinct().zipWithUniqueId().map(reformat)
# print "DEBUG! print ratings before join!"
# print ratings.take(5)
# print "DEBUG! print userids before join!"
# print userids.take(5)
# raings の userid を userid_int に置き換える
# 1. ratings に userids を join する (userid, (itemid+rating, userid_int))
# 2. values でvalue部分だけを取り出し itemid+rating, userid_int
# 3. タプルの順番を入れ替える userid_int, itemid, rating, timestamp (parseJoinedData)
# ratings = ratings.join(userids).values().map(parseJoinedData).union(myRatingsRDD)
ratings = ratings.join(userids).values().map(parseJoinedData)
# print "DEBUG ! print ratings after join."
# print ratings.take(5)
# 各データの件数を数えて表示。
numRatings = ratings.count()
numUsers = ratings.map(lambda r:r[0]).distinct().count()
numRestaurants = ratings.map(lambda r:r[1]).distinct().count()
print "Got %d ratings from %d users on %d restaurants." % (numRatings,numUsers,numRestaurants)
# rating からテスト用と検証用のデータをサンプリング
# 学習用のデータは全件使用
numPartitions = 9
training = ratings.filter(lambda x:x[3]<6)\
.map(lambda x:(x[0],x[1],x[2])).union(myRatingsRDD).repartition(numPartitions).cache()
validation = ratings.filter(lambda x:x[3]>=6 and x[3]<8)\
.map(lambda x:(x[0],x[1],x[2])).repartition(numPartitions).cache()
test = ratings.filter(lambda x:x[3]>=8)\
.map(lambda x:(x[0],x[1],x[2])).repartition(numPartitions).cache()
# 各データの件数を数えて表示。
numTraining = training.count()
numValidation = validation.count()
numTest = test.count()
print "Training: %d, validation: %d, test: %d" % (numTraining,numValidation,numTest)
print training.take(5)
# 学習用各種パラメータ
# 配列なので、複数セット可。複数セットすると最も RMSE の値が良いものを採用する。
ranks = [8,9,10]
lambdas = [0.31,0.32,0.33]
numIters = [3]
# 一時変数
bestValidationRmse = float("inf")
bestModel = None
bestRank = 0
bestLambda = -1.0
bestNumIter = -1
# DEBUG
# print "DEBUG ! #############################################################################"
# print training.take(5)
for rank,lmbda,numIter in itertools.product(ranks,lambdas,numIters):
# 学習の実行
model = ALS.train(training,rank,numIter,lmbda)
# 検証用のデータを使用して RMSE を計算し表示
validationRmse = computeRmse(model, validation, numValidation)
print "RMSE (validation) = %f for the model trained with " % validationRmse + "rank = %d, lambda = %.2f, and numIter %d." % (rank,lmbda,numIter)
# RMSE の値が最もよければ、学習用のパラメータを今のものと置き換える。
if(validationRmse < bestValidationRmse):
bestModel = model
bestRank = rank
bestLambda = lmbda
bestNumIter = numIter
bestValidationRmse = validationRmse
# テスト用データで bestmodel を使い RMSE を計算して表示。
testRmse = computeRmse(bestModel, test, numTest)
print "The best model was trained with rank = %d and lambda = %.2f, " % (bestRank, bestLambda) \
+ "and numIter = %d, and its RMSE on the test set is %f." % (bestNumIter, testRmse)
# rate を推測する。
# レストランの全データ(restaurants)から、まだ訪れたことのない restaurantid を抽出
myRatedRestaurantIds = set([x[1] for x in myRatings])
candidates = sc.parallelize([m for m in restaurants if m not in myRatedRestaurantIds]).repartition(numPartitions).cache()
# 1. userid 0 に対して、訪れたことのないレストランの rating を全て推測
# 2. rate の大きかった順にソートして、Top 50 件を抽出
# 3. 抽出した 50 件を表示
predictions = bestModel.predictAll(candidates.map(lambda x : (0,x))).collect()
recommendations = sorted(predictions, key=lambda x : x[2], reverse=True)[:50]
for i in xrange(len(recommendations)):
print ("%2d: %s, Score: %0.2f , User Id: %s" % (i + 1, restaurants[recommendations[i][1]].encode('utf-8', 'ignore'), recommendations[i][2], recommendations[i][0]))
# clean up
sc.stop()
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment