在這個系列的文章裡，我們將使用 Apache Spark、XGBoost、Elasticsearch 和 MySQL 等工具來搭建一個推薦系統的 Machine Learning Pipeline。推薦系統的組成可以粗略地分成 Candidate Generation 和 Ranking 兩個部分，前者是針對用戶產生候選物品集，常用的方法有 Collaborative Filtering、Content-based、標籤配對、熱門排行或人工精選等；後者則是對這些候選物品排序，以 Top N 的方式呈現最終的推薦結果，常用的方法有 Logistic Regression。
在本篇文章中，我們將以 Ranking 階段常用的方法之一：Logistic Regression 邏輯迴歸為例，利用 Apache Spark 的 Logistic Regression 模型建立一個 GitHub repositories 的推薦系統，以用戶對 repo 的打星紀錄和用戶與 repo 的各項屬性做為特徵，預測出用戶會不會打星某個 repo（分類問題）。最後訓練出來的模型就可以做為我們的推薦系統的 Ranking 模組。不過因為 LR 是線性模型，所以通常需要大量的 Feature Engineering 來習得非線性關係。所以這篇文章的重點是 Spark ML 的 Pipeline 機制和特徵工程，不會在演算法的部分著墨太多。
完整的程式碼可以在 https://github.com/vinta/albedo 找到。
系列文章：
- Build a recommender system with Spark: Implicit ALS
- Build a recommender system with Spark: Content-based and Elasticsearch
- Build a recommender system with Spark: Logistic Regression
- Feature Engineering 特徵工程中常見的方法
- Spark ML cookbook (Scala)
- Spark SQL cookbook (Scala)
- 不定期更新中
Submit the Application
$ spark-submit \
--master spark://192.168.10.100:7077 \
--packages "com.github.fommil.netlib:all:1.1.2,com.hankcs:hanlp:portable-1.3.4,mysql:mysql-connector-java:5.1.41" \
--class ws.vinta.albedo.UserProfileBuilder \
target/albedo-1.0.0-SNAPSHOT.jar
$ spark-submit \
--master spark://192.168.10.100:7077 \
--packages "com.github.fommil.netlib:all:1.1.2,com.hankcs:hanlp:portable-1.3.4,mysql:mysql-connector-java:5.1.41" \
--class ws.vinta.albedo.RepoProfileBuilder \
target/albedo-1.0.0-SNAPSHOT.jar
$ spark-submit \
--master spark://192.168.10.100:7077 \
--packages "com.github.fommil.netlib:all:1.1.2,com.hankcs:hanlp:portable-1.3.4,mysql:mysql-connector-java:5.1.41" \
--class ws.vinta.albedo.LogisticRegressionRanker \
target/albedo-1.0.0-SNAPSHOT.jarref:
https://vinta.ws/code/setup-spark-scala-and-maven-with-intellij-idea.html
https://spark.apache.org/docs/latest/submitting-applications.html
https://spoddutur.github.io/spark-notes/distribution_of_executors_cores_and_memory_for_spark_application
Load Data
我們之前已經利用 GitHub API 和 BigQuery 上的 GitHub Archive 收集了 150 萬筆的打星紀錄和所屬的用戶、repo 數據。目前有以下幾個數據集，大致上是照著 GitHub API 建立的，欄位分別如下：
rawUserInfoDS.printSchema()
// root
// |-- user_id: integer (nullable = true)
// |-- user_login: string (nullable = true)
// |-- user_account_type: string (nullable = true)
// |-- user_name: string (nullable = true)
// |-- user_company: string (nullable = true)
// |-- user_blog: string (nullable = true)
// |-- user_location: string (nullable = true)
// |-- user_email: string (nullable = true)
// |-- user_bio: string (nullable = true)
// |-- user_public_repos_count: integer (nullable = true)
// |-- user_public_gists_count: integer (nullable = true)
// |-- user_followers_count: integer (nullable = true)
// |-- user_following_count: integer (nullable = true)
// |-- user_created_at: timestamp (nullable = true)
// |-- user_updated_at: timestamp (nullable = true)
rawRepoInfoDS.printSchema()
// |-- repo_id: integer (nullable = true)
// |-- repo_owner_id: integer (nullable = true)
// |-- repo_owner_username: string (nullable = true)
// |-- repo_owner_type: string (nullable = true)
// |-- repo_name: string (nullable = true)
// |-- repo_full_name: string (nullable = true)
// |-- repo_description: string (nullable = true)
// |-- repo_language: string (nullable = true)
// |-- repo_created_at: timestamp (nullable = true)
// |-- repo_updated_at: timestamp (nullable = true)
// |-- repo_pushed_at: timestamp (nullable = true)
// |-- repo_homepage: string (nullable = true)
// |-- repo_size: integer (nullable = true)
// |-- repo_stargazers_count: integer (nullable = true)
// |-- repo_forks_count: integer (nullable = true)
// |-- repo_subscribers_count: integer (nullable = true)
// |-- repo_is_fork: boolean (nullable = true)
// |-- repo_has_issues: boolean (nullable = true)
// |-- repo_has_projects: boolean (nullable = true)
// |-- repo_has_downloads: boolean (nullable = true)
// |-- repo_has_wiki: boolean (nullable = true)
// |-- repo_has_pages: boolean (nullable = true)
// |-- repo_open_issues_count: integer (nullable = true)
// |-- repo_topics: string (nullable = true)
rawStarringDS.printSchema()
// root
// |-- user_id: integer (nullable = true)
// |-- repo_id: integer (nullable = true)
// |-- starred_at: timestamp (nullable = true)
// |-- starring: double (nullable = true)ref:
https://www.githubarchive.org/
http://ghtorrent.org/
載入資料之後，要做的第一件事應該就是 Exploratory Data Analysis (EDA) 了，把玩一下手上的數據。建議大家可以試試 Apache Zeppelin 或是 Databricks 的 Notebook，除了內建 Spark 支援的所有語言之外，也整合了 NoSQL 和 JDBC 支援的資料庫，要畫圖表也很方便，簡直比 Jupyter Notebook 還好用了。
ref:
https://zeppelin.apache.org/
https://databricks.com/
Build User Profile / Item Profile
在這個專案中，最主要的數據主體就是 user 和 repo，所以我們針對兩者各自建立 User Profile 和 Item Profile，作為之後在模型訓練階段會使用的特徵。我們把這個步驟跟模型訓練的流程分開，這樣對整個架構的搭建會更有彈性。實務上，我們可以用 user id 或 item id 當 key，直接把製作好的特徵存進 Redis 或其他 schemaless 的 NoSQL 資料庫，方便之後給多個模型取用；在做 real-time 推薦時，也可以很快地拿到特徵，只需要重新計算部份欄位即可。
不過因為這裡主要用的是來自 GitHub API 的資料，某種程度上人家已經幫我們做了很多資料清理和正規化的動作了，但是在現實中，你的系統要處理的數據通常不會這麼乾淨，可能來自各種 data source、有著各種格式，還會隨著時間而改變，通常得花上不少力氣做 Extract, Transform, Load (ETL)，所以最好在寫 log（埋點）的時候就溝通好。而且在 production 環境中，數據是會一直變動的，要確保數據的時效性和容錯性，很重要的一個部分就是 monitoring。
ref:
http://www.algorithmdog.com/ad-rec-deploy
https://tech.meituan.com/online-feature-system02.html
礙於篇幅有限，以下的文章中我們只會挑幾個重要的部分說明。簡單說，在這個步驟的最後，我們會得到 userProfileDF 和 repoProfileDF 這兩個 DataFrame，分別存放製作好的特徵。詳細的程式碼如下：
ref:
https://github.com/vinta/albedo/blob/master/src/main/scala/ws/vinta/albedo/UserProfileBuilder.scala
https://github.com/vinta/albedo/blob/master/src/main/scala/ws/vinta/albedo/RepoProfileBuilder.scala
Feature Engineering
以推薦系統為例，特徵可以分成以下四種：
- 用戶特徵：用戶本身的各種屬性，例如 user id、性別、職業或所在的城市等
- 物品特徵：物品本身的各種屬性，例如 item id、作者、標題、分類、評分或所屬的標籤等
- 交互特徵：用戶對物品做出的某項行為，該行為的 aggregation 或交叉特徵，例如是否看過同類型的電影、最近聽的歌曲的曲風分佈或上週買了多少高單價的商品
- 上下文特徵：用戶對物品做出的某項行為，該行為的 metadata，例如發生的時間、使用的裝置或當前的 GPS 位置等
有些特徵是在資料採集階段就能拿到，有些特徵則會需要額外的步驟（例如透過外部的 API 或是其他模型）才能取得，也有些特徵必須即時更新。順道一提，因為我們要預測的是「某個用戶會不會打星某個 repo」，所以下述特徵裡的 user 可以是 repo stargazer 也可以是 repo owner。
原始特徵：
- 用戶特徵
user_iduser_loginuser_nameuser_emailuser_bloguser_biouser_companyuser_locationuser_followers_counguser_following_countuser_public_repos_countuser_public_gists_countuser_created_atuser_updated_at
- 物品特徵
repo_idrepo_namerepo_ownerrepo_owner_typerepo_languagerepo_descriptionrepo_homepagerepo_subscribers_countrepo_stargazers_countrepo_forks_countrepo_sizerepo_created_atrepo_updated_atrepo_pushed_atrepo_has_issuesrepo_has_projectsrepo_has_downloadsrepo_has_wikirepo_has_pagesrepo_open_issues_countrepo_topics
- 交互特徵
user_stars_repouser_follows_user
- 上下文特徵
user_repo_starred_at
發想特徵：
- 用戶特徵
user_days_between_created_at_today: 該用戶的註冊日期距離今天過了幾年user_days_between_updated_at_today: 該用戶的更新日期距離今天過了幾天user_repos_avg_stargazers_count: 該用戶名下的所有 repo（不含 fork 的）的平均星星數user_organizations: 該用戶屬於哪些組織user_has_null: 該用戶至少有一個欄位是 nulluser_has_blog: 該用戶有沒有網站user_is_freelancer: 該用戶的 bio 中是否包含 Freelancer 等字眼user_is_junior: 該用戶的 bio 中是否包含 Beginner 或 Junior 等字眼user_is_lead: 該用戶的 bio 中是否包含 Team Lead、Architect、Creator、CTO 或 VP of Engineering 等字眼user_is_scholar: 該用戶的 bio 中是否包含 Researcher、Scientist、PhD 或 Professor 等字眼user_is_pm: 該用戶的 bio 中是否包含 Product Manager 等字眼user_knows_backend: 該用戶的 bio 中是否包含 Backend 或 Back end 等字眼user_knows_data: 該用戶的 bio 中是否包含 Machine Learning、Deep Learning 或 Data Science 等字眼user_knows_devops: 該用戶的 bio 中是否包含 DevOps、SRE、SysAdmin 或 Infrastructure 等字眼user_knows_frontend: 該用戶的 bio 中是否包含 Frontend 或 Front end 等字眼user_knows_mobile: 該用戶的 bio 中是否包含 Mobile、iOS 或 Android 等字眼user_knows_recsys: 該用戶的 bio 中是否包含 Recommender System、Data Mining 或 Information Retrieval 等字眼user_knows_web: 該用戶的 bio 中是否包含 Web Development 或 Fullstack 等字眼
- 物品特徵
repo_created_at_days_since_today: 該 repo 的建立日期距離今天過了幾天repo_updated_at_days_since_today: 該 repo 的更新日期距離今天過了幾天repo_pushed_at_days_since_today: 該 repo 的提交日期距離今天過了幾天repo_stargazers_count_in_30days: 該 repo 在 30 天內收到的星星數repo_subscribers_stargazers_ratio: 該 repo 的 watch 數和 star 數的比例repo_forks_stargazers_ratio: 該 repo 的 fork 數和 star 數的比例repo_open_issues_stargazers_ratio: 該 repo 的 數和 star 數的比例repo_releases_count: 該 repo 的 release 或 tag 數repo_lisence: 該 repo 的授權條款repo_readme: 該 repo 的 README 內容repo_has_null: 該 repo 有至少一個欄位是 nullrepo_has_readme: 該 repo 是否有 README 檔案repo_has_changelog: 該 repo 是否有 CHANGELOG 檔案repo_has_contributing: 該 repo 是否有 CONTRIBUTING 檔案repo_has_tests: 該 repo 是否有測試repo_has_ci: 該 repo 是否有 CIrepo_has_dockerfile: 該 repo 是否有 Dockerfilerepo_is_unmaintained: 該 repo 是否不再維護了repo_is_awesome: 該 repo 是否被收錄進任何的 awesome-xxx 列表裡repo_is_vinta_starred: 該 repo 是否被 @vinta aka 本文的作者打星了
- 交互特徵
user_starred_repos_count: 該用戶總共打星了多少 repouser_avg_daily_starred_repos_count: 該用戶平均每天打星多少 repouser_forked_repos_count: 該用戶總共 fork 了多少 repouser_follower_following_count_ratio: 該用戶的 follower 數和 following 數的比例user_recent_searched_keywords: 該用戶最近搜尋的 50 個關鍵字user_recent_commented_repos: 該用戶最近留言的 50 個 repouser_recent_watched_repos: 該用戶最近訂閱的 50 個 repouser_recent_starred_repos_descriptions: 該用戶最近打星的 50 個 repo 的描述user_recent_starred_repos_languages: 該用戶最近打星的 50 個 repo 的語言user_recent_starred_repos_topics: 該用戶最近打星的 50 個 repo 的標籤user_follows_repo_owner: 該用戶是否追蹤該 repo 的作者repo_language_index_in_user_recent_repo_languages: 該 repo 的語言出現在該用戶最近打星的語言列表的順序repo_language_count_in_user_recent_repo_languages: 該 repo 的語言出現在該用戶最近打星的語言列表的次數repo_topics_user_recent_topics_similarity: 該 repo 的標籤與該用戶最近打星的標籤列表的相似度
- 上下文特徵
als_model_prediction: 來自 ALS 模型的預測值，該用戶對該 repo 的偏好程度gbdt_model_index: 來自 GBDT 模型的 tree index，該 observation 的自動特徵
Feature Engineering 特徵工程中常見的方法
https://vinta.ws/code/feature-engineering.html
Detect Outliers
除了缺失值之外，離群值（異常值）也是需要注意的地方。如果是 continuous 特徵，用 box plot 可以很快地發現離群值；如果是 categorical 特徵，可以 SELECT COUNT(*) ... GROUP BY 一下，然後用 bar chart 查看每個 category 的數量。取決於你所要解決的問題，異常值可能可以直接忽略，也可能需要特別對待，例如搞清楚異常值出現的原因，是資料採集時的差錯或是某種隱含的深層的因素之類的。
ref:
https://www.analyticsvidhya.com/blog/2016/01/guide-data-exploration/
https://www.slideshare.net/tw_dsconf/123-70852901
Impute Missing Values
可以利用 df.describe().show() 查看各個欄位的統計數據：count、mean、stddev、min 和 max。除了使用 df.where("some_column IS NULL") 之外，比較不同欄位的 count 差異也可以很快地發現哪些欄位有缺失值。順便觀察一下有缺失值的欄位和 target variable 有沒有什麼關聯。
這裡直接對 null 和 NaN 數據填充缺失值，因為以下幾個欄位都是字串類型，所以直接改成空字串，方便後續的處理。然後順便做一個 has_null 的特徵。
針對 user：
import org.apache.spark.sql.functions._
val nullableColumnNames = Array("user_name", "user_company", "user_blog", "user_location", "user_bio")
val imputedUserInfoDF = rawUserInfoDS
.withColumn("user_has_null", when(nullableColumnNames.map(rawUserInfoDS(_).isNull).reduce(_ || _), true).otherwise(false))
.na.fill("", nullableColumnNames)針對 repo：
import org.apache.spark.sql.functions._
val nullableColumnNames = Array("repo_description", "repo_homepage")
val imputedRepoInfoDF = rawRepoInfoDS
.withColumn("repo_has_null", when(nullableColumnNames.map(rawRepoInfoDS(_).isNull).reduce(_ || _), true).otherwise(false))
.na.fill("", nullableColumnNames)ref:
https://spark.apache.org/docs/latest/api/scala/index.html#org.apache.spark.sql.DataFrameNaFunctions
如果是數值類型的欄位可以考慮使用 Imputer。
ref:
https://spark.apache.org/docs/latest/ml-features.html#imputer
Clean Data
針對 user，用 User-defined Function 對幾個文字欄位做一些正規化的處理：
import org.apache.spark.sql.functions._
val cleanUserInfoDF = imputedUserInfoDF
.withColumn("user_clean_company", cleanCompanyUDF($"user_company"))
.withColumn("user_clean_location", cleanLocationUDF($"user_location"))
.withColumn("user_clean_bio", lower($"user_bio"))針對 repo，過濾掉一些 repo_stargazers_count 太多和太少、description 欄位含有 "unmaintained" 或 "assignment" 等字眼的項目：
val reducedRepoInfo = imputedRepoInfoDF
.where($"repo_is_fork" === false)
.where($"repo_forks_count" <= 90000)
.where($"repo_stargazers_count".between(30, 100000))
val unmaintainedWords = Array("%unmaintained%", "%no longer maintained%", "%deprecated%", "%moved to%")
val assignmentWords = Array("%assignment%", "%作業%", "%作业%")
val demoWords = Array("test", "%demo project%")
val blogWords = Array("my blog")
val cleanRepoInfoDF = reducedRepoInfo
.withColumn("repo_clean_description", lower($"repo_description"))
.withColumn("repo_is_unmaintained", when(unmaintainedWords.map($"repo_clean_description".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("repo_is_assignment", when(assignmentWords.map($"repo_clean_description".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("repo_is_demo", when(demoWords.map($"repo_clean_description".like(_)).reduce(_ or _) and $"repo_stargazers_count" <= 40, true).otherwise(false))
.withColumn("repo_is_blog", when(blogWords.map($"repo_clean_description".like(_)).reduce(_ or _) and $"repo_stargazers_count" <= 40, true).otherwise(false))
.where($"repo_is_unmaintained" === false)
.where($"repo_is_assignment" === false)
.where($"repo_is_demo" === false)
.where($"repo_is_blog" === false)
.withColumn("repo_clean_language", lower($"repo_language"))
.withColumn("repo_clean_topics", lower($"repo_topics"))Construct Features
針對 user，根據上述的「發想特徵」，製作出新的特徵：
import org.apache.spark.sql.expressions.Window
import org.apache.spark.sql.functions._
val webThings = Array("web", "fullstack", "full stack")
val backendThings = Array("backend", "back end", "back-end")
val frontendThings = Array("frontend", "front end", "front-end")
val mobileThings = Array("mobile", "ios", "android")
val devopsThings = Array("devops", "sre", "admin", "infrastructure")
val dataThings = Array("machine learning", "deep learning", "data scien", "data analy")
val recsysThings = Array("data mining", "recommend", "information retrieval")
val leadTitles = Array("team lead", "architect", "creator", "director", "cto", "vp of engineering")
val scholarTitles = Array("researcher", "scientist", "phd", "professor")
val freelancerTitles = Array("freelance")
val juniorTitles = Array("junior", "beginner", "newbie")
val pmTitles = Array("product manager")
val userStarredReposCountDF = rawStarringDS
.groupBy($"user_id")
.agg(count("*").alias("user_starred_repos_count"))
val starringRepoInfoDF = rawStarringDS
.select($"user_id", $"repo_id", $"starred_at")
.join(rawRepoInfoDS, Seq("repo_id"))
val userTopLanguagesDF = starringRepoInfoDF
.withColumn("rank", rank.over(Window.partitionBy($"user_id").orderBy($"starred_at".desc)))
.where($"rank" <= 50)
.groupBy($"user_id")
.agg(collect_list(lower($"repo_language")).alias("user_recent_repo_languages"))
.select($"user_id", $"user_recent_repo_languages")
val userTopTopicsDF = starringRepoInfoDF
.where($"repo_topics" =!= "")
.withColumn("rank", rank.over(Window.partitionBy($"user_id").orderBy($"starred_at".desc)))
.where($"rank" <= 50)
.groupBy($"user_id")
.agg(concat_ws(",", collect_list(lower($"repo_topics"))).alias("temp_user_recent_repo_topics"))
.select($"user_id", split($"temp_user_recent_repo_topics", ",").alias("user_recent_repo_topics"))
val userTopDescriptionDF = starringRepoInfoDF
.where($"repo_description" =!= "")
.withColumn("rank", rank.over(Window.partitionBy($"user_id").orderBy($"starred_at".desc)))
.where($"rank" <= 50)
.groupBy($"user_id")
.agg(concat_ws(" ", collect_list(lower($"repo_description"))).alias("user_recent_repo_descriptions"))
.select($"user_id", $"user_recent_repo_descriptions")
val constructedUserInfoDF = cleanUserInfoDF
.withColumn("user_knows_web", when(webThings.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_knows_backend", when(backendThings.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_knows_frontend", when(frontendThings.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_knows_mobile", when(mobileThings.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_knows_devops", when(devopsThings.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_knows_data", when(dataThings.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_knows_recsys", when(recsysThings.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_is_lead", when(leadTitles.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_is_scholar", when(scholarTitles.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_is_freelancer", when(freelancerTitles.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_is_junior", when(juniorTitles.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_is_pm", when(pmTitles.map($"user_clean_bio".like(_)).reduce(_ or _), true).otherwise(false))
.withColumn("user_followers_following_ratio", round($"user_followers_count" / ($"user_following_count" + lit(1.0)), 3))
.withColumn("user_days_between_created_at_today", datediff(current_date(), $"user_created_at"))
.withColumn("user_days_between_updated_at_today", datediff(current_date(), $"user_updated_at"))
.join(userStarredReposCountDF, Seq("user_id"))
.withColumn("user_avg_daily_starred_repos_count", round($"user_starred_repos_count" / ($"user_days_between_created_at_today" + lit(1.0)), 3))
.join(userTopDescriptionDF, Seq("user_id"))
.join(userTopTopicsDF, Seq("user_id"))
.join(userTopLanguagesDF, Seq("user_id"))針對 repo，根據上述的「發想特徵」，製作出新的特徵，意思到了就好：
import org.apache.spark.sql.functions._
val vintaStarredRepos = rawStarringDS
.where($"user_id" === 652070)
.select($"repo_id".as[Int])
.collect()
.to[List]
val constructedRepoInfoDF = cleanRepoInfoDF
.withColumn("repo_has_activities_in_60days", datediff(current_date(), $"repo_pushed_at") <= 60)
.withColumn("repo_has_homepage", when($"repo_homepage" === "", false).otherwise(true))
.withColumn("repo_is_vinta_starred", when($"repo_id".isin(vintaStarredRepos: _*), true).otherwise(false))
.withColumn("repo_days_between_created_at_today", datediff(current_date(), $"repo_created_at"))
.withColumn("repo_days_between_updated_at_today", datediff(current_date(), $"repo_updated_at"))
.withColumn("repo_days_between_pushed_at_today", datediff(current_date(), $"repo_pushed_at"))
.withColumn("repo_subscribers_stargazers_ratio", round($"repo_subscribers_count" / ($"repo_stargazers_count" + lit(1.0)), 3))
.withColumn("repo_forks_stargazers_ratio", round($"repo_forks_count" / ($"repo_stargazers_count" + lit(1.0)), 3))
.withColumn("repo_open_issues_stargazers_ratio", round($"repo_open_issues_count" / ($"repo_stargazers_count" + lit(1.0)), 3))
.withColumn("repo_text", lower(concat_ws(" ", $"repo_owner_username", $"repo_name", $"repo_language", $"repo_description")))ref:
https://databricks.com/blog/2015/09/16/apache-spark-1-5-dataframe-api-highlights.html
Convert Features
針對 user，這裡主要是對一些 categorical 特徵作 binning：
import org.apache.spark.sql.functions._
val companyCountDF = cleanUserInfoDF
.groupBy($"user_clean_company")
.agg(count("*").alias("count_per_user_company"))
val locationCountDF = cleanUserInfoDF
.groupBy($"user_clean_location")
.agg(count("*").alias("count_per_user_location"))
val transformedUserInfoDF = constructedUserInfoDF
.join(companyCountDF, Seq("user_clean_company"))
.join(locationCountDF, Seq("user_clean_location"))
.withColumn("user_has_blog", when($"user_blog" === "", 0.0).otherwise(1.0))
.withColumn("user_binned_company", when($"count_per_user_company" <= 5, "__other").otherwise($"user_clean_company"))
.withColumn("user_binned_location", when($"count_per_user_location" <= 50, "__other").otherwise($"user_clean_location"))針對 repo：
import org.apache.spark.sql.functions._
val languagesDF = cleanRepoInfoDF
.groupBy($"repo_clean_language")
.agg(count("*").alias("count_per_repo_language"))
.select($"repo_clean_language", $"count_per_repo_language")
.cache()
val transformedRepoInfoDF = constructedRepoInfoDF
.join(languagesDF, Seq("repo_clean_language"))
.withColumn("repo_binned_language", when($"count_per_repo_language" <= 30, "__other").otherwise($"repo_clean_language"))
.withColumn("repo_clean_topics", split($"repo_topics", ","))ref:
https://docs.databricks.com/spark/latest/mllib/binary-classification-mllib-pipelines.html
Prepare the Feature Pipeline
我們過濾掉那些打星了超多 repo 的用戶。從收集到的數據發現，有些用戶甚至打星了一兩萬個 repo，這些用戶可能是個爬蟲專用帳號或是他看到什麼就打星什麼，推薦系統對這樣的用戶來說可能沒什麼意義，還不如從數據集中拿掉。
import org.apache.spark.sql.functions._
val maxStarredReposCount = 2000
val userStarredReposCountDF = rawStarringDS
.groupBy($"user_id")
.agg(count("*").alias("user_starred_repos_count"))
val reducedStarringDF = rawStarringDS
.join(userStarredReposCountDF, Seq("user_id"))
.where($"user_starred_repos_count" <= maxStarredReposCount)
.select($"user_id", $"repo_id", $"starred_at", $"starring")
val profileStarringDF = reducedStarringDF
.join(userProfileDF, Seq("user_id"))
.join(repoProfileDF, Seq("repo_id"))Build the Feature Pipeline
把處理特徵的一連串流程寫成 Spark ML Pipeline，方便抽換或是加入新的 Transformer，例如 Standardization、One-hot Encoding 和 Word2Vec，也把 ALS 模型的預測值做為其中一項特徵。
import org.apache.spark.ml.feature._
import org.apache.spark.ml.recommendation.ALSModel
import ws.vinta.albedo.transformers.UserRepoTransformer
val profileStarringDF = reducedStarringDF
.join(userProfileDF, Seq("user_id"))
.join(repoProfileDF, Seq("repo_id"))
.cache()
categoricalColumnNames += "user_id"
categoricalColumnNames += "repo_id"
val userRepoTransformer = new UserRepoTransformer()
.setInputCols(Array("repo_language", "user_recent_repo_languages"))
continuousColumnNames += "repo_language_index_in_user_recent_repo_languages"
continuousColumnNames += "repo_language_count_in_user_recent_repo_languages"
val alsModelPath = s"${settings.dataDir}/${settings.today}/alsModel.parquet"
val alsModel = ALSModel.load(alsModelPath)
.setUserCol("user_id")
.setItemCol("repo_id")
.setPredictionCol("als_score")
.setColdStartStrategy("drop")
continuousColumnNames += "als_score"
val categoricalTransformers = categoricalColumnNames.flatMap((columnName: String) => {
val stringIndexer = new StringIndexer()
.setInputCol(columnName)
.setOutputCol(s"${columnName}__idx")
.setHandleInvalid("keep")
val oneHotEncoder = new OneHotEncoder()
.setInputCol(s"${columnName}__idx")
.setOutputCol(s"${columnName}__ohe")
.setDropLast(false)
Array(stringIndexer, oneHotEncoder)
})
val listTransformers = listColumnNames.flatMap((columnName: String) => {
val countVectorizerModel = new CountVectorizer()
.setInputCol(columnName)
.setOutputCol(s"${columnName}__cv")
.setMinDF(10)
.setMinTF(1)
Array(countVectorizerModel)
})
val textTransformers = textColumnNames.flatMap((columnName: String) => {
val hanLPTokenizer = new HanLPTokenizer()
.setInputCol(columnName)
.setOutputCol(s"${columnName}__words")
.setShouldRemoveStopWords(true)
val stopWordsRemover = new StopWordsRemover()
.setInputCol(s"${columnName}__words")
.setOutputCol(s"${columnName}__filtered_words")
.setStopWords(StopWordsRemover.loadDefaultStopWords("english"))
val word2VecModelPath = s"${settings.dataDir}/${settings.today}/word2VecModel.parquet"
val word2VecModel = Word2VecModel.load(word2VecModelPath)
.setInputCol(s"${columnName}__filtered_words")
.setOutputCol(s"${columnName}__w2v")
Array(hanLPTokenizer, stopWordsRemover, word2VecModel)
})
val finalBooleanColumnNames = booleanColumnNames.toArray
val finalContinuousColumnNames = continuousColumnNames.toArray
val finalCategoricalColumnNames = categoricalColumnNames.map(columnName => s"${columnName}__ohe").toArray
val finalListColumnNames = listColumnNames.map(columnName => s"${columnName}__cv").toArray
val finalTextColumnNames = textColumnNames.map(columnName => s"${columnName}__w2v").toArray
val vectorAssembler = new SimpleVectorAssembler()
.setInputCols(finalBooleanColumnNames ++ finalContinuousColumnNames ++ finalCategoricalColumnNames ++ finalListColumnNames ++ finalTextColumnNames)
.setOutputCol("features")
val featureStages = mutable.ArrayBuffer.empty[PipelineStage]
featureStages += userRepoTransformer
featureStages += alsModel
featureStages ++= categoricalTransformers
featureStages ++= listTransformers
featureStages ++= textTransformers
featureStages += vectorAssembler
val featurePipeline = new Pipeline().setStages(featureStages.toArray)
val featurePipelineModel = featurePipeline.fit(profileStarringDF)ref:
https://spark.apache.org/docs/latest/ml-pipeline.html
https://spark.apache.org/docs/latest/ml-features.html
Handle Imbalanced Data
因為我們要訓練一個 Binary Classification 二元分類模型，會同時需要 positive（正樣本）和 negative（負樣本）。但是我們的原始數據 rawStarringDS 都是正樣本，也就是說我們只有「用戶有對哪些 repo 打星的資料」（正樣本），卻沒有「用戶沒有對哪些 repo 打星的資料」（負樣本）。我們當然是可以用「所有用戶沒有打星的 repo 做為負樣本」，但是考慮到這種做法產生的負樣本的數量實在太大，而且也不太合理，因為那些用戶沒有打星的 repo 不見得是因為他不喜歡，可能只是因為他不知道有那個 repo 存在。
我們後來採用的做法是「用熱門但是用戶沒有打星的 repo 做為負樣本」，我們寫了一個 Spark Transformer 來做這件事：
import ws.vinta.albedo.transformers.NegativeBalancer
import scala.collection.mutable
val sc = spark.sparkContext
val popularReposDS = loadPopularRepoDF()
val popularRepos = popularReposDS
.select($"repo_id".as[Int])
.collect()
.to[mutable.LinkedHashSet]
val bcPopularRepos = sc.broadcast(popularRepos)
val negativeBalancer = new NegativeBalancer(bcPopularRepos)
.setUserCol("user_id")
.setItemCol("repo_id")
.setTimeCol("starred_at")
.setLabelCol("starring")
.setNegativeValue(0.0)
.setNegativePositiveRatio(2.0)
val balancedStarringDF = negativeBalancer.transform(reducedStarringDF)ref:
https://github.com/vinta/albedo/blob/master/src/main/scala/ws/vinta/albedo/evaluators/RankingEvaluator.scala
http://www.kdnuggets.com/2017/06/7-techniques-handle-imbalanced-data.html
Split Data
直接使用 holdout 的方式，隨機分配不同的 row 到 training set 和 test set。其他的做法可能是根據時間來拆分，用以前的數據來預測之後的行為。
val profileBalancedStarringDF = balancedStarringDF
.join(userProfileDF, Seq("user_id"))
.join(repoProfileDF, Seq("repo_id"))
val tmpDF = featurePipelineModel.transform(profileBalancedStarringDF)
val keepColumnName = tmpDF.columns.filter((columnName: String) => {
!columnName.endsWith("__idx") &&
!columnName.endsWith("__ohe") &&
!columnName.endsWith("__cv") &&
!columnName.endsWith("__words") &&
!columnName.endsWith("__filtered_words") &&
!columnName.endsWith("__w2v")
})
val featuredBalancedStarringDF = tmpDF.select(keepColumnName.map(col): _*)
val Array(trainingFeaturedDF, testFeaturedDF) = featuredBalancedStarringDF.randomSplit(Array(0.9, 0.1))Build the Model Pipeline
為了方便之後的擴充性，這裡也使用 Spark ML Pipeline 的寫法。Spark ML 的 LogisticRegression 可以額外設置一個 weightCol 來調整不同 row 的權重。
import org.apache.spark.ml.classification.LogisticRegression
import org.apache.spark.ml.{Pipeline, PipelineStage}
import scala.collection.mutable
val weightSQL = """
SELECT *,
1.0 AS default_weight,
IF (starring = 1.0, 0.9, 0.1) AS positive_weight,
IF (starring = 1.0 AND datediff(current_date(), starred_at) <= 365, 0.9, 0.1) AS recent_starred_weight
FROM __THIS__
""".stripMargin
val weightTransformer = new SQLTransformer()
.setStatement(weightSQL)
val lr = new LogisticRegression()
.setMaxIter(200)
.setRegParam(0.7)
.setElasticNetParam(0.0)
.setStandardization(true)
.setLabelCol("starring")
.setFeaturesCol("standard_features")
.setWeightCol("recent_starred_weight")
val modelStages = mutable.ArrayBuffer.empty[PipelineStage]
modelStages += weightTransformer
modelStages += lr
val modelPipeline = new Pipeline().setStages(modelStages.toArray)
val modelPipelineModel = modelPipeline.fit(trainingFeaturedDF)ref:
https://spark.apache.org/docs/latest/ml-classification-regression.html
Evaluate the Model: Classification
因為 Logistic Regression 是二元分類模型，所以我們可以用 Spark ML 的 BinaryClassificationEvaluator 來評估結果。不過因為我們做的是推薦系統，真正在乎的是 Top N 的排序問題，所以這裡的 AUC 的數值參考一下就好。
import org.apache.spark.ml.evaluation.BinaryClassificationEvaluator
val testRankedDF = modelPipelineModel.transform(testFeaturedDF)
val binaryClassificationEvaluator = new BinaryClassificationEvaluator()
.setMetricName("areaUnderROC")
.setRawPredictionCol("rawPrediction")
.setLabelCol("starring")
val classificationMetric = binaryClassificationEvaluator.evaluate(testRankedDF)
println(s"${binaryClassificationEvaluator.getMetricName} = $classificationMetric")
// areaUnderROC = 0.9450631491281277ref:
https://spark.apache.org/docs/latest/api/scala/index.html#org.apache.spark.ml.evaluation.BinaryClassificationEvaluator
https://docs.databricks.com/spark/latest/mllib/binary-classification-mllib-pipelines.html
Generate Candidates
推薦系統的另外一個重要部分就是產生候選物品集，這裡我們使用以下幾種方式：
- ALS: 協同過濾的推薦
- Content-based: 基於內容的推薦
- Popularity: 基於熱門的推薦
不過因為這篇文章的主題是排序和特徵工程的 Machine Learning Pipeline，所以產生候選物品集的部分就不多說了，有興趣的人可以直接看底下連結的 source code 或是這個系列的其他文章。
import ws.vinta.albedo.recommenders.ALSRecommender
import ws.vinta.albedo.recommenders.ContentRecommender
import ws.vinta.albedo.recommenders.PopularityRecommender
val topK = 30
val alsRecommender = new ALSRecommender()
.setUserCol("user_id")
.setItemCol("repo_id")
.setTopK(topK)
val contentRecommender = new ContentRecommender()
.setUserCol("user_id")
.setItemCol("repo_id")
.setTopK(topK)
.setEnableEvaluationMode(true)
val popularityRecommender = new PopularityRecommender()
.setUserCol("user_id")
.setItemCol("repo_id")
.setTopK(topK)
val recommenders = mutable.ArrayBuffer.empty[Recommender]
recommenders += alsRecommender
recommenders += contentRecommender
recommenders += popularityRecommender
val candidateDF = recommenders
.map((recommender: Recommender) => recommender.recommendForUsers(testUserDF))
.reduce(_ union _)
.select($"user_id", $"repo_id")
.distinct()
// 每個 Recommender 的結果類似這樣：
// +-------+-------+----------+------+
// |user_id|repo_id|score |source|
// +-------+-------+----------+------+
// |652070 |1239728|0.6731846 |als |
// |652070 |854078 |0.7187486 |als |
// |652070 |1502338|0.70165294|als |
// |652070 |1184678|0.7434903 |als |
// |652070 |547708 |0.7956538 |als |
// +-------+-------+----------+------+ref:
https://github.com/vinta/albedo/blob/master/src/main/scala/ws/vinta/albedo/recommenders/ALSRecommender.scala
https://github.com/vinta/albedo/blob/master/src/main/scala/ws/vinta/albedo/recommenders/ContentRecommender.scala
https://github.com/vinta/albedo/blob/master/src/main/scala/ws/vinta/albedo/recommenders/PopularityRecommender.scala
Predict the Ranking
把這些候選物品集丟給我們訓練好的 Logistic Regression 模型來排序。結果中的 probability 欄位的第 0 項表示結果為 0 的機率（negative）、第 1 項表示結果為 1 的機率（positive）。
val profileCandidateDF = candidateDF
.join(userProfileDF, Seq("user_id"))
.join(repoProfileDF, Seq("repo_id"))
val featuredCandidateDF = featurePipelineModel
.transform(profileCandidateDF)
val rankedCandidateDF = modelPipelineModel
.transform(featuredCandidateDF)
// rankedCandidateDF 的結果類似這樣：
// +-------+--------+----------+----------------------------------------+
// |user_id|repo_id |prediction|probability |
// +-------+--------+----------+----------------------------------------+
// |652070 |83467664|1.0 |[0.12711894229094317,0.8728810577090568]|
// |652070 |55099616|1.0 |[0.1422859437320775,0.8577140562679224] |
// |652070 |42266235|1.0 |[0.1462014853157966,0.8537985146842034] |
// |652070 |78012800|1.0 |[0.15576081067098502,0.844239189329015] |
// |652070 |5928761 |1.0 |[0.16149848941925066,0.8385015105807493]|
// +-------+--------+----------+----------------------------------------+Evaluate the Model: Ranking
最後我們使用 Information Retrieval 領域中用來評價排序能力的指標 NDCG (Normalized Discounted Cumulative Gain) 來評估排序的結果。Spark MLlib 有現成的 RankingMetrics 可以用，但是它只適用於 RDD-based 的 API，所以我們改寫成適合 DataFrame-based 的 Evaluator。
import org.apache.spark.sql.expressions.Window
import org.apache.spark.sql.functions._
import ws.vinta.albedo.evaluators.RankingEvaluator
val userActualItemsDF = reducedStarringDF
.withColumn("rank", rank().over(Window.partitionBy($"user_id").orderBy($"starred_at".desc)))
.where($"rank" <= topK)
.groupBy($"user_id")
.agg(collect_list($"repo_id").alias("items"))
val userPredictedItemsDF = rankedCandidateDF
.withColumn("rank", rank().over(Window.partitionBy($"user_id").orderBy(toArrayUDF($"probability").getItem(1).desc)))
.where($"rank" <= topK)
.groupBy($"user_id")
.agg(collect_list($"repo_id").alias("items"))
val rankingEvaluator = new RankingEvaluator(userActualItemsDF)
.setMetricName("NDCG@k")
.setK(topK)
.setUserCol("user_id")
.setItemsCol("items")
val rankingMetric = rankingEvaluator.evaluate(userPredictedItemsDF)
println(s"${rankingEvaluator.getFormattedMetricName} = $rankingMetric")
// NDCG@30 = 0.021114356461615493ref:
https://github.com/vinta/albedo/blob/master/src/main/scala/ws/vinta/albedo/evaluators/RankingEvaluator.scala
https://spark.apache.org/docs/latest/mllib-evaluation-metrics.html#ranking-systems
https://weekly.codetengu.com/issues/83#kOxuVxW
