{"id":456,"date":"2017-10-31T19:35:31","date_gmt":"2017-10-31T11:35:31","guid":{"rendered":"https:\/\/vinta.ws\/code\/?p=456"},"modified":"2026-02-18T01:20:35","modified_gmt":"2026-02-17T17:20:35","slug":"build-a-recommender-system-with-spark-logistic-regression","status":"publish","type":"post","link":"https:\/\/vinta.ws\/code\/build-a-recommender-system-with-spark-logistic-regression.html","title":{"rendered":"Build a recommender system with Spark: Logistic Regression"},"content":{"rendered":"<p>\u5728\u9019\u500b\u7cfb\u5217\u7684\u6587\u7ae0\u88e1\uff0c\u6211\u5011\u5c07\u4f7f\u7528 Apache Spark\u3001XGBoost\u3001Elasticsearch \u548c MySQL \u7b49\u5de5\u5177\u4f86\u642d\u5efa\u4e00\u500b\u63a8\u85a6\u7cfb\u7d71\u7684 Machine Learning Pipeline\u3002\u63a8\u85a6\u7cfb\u7d71\u7684\u7d44\u6210\u53ef\u4ee5\u7c97\u7565\u5730\u5206\u6210 Candidate Generation \u548c Ranking \u5169\u500b\u90e8\u5206\uff0c\u524d\u8005\u662f\u91dd\u5c0d\u7528\u6236\u7522\u751f\u5019\u9078\u7269\u54c1\u96c6\uff0c\u5e38\u7528\u7684\u65b9\u6cd5\u6709 Collaborative Filtering\u3001Content-based\u3001\u6a19\u7c64\u914d\u5c0d\u3001\u71b1\u9580\u6392\u884c\u6216\u4eba\u5de5\u7cbe\u9078\u7b49\uff1b\u5f8c\u8005\u5247\u662f\u5c0d\u9019\u4e9b\u5019\u9078\u7269\u54c1\u6392\u5e8f\uff0c\u4ee5 Top N \u7684\u65b9\u5f0f\u5448\u73fe\u6700\u7d42\u7684\u63a8\u85a6\u7d50\u679c\uff0c\u5e38\u7528\u7684\u65b9\u6cd5\u6709 Logistic Regression\u3002<\/p>\n<p>\u5728\u672c\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u5011\u5c07\u4ee5 Ranking \u968e\u6bb5\u5e38\u7528\u7684\u65b9\u6cd5\u4e4b\u4e00\uff1aLogistic Regression \u908f\u8f2f\u8ff4\u6b78\u70ba\u4f8b\uff0c\u5229\u7528 Apache Spark \u7684 Logistic Regression \u6a21\u578b\u5efa\u7acb\u4e00\u500b GitHub repositories \u7684\u63a8\u85a6\u7cfb\u7d71\uff0c\u4ee5\u7528\u6236\u5c0d repo \u7684\u6253\u661f\u7d00\u9304\u548c\u7528\u6236\u8207 repo \u7684\u5404\u9805\u5c6c\u6027\u505a\u70ba\u7279\u5fb5\uff0c\u9810\u6e2c\u51fa\u7528\u6236\u6703\u4e0d\u6703\u6253\u661f\u67d0\u500b repo\uff08\u5206\u985e\u554f\u984c\uff09\u3002\u6700\u5f8c\u8a13\u7df4\u51fa\u4f86\u7684\u6a21\u578b\u5c31\u53ef\u4ee5\u505a\u70ba\u6211\u5011\u7684\u63a8\u85a6\u7cfb\u7d71\u7684 Ranking \u6a21\u7d44\u3002\u4e0d\u904e\u56e0\u70ba LR \u662f\u7dda\u6027\u6a21\u578b\uff0c\u6240\u4ee5\u901a\u5e38\u9700\u8981\u5927\u91cf\u7684 Feature Engineering \u4f86\u7fd2\u5f97\u975e\u7dda\u6027\u95dc\u4fc2\u3002\u6240\u4ee5\u9019\u7bc7\u6587\u7ae0\u7684\u91cd\u9ede\u662f Spark ML \u7684 Pipeline \u6a5f\u5236\u548c\u7279\u5fb5\u5de5\u7a0b\uff0c\u4e0d\u6703\u5728\u6f14\u7b97\u6cd5\u7684\u90e8\u5206\u8457\u58a8\u592a\u591a\u3002<\/p>\n<p>\u5b8c\u6574\u7684\u7a0b\u5f0f\u78bc\u53ef\u4ee5\u5728 <a href=\"https:\/\/github.com\/vinta\/albedo\">https:\/\/github.com\/vinta\/albedo<\/a> \u627e\u5230\u3002<\/p>\n<p>\u7cfb\u5217\u6587\u7ae0\uff1a<\/p>\n<ul>\n<li><a href=\"https:\/\/vinta.ws\/code\/build-a-recommender-system-with-pyspark-implicit-als.html\">Build a recommender system with Spark: Implicit ALS<\/a><\/li>\n<li><a href=\"https:\/\/vinta.ws\/code\/build-a-recommender-system-with-spark-and-elasticsearch-content-based.html\">Build a recommender system with Spark: Content-based and Elasticsearch<\/a><\/li>\n<li><a href=\"https:\/\/vinta.ws\/code\/build-a-recommender-system-with-spark-logistic-regression.html\">Build a recommender system with Spark: Logistic Regression<\/a><\/li>\n<li><a href=\"https:\/\/vinta.ws\/code\/feature-engineering.html\">Feature Engineering \u7279\u5fb5\u5de5\u7a0b\u4e2d\u5e38\u898b\u7684\u65b9\u6cd5<\/a><\/li>\n<li><a href=\"https:\/\/vinta.ws\/code\/spark-ml-cookbook-scala.html\">Spark ML cookbook (Scala)<\/a><\/li>\n<li><a href=\"https:\/\/vinta.ws\/code\/spark-sql-cookbook-scala.html\">Spark SQL cookbook (Scala)<\/a><\/li>\n<li>\u4e0d\u5b9a\u671f\u66f4\u65b0\u4e2d<\/li>\n<\/ul>\n<h2>Submit the Application<\/h2>\n<pre class=\"line-numbers\"><code class=\"language-bash\">$ spark-submit \n--master spark:\/\/192.168.10.100:7077 \n--packages \"com.github.fommil.netlib:all:1.1.2,com.hankcs:hanlp:portable-1.3.4,mysql:mysql-connector-java:5.1.41\" \n--class ws.vinta.albedo.UserProfileBuilder \ntarget\/albedo-1.0.0-SNAPSHOT.jar\n\n$ spark-submit \n--master spark:\/\/192.168.10.100:7077 \n--packages \"com.github.fommil.netlib:all:1.1.2,com.hankcs:hanlp:portable-1.3.4,mysql:mysql-connector-java:5.1.41\" \n--class ws.vinta.albedo.RepoProfileBuilder \ntarget\/albedo-1.0.0-SNAPSHOT.jar\n\n$ spark-submit \n--master spark:\/\/192.168.10.100:7077 \n--packages \"com.github.fommil.netlib:all:1.1.2,com.hankcs:hanlp:portable-1.3.4,mysql:mysql-connector-java:5.1.41\" \n--class ws.vinta.albedo.LogisticRegressionRanker \ntarget\/albedo-1.0.0-SNAPSHOT.jar<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"https:\/\/vinta.ws\/code\/setup-spark-scala-and-maven-with-intellij-idea.html\">https:\/\/vinta.ws\/code\/setup-spark-scala-and-maven-with-intellij-idea.html<\/a><br \/>\n<a href=\"https:\/\/spark.apache.org\/docs\/latest\/submitting-applications.html\">https:\/\/spark.apache.org\/docs\/latest\/submitting-applications.html<\/a><br \/>\n<a href=\"https:\/\/spoddutur.github.io\/spark-notes\/distribution_of_executors_cores_and_memory_for_spark_application\">https:\/\/spoddutur.github.io\/spark-notes\/distribution_of_executors_cores_and_memory_for_spark_application<\/a><\/p>\n<h2>Load Data<\/h2>\n<p>\u6211\u5011\u4e4b\u524d\u5df2\u7d93\u5229\u7528 GitHub API \u548c BigQuery \u4e0a\u7684 GitHub Archive \u6536\u96c6\u4e86 150 \u842c\u7b46\u7684\u6253\u661f\u7d00\u9304\u548c\u6240\u5c6c\u7684\u7528\u6236\u3001repo \u6578\u64da\u3002\u76ee\u524d\u6709\u4ee5\u4e0b\u5e7e\u500b\u6578\u64da\u96c6\uff0c\u5927\u81f4\u4e0a\u662f\u7167\u8457 GitHub API \u5efa\u7acb\u7684\uff0c\u6b04\u4f4d\u5206\u5225\u5982\u4e0b\uff1a<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">rawUserInfoDS.printSchema()\n\/\/ root\n \/\/ |-- user_id: integer (nullable = true)\n \/\/ |-- user_login: string (nullable = true)\n \/\/ |-- user_account_type: string (nullable = true)\n \/\/ |-- user_name: string (nullable = true)\n \/\/ |-- user_company: string (nullable = true)\n \/\/ |-- user_blog: string (nullable = true)\n \/\/ |-- user_location: string (nullable = true)\n \/\/ |-- user_email: string (nullable = true)\n \/\/ |-- user_bio: string (nullable = true)\n \/\/ |-- user_public_repos_count: integer (nullable = true)\n \/\/ |-- user_public_gists_count: integer (nullable = true)\n \/\/ |-- user_followers_count: integer (nullable = true)\n \/\/ |-- user_following_count: integer (nullable = true)\n \/\/ |-- user_created_at: timestamp (nullable = true)\n \/\/ |-- user_updated_at: timestamp (nullable = true)\n\nrawRepoInfoDS.printSchema()\n\/\/ |-- repo_id: integer (nullable = true)\n \/\/ |-- repo_owner_id: integer (nullable = true)\n \/\/ |-- repo_owner_username: string (nullable = true)\n \/\/ |-- repo_owner_type: string (nullable = true)\n \/\/ |-- repo_name: string (nullable = true)\n \/\/ |-- repo_full_name: string (nullable = true)\n \/\/ |-- repo_description: string (nullable = true)\n \/\/ |-- repo_language: string (nullable = true)\n \/\/ |-- repo_created_at: timestamp (nullable = true)\n \/\/ |-- repo_updated_at: timestamp (nullable = true)\n \/\/ |-- repo_pushed_at: timestamp (nullable = true)\n \/\/ |-- repo_homepage: string (nullable = true)\n \/\/ |-- repo_size: integer (nullable = true)\n \/\/ |-- repo_stargazers_count: integer (nullable = true)\n \/\/ |-- repo_forks_count: integer (nullable = true)\n \/\/ |-- repo_subscribers_count: integer (nullable = true)\n \/\/ |-- repo_is_fork: boolean (nullable = true)\n \/\/ |-- repo_has_issues: boolean (nullable = true)\n \/\/ |-- repo_has_projects: boolean (nullable = true)\n \/\/ |-- repo_has_downloads: boolean (nullable = true)\n \/\/ |-- repo_has_wiki: boolean (nullable = true)\n \/\/ |-- repo_has_pages: boolean (nullable = true)\n \/\/ |-- repo_open_issues_count: integer (nullable = true)\n \/\/ |-- repo_topics: string (nullable = true)\n\nrawStarringDS.printSchema()\n\/\/ root\n \/\/ |-- user_id: integer (nullable = true)\n \/\/ |-- repo_id: integer (nullable = true)\n \/\/ |-- starred_at: timestamp (nullable = true)\n \/\/ |-- starring: double (nullable = true)<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"https:\/\/www.githubarchive.org\/\">https:\/\/www.githubarchive.org\/<\/a><br \/>\n<a href=\"http:\/\/ghtorrent.org\/\">http:\/\/ghtorrent.org\/<\/a><\/p>\n<p>\u8f09\u5165\u8cc7\u6599\u4e4b\u5f8c\uff0c\u8981\u505a\u7684\u7b2c\u4e00\u4ef6\u4e8b\u61c9\u8a72\u5c31\u662f Exploratory Data Analysis (EDA) \u4e86\uff0c\u628a\u73a9\u4e00\u4e0b\u624b\u4e0a\u7684\u6578\u64da\u3002\u5efa\u8b70\u5927\u5bb6\u53ef\u4ee5\u8a66\u8a66 Apache Zeppelin \u6216\u662f Databricks \u7684 Notebook\uff0c\u9664\u4e86\u5167\u5efa Spark \u652f\u63f4\u7684\u6240\u6709\u8a9e\u8a00\u4e4b\u5916\uff0c\u4e5f\u6574\u5408\u4e86 NoSQL \u548c JDBC \u652f\u63f4\u7684\u8cc7\u6599\u5eab\uff0c\u8981\u756b\u5716\u8868\u4e5f\u5f88\u65b9\u4fbf\uff0c\u7c21\u76f4\u6bd4 Jupyter Notebook \u9084\u597d\u7528\u4e86\u3002<\/p>\n<p>ref:<br \/>\n<a href=\"https:\/\/zeppelin.apache.org\/\">https:\/\/zeppelin.apache.org\/<\/a><br \/>\n<a href=\"https:\/\/databricks.com\/\">https:\/\/databricks.com\/<\/a><\/p>\n<h2>Build User Profile \/ Item Profile<\/h2>\n<p>\u5728\u9019\u500b\u5c08\u6848\u4e2d\uff0c\u6700\u4e3b\u8981\u7684\u6578\u64da\u4e3b\u9ad4\u5c31\u662f user \u548c repo\uff0c\u6240\u4ee5\u6211\u5011\u91dd\u5c0d\u5169\u8005\u5404\u81ea\u5efa\u7acb User Profile \u548c Item Profile\uff0c\u4f5c\u70ba\u4e4b\u5f8c\u5728\u6a21\u578b\u8a13\u7df4\u968e\u6bb5\u6703\u4f7f\u7528\u7684\u7279\u5fb5\u3002\u6211\u5011\u628a\u9019\u500b\u6b65\u9a5f\u8ddf\u6a21\u578b\u8a13\u7df4\u7684\u6d41\u7a0b\u5206\u958b\uff0c\u9019\u6a23\u5c0d\u6574\u500b\u67b6\u69cb\u7684\u642d\u5efa\u6703\u66f4\u6709\u5f48\u6027\u3002\u5be6\u52d9\u4e0a\uff0c\u6211\u5011\u53ef\u4ee5\u7528 user id \u6216 item id \u7576 key\uff0c\u76f4\u63a5\u628a\u88fd\u4f5c\u597d\u7684\u7279\u5fb5\u5b58\u9032 Redis \u6216\u5176\u4ed6 schemaless \u7684 NoSQL \u8cc7\u6599\u5eab\uff0c\u65b9\u4fbf\u4e4b\u5f8c\u7d66\u591a\u500b\u6a21\u578b\u53d6\u7528\uff1b\u5728\u505a real-time \u63a8\u85a6\u6642\uff0c\u4e5f\u53ef\u4ee5\u5f88\u5feb\u5730\u62ff\u5230\u7279\u5fb5\uff0c\u53ea\u9700\u8981\u91cd\u65b0\u8a08\u7b97\u90e8\u4efd\u6b04\u4f4d\u5373\u53ef\u3002<\/p>\n<p>\u4e0d\u904e\u56e0\u70ba\u9019\u88e1\u4e3b\u8981\u7528\u7684\u662f\u4f86\u81ea GitHub API \u7684\u8cc7\u6599\uff0c\u67d0\u7a2e\u7a0b\u5ea6\u4e0a\u4eba\u5bb6\u5df2\u7d93\u5e6b\u6211\u5011\u505a\u4e86\u5f88\u591a\u8cc7\u6599\u6e05\u7406\u548c\u6b63\u898f\u5316\u7684\u52d5\u4f5c\u4e86\uff0c\u4f46\u662f\u5728\u73fe\u5be6\u4e2d\uff0c\u4f60\u7684\u7cfb\u7d71\u8981\u8655\u7406\u7684\u6578\u64da\u901a\u5e38\u4e0d\u6703\u9019\u9ebc\u4e7e\u6de8\uff0c\u53ef\u80fd\u4f86\u81ea\u5404\u7a2e data source\u3001\u6709\u8457\u5404\u7a2e\u683c\u5f0f\uff0c\u9084\u6703\u96a8\u8457\u6642\u9593\u800c\u6539\u8b8a\uff0c\u901a\u5e38\u5f97\u82b1\u4e0a\u4e0d\u5c11\u529b\u6c23\u505a Extract, Transform, Load (ETL)\uff0c\u6240\u4ee5\u6700\u597d\u5728\u5beb log\uff08\u57cb\u9ede\uff09\u7684\u6642\u5019\u5c31\u6e9d\u901a\u597d\u3002\u800c\u4e14\u5728 production \u74b0\u5883\u4e2d\uff0c\u6578\u64da\u662f\u6703\u4e00\u76f4\u8b8a\u52d5\u7684\uff0c\u8981\u78ba\u4fdd\u6578\u64da\u7684\u6642\u6548\u6027\u548c\u5bb9\u932f\u6027\uff0c\u5f88\u91cd\u8981\u7684\u4e00\u500b\u90e8\u5206\u5c31\u662f monitoring\u3002<\/p>\n<p>ref:<br \/>\n<a href=\"http:\/\/www.algorithmdog.com\/ad-rec-deploy\">http:\/\/www.algorithmdog.com\/ad-rec-deploy<\/a><br \/>\n<a href=\"https:\/\/tech.meituan.com\/online-feature-system02.html\">https:\/\/tech.meituan.com\/online-feature-system02.html<\/a><\/p>\n<p>\u7919\u65bc\u7bc7\u5e45\u6709\u9650\uff0c\u4ee5\u4e0b\u7684\u6587\u7ae0\u4e2d\u6211\u5011\u53ea\u6703\u6311\u5e7e\u500b\u91cd\u8981\u7684\u90e8\u5206\u8aaa\u660e\u3002\u7c21\u55ae\u8aaa\uff0c\u5728\u9019\u500b\u6b65\u9a5f\u7684\u6700\u5f8c\uff0c\u6211\u5011\u6703\u5f97\u5230 <code>userProfileDF<\/code> \u548c <code>repoProfileDF<\/code> \u9019\u5169\u500b DataFrame\uff0c\u5206\u5225\u5b58\u653e\u88fd\u4f5c\u597d\u7684\u7279\u5fb5\u3002\u8a73\u7d30\u7684\u7a0b\u5f0f\u78bc\u5982\u4e0b\uff1a<\/p>\n<p>ref:<br \/>\n<a href=\"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\/UserProfileBuilder.scala<\/a><br \/>\n<a href=\"https:\/\/github.com\/vinta\/albedo\/blob\/master\/src\/main\/scala\/ws\/vinta\/albedo\/RepoProfileBuilder.scala\">https:\/\/github.com\/vinta\/albedo\/blob\/master\/src\/main\/scala\/ws\/vinta\/albedo\/RepoProfileBuilder.scala<\/a><\/p>\n<h2>Feature Engineering<\/h2>\n<p>\u4ee5\u63a8\u85a6\u7cfb\u7d71\u70ba\u4f8b\uff0c\u7279\u5fb5\u53ef\u4ee5\u5206\u6210\u4ee5\u4e0b\u56db\u7a2e\uff1a<\/p>\n<ul>\n<li>\u7528\u6236\u7279\u5fb5\uff1a\u7528\u6236\u672c\u8eab\u7684\u5404\u7a2e\u5c6c\u6027\uff0c\u4f8b\u5982 user id\u3001\u6027\u5225\u3001\u8077\u696d\u6216\u6240\u5728\u7684\u57ce\u5e02\u7b49<\/li>\n<li>\u7269\u54c1\u7279\u5fb5\uff1a\u7269\u54c1\u672c\u8eab\u7684\u5404\u7a2e\u5c6c\u6027\uff0c\u4f8b\u5982 item id\u3001\u4f5c\u8005\u3001\u6a19\u984c\u3001\u5206\u985e\u3001\u8a55\u5206\u6216\u6240\u5c6c\u7684\u6a19\u7c64\u7b49<\/li>\n<li>\u4ea4\u4e92\u7279\u5fb5\uff1a\u7528\u6236\u5c0d\u7269\u54c1\u505a\u51fa\u7684\u67d0\u9805\u884c\u70ba\uff0c\u8a72\u884c\u70ba\u7684 aggregation \u6216\u4ea4\u53c9\u7279\u5fb5\uff0c\u4f8b\u5982\u662f\u5426\u770b\u904e\u540c\u985e\u578b\u7684\u96fb\u5f71\u3001\u6700\u8fd1\u807d\u7684\u6b4c\u66f2\u7684\u66f2\u98a8\u5206\u4f48\u6216\u4e0a\u9031\u8cb7\u4e86\u591a\u5c11\u9ad8\u55ae\u50f9\u7684\u5546\u54c1<\/li>\n<li>\u4e0a\u4e0b\u6587\u7279\u5fb5\uff1a\u7528\u6236\u5c0d\u7269\u54c1\u505a\u51fa\u7684\u67d0\u9805\u884c\u70ba\uff0c\u8a72\u884c\u70ba\u7684 metadata\uff0c\u4f8b\u5982\u767c\u751f\u7684\u6642\u9593\u3001\u4f7f\u7528\u7684\u88dd\u7f6e\u6216\u7576\u524d\u7684 GPS \u4f4d\u7f6e\u7b49<\/li>\n<\/ul>\n<p>\u6709\u4e9b\u7279\u5fb5\u662f\u5728\u8cc7\u6599\u63a1\u96c6\u968e\u6bb5\u5c31\u80fd\u62ff\u5230\uff0c\u6709\u4e9b\u7279\u5fb5\u5247\u6703\u9700\u8981\u984d\u5916\u7684\u6b65\u9a5f\uff08\u4f8b\u5982\u900f\u904e\u5916\u90e8\u7684 API \u6216\u662f\u5176\u4ed6\u6a21\u578b\uff09\u624d\u80fd\u53d6\u5f97\uff0c\u4e5f\u6709\u4e9b\u7279\u5fb5\u5fc5\u9808\u5373\u6642\u66f4\u65b0\u3002\u9806\u9053\u4e00\u63d0\uff0c\u56e0\u70ba\u6211\u5011\u8981\u9810\u6e2c\u7684\u662f\u300c\u67d0\u500b\u7528\u6236\u6703\u4e0d\u6703\u6253\u661f\u67d0\u500b repo\u300d\uff0c\u6240\u4ee5\u4e0b\u8ff0\u7279\u5fb5\u88e1\u7684 user \u53ef\u4ee5\u662f repo stargazer \u4e5f\u53ef\u4ee5\u662f repo owner\u3002<\/p>\n<p>\u539f\u59cb\u7279\u5fb5\uff1a<\/p>\n<ul>\n<li>\u7528\u6236\u7279\u5fb5\n<ul>\n<li><code>user_id<\/code><\/li>\n<li><code>user_login<\/code><\/li>\n<li><code>user_name<\/code><\/li>\n<li><code>user_email<\/code><\/li>\n<li><code>user_blog<\/code><\/li>\n<li><code>user_bio<\/code><\/li>\n<li><code>user_company<\/code><\/li>\n<li><code>user_location<\/code><\/li>\n<li><code>user_followers_coung<\/code><\/li>\n<li><code>user_following_count<\/code><\/li>\n<li><code>user_public_repos_count<\/code><\/li>\n<li><code>user_public_gists_count<\/code><\/li>\n<li><code>user_created_at<\/code><\/li>\n<li><code>user_updated_at<\/code><\/li>\n<\/ul>\n<\/li>\n<li>\u7269\u54c1\u7279\u5fb5\n<ul>\n<li><code>repo_id<\/code><\/li>\n<li><code>repo_name<\/code><\/li>\n<li><code>repo_owner<\/code><\/li>\n<li><code>repo_owner_type<\/code><\/li>\n<li><code>repo_language<\/code><\/li>\n<li><code>repo_description<\/code><\/li>\n<li><code>repo_homepage<\/code><\/li>\n<li><code>repo_subscribers_count<\/code><\/li>\n<li><code>repo_stargazers_count<\/code><\/li>\n<li><code>repo_forks_count<\/code><\/li>\n<li><code>repo_size<\/code><\/li>\n<li><code>repo_created_at<\/code><\/li>\n<li><code>repo_updated_at<\/code><\/li>\n<li><code>repo_pushed_at<\/code><\/li>\n<li><code>repo_has_issues<\/code><\/li>\n<li><code>repo_has_projects<\/code><\/li>\n<li><code>repo_has_downloads<\/code><\/li>\n<li><code>repo_has_wiki<\/code><\/li>\n<li><code>repo_has_pages<\/code><\/li>\n<li><code>repo_open_issues_count<\/code><\/li>\n<li><code>repo_topics<\/code><\/li>\n<\/ul>\n<\/li>\n<li>\u4ea4\u4e92\u7279\u5fb5\n<ul>\n<li><code>user_stars_repo<\/code><\/li>\n<li><code>user_follows_user<\/code><\/li>\n<\/ul>\n<\/li>\n<li>\u4e0a\u4e0b\u6587\u7279\u5fb5\n<ul>\n<li><code>user_repo_starred_at<\/code><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>\u767c\u60f3\u7279\u5fb5\uff1a<\/p>\n<ul>\n<li>\u7528\u6236\u7279\u5fb5\n<ul>\n<li><code>user_days_between_created_at_today<\/code>: \u8a72\u7528\u6236\u7684\u8a3b\u518a\u65e5\u671f\u8ddd\u96e2\u4eca\u5929\u904e\u4e86\u5e7e\u5e74<\/li>\n<li><code>user_days_between_updated_at_today<\/code>: \u8a72\u7528\u6236\u7684\u66f4\u65b0\u65e5\u671f\u8ddd\u96e2\u4eca\u5929\u904e\u4e86\u5e7e\u5929<\/li>\n<li><code>user_repos_avg_stargazers_count<\/code>: \u8a72\u7528\u6236\u540d\u4e0b\u7684\u6240\u6709 repo\uff08\u4e0d\u542b fork \u7684\uff09\u7684\u5e73\u5747\u661f\u661f\u6578<\/li>\n<li><code>user_organizations<\/code>: \u8a72\u7528\u6236\u5c6c\u65bc\u54ea\u4e9b\u7d44\u7e54<\/li>\n<li><code>user_has_null<\/code>: \u8a72\u7528\u6236\u81f3\u5c11\u6709\u4e00\u500b\u6b04\u4f4d\u662f null<\/li>\n<li><code>user_has_blog<\/code>: \u8a72\u7528\u6236\u6709\u6c92\u6709\u7db2\u7ad9<\/li>\n<li><code>user_is_freelancer<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b Freelancer \u7b49\u5b57\u773c<\/li>\n<li><code>user_is_junior<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b Beginner \u6216 Junior \u7b49\u5b57\u773c<\/li>\n<li><code>user_is_lead<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b Team Lead\u3001Architect\u3001Creator\u3001CTO \u6216 VP of Engineering \u7b49\u5b57\u773c<\/li>\n<li><code>user_is_scholar<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b Researcher\u3001Scientist\u3001PhD \u6216 Professor \u7b49\u5b57\u773c<\/li>\n<li><code>user_is_pm<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b Product Manager \u7b49\u5b57\u773c<\/li>\n<li><code>user_knows_backend<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b Backend \u6216 Back end \u7b49\u5b57\u773c<\/li>\n<li><code>user_knows_data<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b Machine Learning\u3001Deep Learning \u6216 Data Science \u7b49\u5b57\u773c<\/li>\n<li><code>user_knows_devops<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b DevOps\u3001SRE\u3001SysAdmin \u6216 Infrastructure \u7b49\u5b57\u773c<\/li>\n<li><code>user_knows_frontend<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b Frontend \u6216 Front end \u7b49\u5b57\u773c<\/li>\n<li><code>user_knows_mobile<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b Mobile\u3001iOS \u6216 Android \u7b49\u5b57\u773c<\/li>\n<li><code>user_knows_recsys<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b Recommender System\u3001Data Mining \u6216 Information Retrieval \u7b49\u5b57\u773c<\/li>\n<li><code>user_knows_web<\/code>: \u8a72\u7528\u6236\u7684 bio \u4e2d\u662f\u5426\u5305\u542b Web Development \u6216 Fullstack \u7b49\u5b57\u773c<\/li>\n<\/ul>\n<\/li>\n<li>\u7269\u54c1\u7279\u5fb5\n<ul>\n<li><code>repo_created_at_days_since_today<\/code>: \u8a72 repo \u7684\u5efa\u7acb\u65e5\u671f\u8ddd\u96e2\u4eca\u5929\u904e\u4e86\u5e7e\u5929<\/li>\n<li><code>repo_updated_at_days_since_today<\/code>: \u8a72 repo \u7684\u66f4\u65b0\u65e5\u671f\u8ddd\u96e2\u4eca\u5929\u904e\u4e86\u5e7e\u5929<\/li>\n<li><code>repo_pushed_at_days_since_today<\/code>: \u8a72 repo \u7684\u63d0\u4ea4\u65e5\u671f\u8ddd\u96e2\u4eca\u5929\u904e\u4e86\u5e7e\u5929<\/li>\n<li><code>repo_stargazers_count_in_30days<\/code>: \u8a72 repo \u5728 30 \u5929\u5167\u6536\u5230\u7684\u661f\u661f\u6578<\/li>\n<li><code>repo_subscribers_stargazers_ratio<\/code>: \u8a72 repo \u7684 watch \u6578\u548c star \u6578\u7684\u6bd4\u4f8b<\/li>\n<li><code>repo_forks_stargazers_ratio<\/code>: \u8a72 repo \u7684 fork \u6578\u548c star \u6578\u7684\u6bd4\u4f8b<\/li>\n<li><code>repo_open_issues_stargazers_ratio<\/code>: \u8a72 repo \u7684 \u6578\u548c star \u6578\u7684\u6bd4\u4f8b<\/li>\n<li><code>repo_releases_count<\/code>: \u8a72 repo \u7684 release \u6216 tag \u6578<\/li>\n<li><code>repo_lisence<\/code>: \u8a72 repo \u7684\u6388\u6b0a\u689d\u6b3e<\/li>\n<li><code>repo_readme<\/code>: \u8a72 repo \u7684 README \u5167\u5bb9<\/li>\n<li><code>repo_has_null<\/code>: \u8a72 repo \u6709\u81f3\u5c11\u4e00\u500b\u6b04\u4f4d\u662f null<\/li>\n<li><code>repo_has_readme<\/code>: \u8a72 repo \u662f\u5426\u6709 README \u6a94\u6848<\/li>\n<li><code>repo_has_changelog<\/code>: \u8a72 repo \u662f\u5426\u6709 CHANGELOG \u6a94\u6848<\/li>\n<li><code>repo_has_contributing<\/code>: \u8a72 repo \u662f\u5426\u6709 CONTRIBUTING \u6a94\u6848<\/li>\n<li><code>repo_has_tests<\/code>: \u8a72 repo \u662f\u5426\u6709\u6e2c\u8a66<\/li>\n<li><code>repo_has_ci<\/code>: \u8a72 repo \u662f\u5426\u6709 CI<\/li>\n<li><code>repo_has_dockerfile<\/code>: \u8a72 repo \u662f\u5426\u6709 Dockerfile<\/li>\n<li><code>repo_is_unmaintained<\/code>: \u8a72 repo \u662f\u5426\u4e0d\u518d\u7dad\u8b77\u4e86<\/li>\n<li><code>repo_is_awesome<\/code>: \u8a72 repo \u662f\u5426\u88ab\u6536\u9304\u9032\u4efb\u4f55\u7684 awesome-xxx \u5217\u8868\u88e1<\/li>\n<li><code>repo_is_vinta_starred<\/code>: \u8a72 repo \u662f\u5426\u88ab @vinta aka \u672c\u6587\u7684\u4f5c\u8005\u6253\u661f\u4e86<\/li>\n<\/ul>\n<\/li>\n<li>\u4ea4\u4e92\u7279\u5fb5\n<ul>\n<li><code>user_starred_repos_count<\/code>: \u8a72\u7528\u6236\u7e3d\u5171\u6253\u661f\u4e86\u591a\u5c11 repo<\/li>\n<li><code>user_avg_daily_starred_repos_count<\/code>: \u8a72\u7528\u6236\u5e73\u5747\u6bcf\u5929\u6253\u661f\u591a\u5c11 repo<\/li>\n<li><code>user_forked_repos_count<\/code>: \u8a72\u7528\u6236\u7e3d\u5171 fork \u4e86\u591a\u5c11 repo<\/li>\n<li><code>user_follower_following_count_ratio<\/code>: \u8a72\u7528\u6236\u7684 follower \u6578\u548c following \u6578\u7684\u6bd4\u4f8b<\/li>\n<li><code>user_recent_searched_keywords<\/code>: \u8a72\u7528\u6236\u6700\u8fd1\u641c\u5c0b\u7684 50 \u500b\u95dc\u9375\u5b57<\/li>\n<li><code>user_recent_commented_repos<\/code>: \u8a72\u7528\u6236\u6700\u8fd1\u7559\u8a00\u7684 50 \u500b repo<\/li>\n<li><code>user_recent_watched_repos<\/code>: \u8a72\u7528\u6236\u6700\u8fd1\u8a02\u95b1\u7684 50 \u500b repo<\/li>\n<li><code>user_recent_starred_repos_descriptions<\/code>: \u8a72\u7528\u6236\u6700\u8fd1\u6253\u661f\u7684 50 \u500b repo \u7684\u63cf\u8ff0<\/li>\n<li><code>user_recent_starred_repos_languages<\/code>: \u8a72\u7528\u6236\u6700\u8fd1\u6253\u661f\u7684 50 \u500b repo \u7684\u8a9e\u8a00<\/li>\n<li><code>user_recent_starred_repos_topics<\/code>: \u8a72\u7528\u6236\u6700\u8fd1\u6253\u661f\u7684 50 \u500b repo \u7684\u6a19\u7c64<\/li>\n<li><code>user_follows_repo_owner<\/code>: \u8a72\u7528\u6236\u662f\u5426\u8ffd\u8e64\u8a72 repo \u7684\u4f5c\u8005<\/li>\n<li><code>repo_language_index_in_user_recent_repo_languages<\/code>: \u8a72 repo \u7684\u8a9e\u8a00\u51fa\u73fe\u5728\u8a72\u7528\u6236\u6700\u8fd1\u6253\u661f\u7684\u8a9e\u8a00\u5217\u8868\u7684\u9806\u5e8f<\/li>\n<li><code>repo_language_count_in_user_recent_repo_languages<\/code>: \u8a72 repo \u7684\u8a9e\u8a00\u51fa\u73fe\u5728\u8a72\u7528\u6236\u6700\u8fd1\u6253\u661f\u7684\u8a9e\u8a00\u5217\u8868\u7684\u6b21\u6578<\/li>\n<li><code>repo_topics_user_recent_topics_similarity<\/code>: \u8a72 repo \u7684\u6a19\u7c64\u8207\u8a72\u7528\u6236\u6700\u8fd1\u6253\u661f\u7684\u6a19\u7c64\u5217\u8868\u7684\u76f8\u4f3c\u5ea6<\/li>\n<\/ul>\n<\/li>\n<li>\u4e0a\u4e0b\u6587\u7279\u5fb5\n<ul>\n<li><code>als_model_prediction<\/code>: \u4f86\u81ea ALS \u6a21\u578b\u7684\u9810\u6e2c\u503c\uff0c\u8a72\u7528\u6236\u5c0d\u8a72 repo \u7684\u504f\u597d\u7a0b\u5ea6<\/li>\n<li><code>gbdt_model_index<\/code>: \u4f86\u81ea GBDT \u6a21\u578b\u7684 tree index\uff0c\u8a72 observation \u7684\u81ea\u52d5\u7279\u5fb5<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>Feature Engineering \u7279\u5fb5\u5de5\u7a0b\u4e2d\u5e38\u898b\u7684\u65b9\u6cd5<br \/>\n<a href=\"https:\/\/vinta.ws\/code\/feature-engineering.html\">https:\/\/vinta.ws\/code\/feature-engineering.html<\/a><\/p>\n<h3>Detect Outliers<\/h3>\n<p>\u9664\u4e86\u7f3a\u5931\u503c\u4e4b\u5916\uff0c\u96e2\u7fa4\u503c\uff08\u7570\u5e38\u503c\uff09\u4e5f\u662f\u9700\u8981\u6ce8\u610f\u7684\u5730\u65b9\u3002\u5982\u679c\u662f continuous \u7279\u5fb5\uff0c\u7528 box plot \u53ef\u4ee5\u5f88\u5feb\u5730\u767c\u73fe\u96e2\u7fa4\u503c\uff1b\u5982\u679c\u662f categorical \u7279\u5fb5\uff0c\u53ef\u4ee5 <code>SELECT COUNT(*) ... GROUP BY<\/code> \u4e00\u4e0b\uff0c\u7136\u5f8c\u7528 bar chart \u67e5\u770b\u6bcf\u500b category \u7684\u6578\u91cf\u3002\u53d6\u6c7a\u65bc\u4f60\u6240\u8981\u89e3\u6c7a\u7684\u554f\u984c\uff0c\u7570\u5e38\u503c\u53ef\u80fd\u53ef\u4ee5\u76f4\u63a5\u5ffd\u7565\uff0c\u4e5f\u53ef\u80fd\u9700\u8981\u7279\u5225\u5c0d\u5f85\uff0c\u4f8b\u5982\u641e\u6e05\u695a\u7570\u5e38\u503c\u51fa\u73fe\u7684\u539f\u56e0\uff0c\u662f\u8cc7\u6599\u63a1\u96c6\u6642\u7684\u5dee\u932f\u6216\u662f\u67d0\u7a2e\u96b1\u542b\u7684\u6df1\u5c64\u7684\u56e0\u7d20\u4e4b\u985e\u7684\u3002<\/p>\n<p>ref:<br \/>\n<a href=\"https:\/\/www.analyticsvidhya.com\/blog\/2016\/01\/guide-data-exploration\/\">https:\/\/www.analyticsvidhya.com\/blog\/2016\/01\/guide-data-exploration\/<\/a><br \/>\n<a href=\"https:\/\/www.slideshare.net\/tw_dsconf\/123-70852901\">https:\/\/www.slideshare.net\/tw_dsconf\/123-70852901<\/a><\/p>\n<h3>Impute Missing Values<\/h3>\n<p>\u53ef\u4ee5\u5229\u7528 <code>df.describe().show()<\/code> \u67e5\u770b\u5404\u500b\u6b04\u4f4d\u7684\u7d71\u8a08\u6578\u64da\uff1a<code>count<\/code>\u3001<code>mean<\/code>\u3001<code>stddev<\/code>\u3001<code>min<\/code> \u548c <code>max<\/code>\u3002\u9664\u4e86\u4f7f\u7528 <code>df.where(&quot;some_column IS NULL&quot;)<\/code> \u4e4b\u5916\uff0c\u6bd4\u8f03\u4e0d\u540c\u6b04\u4f4d\u7684 <code>count<\/code> \u5dee\u7570\u4e5f\u53ef\u4ee5\u5f88\u5feb\u5730\u767c\u73fe\u54ea\u4e9b\u6b04\u4f4d\u6709\u7f3a\u5931\u503c\u3002\u9806\u4fbf\u89c0\u5bdf\u4e00\u4e0b\u6709\u7f3a\u5931\u503c\u7684\u6b04\u4f4d\u548c target variable \u6709\u6c92\u6709\u4ec0\u9ebc\u95dc\u806f\u3002<\/p>\n<p>\u9019\u88e1\u76f4\u63a5\u5c0d <code>null<\/code> \u548c <code>NaN<\/code> \u6578\u64da\u586b\u5145\u7f3a\u5931\u503c\uff0c\u56e0\u70ba\u4ee5\u4e0b\u5e7e\u500b\u6b04\u4f4d\u90fd\u662f\u5b57\u4e32\u985e\u578b\uff0c\u6240\u4ee5\u76f4\u63a5\u6539\u6210\u7a7a\u5b57\u4e32\uff0c\u65b9\u4fbf\u5f8c\u7e8c\u7684\u8655\u7406\u3002\u7136\u5f8c\u9806\u4fbf\u505a\u4e00\u500b <code>has_null<\/code> \u7684\u7279\u5fb5\u3002<\/p>\n<p>\u91dd\u5c0d user\uff1a<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.sql.functions._\n\nval nullableColumnNames = Array(\"user_name\", \"user_company\", \"user_blog\", \"user_location\", \"user_bio\")\n\nval imputedUserInfoDF = rawUserInfoDS\n  .withColumn(\"user_has_null\", when(nullableColumnNames.map(rawUserInfoDS(_).isNull).reduce(_ || _), true).otherwise(false))\n  .na.fill(\"\", nullableColumnNames)<\/code><\/pre>\n<p>\u91dd\u5c0d repo\uff1a<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.sql.functions._\n\nval nullableColumnNames = Array(\"repo_description\", \"repo_homepage\")\n\nval imputedRepoInfoDF = rawRepoInfoDS\n  .withColumn(\"repo_has_null\", when(nullableColumnNames.map(rawRepoInfoDS(_).isNull).reduce(_ || _), true).otherwise(false))\n  .na.fill(\"\", nullableColumnNames)<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"https:\/\/spark.apache.org\/docs\/latest\/api\/scala\/index.html#org.apache.spark.sql.DataFrameNaFunctions\">https:\/\/spark.apache.org\/docs\/latest\/api\/scala\/index.html#org.apache.spark.sql.DataFrameNaFunctions<\/a><\/p>\n<p>\u5982\u679c\u662f\u6578\u503c\u985e\u578b\u7684\u6b04\u4f4d\u53ef\u4ee5\u8003\u616e\u4f7f\u7528 <code>Imputer<\/code>\u3002<\/p>\n<p>ref:<br \/>\n<a href=\"https:\/\/spark.apache.org\/docs\/latest\/ml-features.html#imputer\">https:\/\/spark.apache.org\/docs\/latest\/ml-features.html#imputer<\/a><\/p>\n<h3>Clean Data<\/h3>\n<p>\u91dd\u5c0d user\uff0c\u7528 User-defined Function \u5c0d\u5e7e\u500b\u6587\u5b57\u6b04\u4f4d\u505a\u4e00\u4e9b\u6b63\u898f\u5316\u7684\u8655\u7406\uff1a<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.sql.functions._\n\nval cleanUserInfoDF = imputedUserInfoDF\n  .withColumn(\"user_clean_company\", cleanCompanyUDF($\"user_company\"))\n  .withColumn(\"user_clean_location\", cleanLocationUDF($\"user_location\"))\n  .withColumn(\"user_clean_bio\", lower($\"user_bio\"))<\/code><\/pre>\n<p>\u91dd\u5c0d repo\uff0c\u904e\u6ffe\u6389\u4e00\u4e9b <code>repo_stargazers_count<\/code> \u592a\u591a\u548c\u592a\u5c11\u3001<code>description<\/code> \u6b04\u4f4d\u542b\u6709 &quot;unmaintained&quot; \u6216 &quot;assignment&quot; \u7b49\u5b57\u773c\u7684\u9805\u76ee\uff1a<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">val reducedRepoInfo = imputedRepoInfoDF\n  .where($\"repo_is_fork\" === false)\n  .where($\"repo_forks_count\" &lt;= 90000)\n  .where($\"repo_stargazers_count\".between(30, 100000))\n\nval unmaintainedWords = Array(\"%unmaintained%\", \"%no longer maintained%\", \"%deprecated%\", \"%moved to%\")\nval assignmentWords = Array(\"%assignment%\", \"%\u4f5c\u696d%\", \"%\u4f5c\u4e1a%\")\nval demoWords = Array(\"test\", \"%demo project%\")\nval blogWords = Array(\"my blog\")\n\nval cleanRepoInfoDF = reducedRepoInfo\n  .withColumn(\"repo_clean_description\", lower($\"repo_description\"))\n  .withColumn(\"repo_is_unmaintained\", when(unmaintainedWords.map($\"repo_clean_description\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"repo_is_assignment\", when(assignmentWords.map($\"repo_clean_description\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"repo_is_demo\", when(demoWords.map($\"repo_clean_description\".like(_)).reduce(_ or _) and $\"repo_stargazers_count\" &lt;= 40, true).otherwise(false))\n  .withColumn(\"repo_is_blog\", when(blogWords.map($\"repo_clean_description\".like(_)).reduce(_ or _) and $\"repo_stargazers_count\" &lt;= 40, true).otherwise(false))\n  .where($\"repo_is_unmaintained\" === false)\n  .where($\"repo_is_assignment\" === false)\n  .where($\"repo_is_demo\" === false)\n  .where($\"repo_is_blog\" === false)\n  .withColumn(\"repo_clean_language\", lower($\"repo_language\"))\n  .withColumn(\"repo_clean_topics\", lower($\"repo_topics\"))<\/code><\/pre>\n<h3>Construct Features<\/h3>\n<p>\u91dd\u5c0d user\uff0c\u6839\u64da\u4e0a\u8ff0\u7684\u300c\u767c\u60f3\u7279\u5fb5\u300d\uff0c\u88fd\u4f5c\u51fa\u65b0\u7684\u7279\u5fb5\uff1a<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.sql.expressions.Window\nimport org.apache.spark.sql.functions._\n\nval webThings = Array(\"web\", \"fullstack\", \"full stack\")\nval backendThings = Array(\"backend\", \"back end\", \"back-end\")\nval frontendThings = Array(\"frontend\", \"front end\", \"front-end\")\nval mobileThings = Array(\"mobile\", \"ios\", \"android\")\nval devopsThings = Array(\"devops\", \"sre\", \"admin\", \"infrastructure\")\nval dataThings = Array(\"machine learning\", \"deep learning\", \"data scien\", \"data analy\")\nval recsysThings = Array(\"data mining\", \"recommend\", \"information retrieval\")\n\nval leadTitles = Array(\"team lead\", \"architect\", \"creator\", \"director\", \"cto\", \"vp of engineering\")\nval scholarTitles = Array(\"researcher\", \"scientist\", \"phd\", \"professor\")\nval freelancerTitles = Array(\"freelance\")\nval juniorTitles = Array(\"junior\", \"beginner\", \"newbie\")\nval pmTitles = Array(\"product manager\")\n\nval userStarredReposCountDF = rawStarringDS\n  .groupBy($\"user_id\")\n  .agg(count(\"*\").alias(\"user_starred_repos_count\"))\n\nval starringRepoInfoDF = rawStarringDS\n  .select($\"user_id\", $\"repo_id\", $\"starred_at\")\n  .join(rawRepoInfoDS, Seq(\"repo_id\"))\n\nval userTopLanguagesDF = starringRepoInfoDF\n  .withColumn(\"rank\", rank.over(Window.partitionBy($\"user_id\").orderBy($\"starred_at\".desc)))\n  .where($\"rank\" &lt;= 50)\n  .groupBy($\"user_id\")\n  .agg(collect_list(lower($\"repo_language\")).alias(\"user_recent_repo_languages\"))\n  .select($\"user_id\", $\"user_recent_repo_languages\")\n\nval userTopTopicsDF = starringRepoInfoDF\n  .where($\"repo_topics\" =!= \"\")\n  .withColumn(\"rank\", rank.over(Window.partitionBy($\"user_id\").orderBy($\"starred_at\".desc)))\n  .where($\"rank\" &lt;= 50)\n  .groupBy($\"user_id\")\n  .agg(concat_ws(\",\", collect_list(lower($\"repo_topics\"))).alias(\"temp_user_recent_repo_topics\"))\n  .select($\"user_id\", split($\"temp_user_recent_repo_topics\", \",\").alias(\"user_recent_repo_topics\"))\n\nval userTopDescriptionDF = starringRepoInfoDF\n  .where($\"repo_description\" =!= \"\")\n  .withColumn(\"rank\", rank.over(Window.partitionBy($\"user_id\").orderBy($\"starred_at\".desc)))\n  .where($\"rank\" &lt;= 50)\n  .groupBy($\"user_id\")\n  .agg(concat_ws(\" \", collect_list(lower($\"repo_description\"))).alias(\"user_recent_repo_descriptions\"))\n  .select($\"user_id\", $\"user_recent_repo_descriptions\")\n\nval constructedUserInfoDF = cleanUserInfoDF\n  .withColumn(\"user_knows_web\", when(webThings.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_knows_backend\", when(backendThings.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_knows_frontend\", when(frontendThings.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_knows_mobile\", when(mobileThings.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_knows_devops\", when(devopsThings.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_knows_data\", when(dataThings.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_knows_recsys\", when(recsysThings.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_is_lead\", when(leadTitles.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_is_scholar\", when(scholarTitles.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_is_freelancer\", when(freelancerTitles.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_is_junior\", when(juniorTitles.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_is_pm\", when(pmTitles.map($\"user_clean_bio\".like(_)).reduce(_ or _), true).otherwise(false))\n  .withColumn(\"user_followers_following_ratio\", round($\"user_followers_count\" \/ ($\"user_following_count\" + lit(1.0)), 3))\n  .withColumn(\"user_days_between_created_at_today\", datediff(current_date(), $\"user_created_at\"))\n  .withColumn(\"user_days_between_updated_at_today\", datediff(current_date(), $\"user_updated_at\"))\n  .join(userStarredReposCountDF, Seq(\"user_id\"))\n  .withColumn(\"user_avg_daily_starred_repos_count\", round($\"user_starred_repos_count\" \/ ($\"user_days_between_created_at_today\" + lit(1.0)), 3))\n  .join(userTopDescriptionDF, Seq(\"user_id\"))\n  .join(userTopTopicsDF, Seq(\"user_id\"))\n  .join(userTopLanguagesDF, Seq(\"user_id\"))<\/code><\/pre>\n<p>\u91dd\u5c0d repo\uff0c\u6839\u64da\u4e0a\u8ff0\u7684\u300c\u767c\u60f3\u7279\u5fb5\u300d\uff0c\u88fd\u4f5c\u51fa\u65b0\u7684\u7279\u5fb5\uff0c\u610f\u601d\u5230\u4e86\u5c31\u597d\uff1a<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.sql.functions._\n\nval vintaStarredRepos = rawStarringDS\n  .where($\"user_id\" === 652070)\n  .select($\"repo_id\".as[Int])\n  .collect()\n  .to[List]\n\nval constructedRepoInfoDF = cleanRepoInfoDF\n  .withColumn(\"repo_has_activities_in_60days\", datediff(current_date(), $\"repo_pushed_at\") &lt;= 60)\n  .withColumn(\"repo_has_homepage\", when($\"repo_homepage\" === \"\", false).otherwise(true))\n  .withColumn(\"repo_is_vinta_starred\", when($\"repo_id\".isin(vintaStarredRepos: _*), true).otherwise(false))\n  .withColumn(\"repo_days_between_created_at_today\", datediff(current_date(), $\"repo_created_at\"))\n  .withColumn(\"repo_days_between_updated_at_today\", datediff(current_date(), $\"repo_updated_at\"))\n  .withColumn(\"repo_days_between_pushed_at_today\", datediff(current_date(), $\"repo_pushed_at\"))\n  .withColumn(\"repo_subscribers_stargazers_ratio\", round($\"repo_subscribers_count\" \/ ($\"repo_stargazers_count\" + lit(1.0)), 3))\n  .withColumn(\"repo_forks_stargazers_ratio\", round($\"repo_forks_count\" \/ ($\"repo_stargazers_count\" + lit(1.0)), 3))\n  .withColumn(\"repo_open_issues_stargazers_ratio\", round($\"repo_open_issues_count\" \/ ($\"repo_stargazers_count\" + lit(1.0)), 3))\n  .withColumn(\"repo_text\", lower(concat_ws(\" \", $\"repo_owner_username\", $\"repo_name\", $\"repo_language\", $\"repo_description\")))<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"https:\/\/databricks.com\/blog\/2015\/09\/16\/apache-spark-1-5-dataframe-api-highlights.html\">https:\/\/databricks.com\/blog\/2015\/09\/16\/apache-spark-1-5-dataframe-api-highlights.html<\/a><\/p>\n<h3>Convert Features<\/h3>\n<p>\u91dd\u5c0d user\uff0c\u9019\u88e1\u4e3b\u8981\u662f\u5c0d\u4e00\u4e9b categorical \u7279\u5fb5\u4f5c binning\uff1a<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.sql.functions._\n\nval companyCountDF = cleanUserInfoDF\n  .groupBy($\"user_clean_company\")\n  .agg(count(\"*\").alias(\"count_per_user_company\"))\n\nval locationCountDF = cleanUserInfoDF\n  .groupBy($\"user_clean_location\")\n  .agg(count(\"*\").alias(\"count_per_user_location\"))\n\nval transformedUserInfoDF = constructedUserInfoDF\n  .join(companyCountDF, Seq(\"user_clean_company\"))\n  .join(locationCountDF, Seq(\"user_clean_location\"))\n  .withColumn(\"user_has_blog\", when($\"user_blog\" === \"\", 0.0).otherwise(1.0))\n  .withColumn(\"user_binned_company\", when($\"count_per_user_company\" &lt;= 5, \"__other\").otherwise($\"user_clean_company\"))\n  .withColumn(\"user_binned_location\", when($\"count_per_user_location\" &lt;= 50, \"__other\").otherwise($\"user_clean_location\"))<\/code><\/pre>\n<p>\u91dd\u5c0d repo\uff1a<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.sql.functions._\n\nval languagesDF = cleanRepoInfoDF\n  .groupBy($\"repo_clean_language\")\n  .agg(count(\"*\").alias(\"count_per_repo_language\"))\n  .select($\"repo_clean_language\", $\"count_per_repo_language\")\n  .cache()\n\nval transformedRepoInfoDF = constructedRepoInfoDF\n  .join(languagesDF, Seq(\"repo_clean_language\"))\n  .withColumn(\"repo_binned_language\", when($\"count_per_repo_language\" &lt;= 30, \"__other\").otherwise($\"repo_clean_language\"))\n  .withColumn(\"repo_clean_topics\", split($\"repo_topics\", \",\"))<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"https:\/\/docs.databricks.com\/spark\/latest\/mllib\/binary-classification-mllib-pipelines.html\">https:\/\/docs.databricks.com\/spark\/latest\/mllib\/binary-classification-mllib-pipelines.html<\/a><\/p>\n<h2>Prepare the Feature Pipeline<\/h2>\n<p>\u6211\u5011\u904e\u6ffe\u6389\u90a3\u4e9b\u6253\u661f\u4e86\u8d85\u591a repo \u7684\u7528\u6236\u3002\u5f9e\u6536\u96c6\u5230\u7684\u6578\u64da\u767c\u73fe\uff0c\u6709\u4e9b\u7528\u6236\u751a\u81f3\u6253\u661f\u4e86\u4e00\u5169\u842c\u500b repo\uff0c\u9019\u4e9b\u7528\u6236\u53ef\u80fd\u662f\u500b\u722c\u87f2\u5c08\u7528\u5e33\u865f\u6216\u662f\u4ed6\u770b\u5230\u4ec0\u9ebc\u5c31\u6253\u661f\u4ec0\u9ebc\uff0c\u63a8\u85a6\u7cfb\u7d71\u5c0d\u9019\u6a23\u7684\u7528\u6236\u4f86\u8aaa\u53ef\u80fd\u6c92\u4ec0\u9ebc\u610f\u7fa9\uff0c\u9084\u4e0d\u5982\u5f9e\u6578\u64da\u96c6\u4e2d\u62ff\u6389\u3002<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.sql.functions._\n\nval maxStarredReposCount = 2000\n\nval userStarredReposCountDF = rawStarringDS\n  .groupBy($\"user_id\")\n  .agg(count(\"*\").alias(\"user_starred_repos_count\"))\n\nval reducedStarringDF = rawStarringDS\n  .join(userStarredReposCountDF, Seq(\"user_id\"))\n  .where($\"user_starred_repos_count\" &lt;= maxStarredReposCount)\n  .select($\"user_id\", $\"repo_id\", $\"starred_at\", $\"starring\")\n\nval profileStarringDF = reducedStarringDF\n  .join(userProfileDF, Seq(\"user_id\"))\n  .join(repoProfileDF, Seq(\"repo_id\"))<\/code><\/pre>\n<h2>Build the Feature Pipeline<\/h2>\n<p>\u628a\u8655\u7406\u7279\u5fb5\u7684\u4e00\u9023\u4e32\u6d41\u7a0b\u5beb\u6210 Spark ML Pipeline\uff0c\u65b9\u4fbf\u62bd\u63db\u6216\u662f\u52a0\u5165\u65b0\u7684 <code>Transformer<\/code>\uff0c\u4f8b\u5982 Standardization\u3001One-hot Encoding \u548c Word2Vec\uff0c\u4e5f\u628a ALS \u6a21\u578b\u7684\u9810\u6e2c\u503c\u505a\u70ba\u5176\u4e2d\u4e00\u9805\u7279\u5fb5\u3002<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.ml.feature._\nimport org.apache.spark.ml.recommendation.ALSModel\nimport ws.vinta.albedo.transformers.UserRepoTransformer\n\nval profileStarringDF = reducedStarringDF\n  .join(userProfileDF, Seq(\"user_id\"))\n  .join(repoProfileDF, Seq(\"repo_id\"))\n  .cache()\n\ncategoricalColumnNames += \"user_id\"\ncategoricalColumnNames += \"repo_id\"\n\nval userRepoTransformer = new UserRepoTransformer()\n  .setInputCols(Array(\"repo_language\", \"user_recent_repo_languages\"))\n\ncontinuousColumnNames += \"repo_language_index_in_user_recent_repo_languages\"\ncontinuousColumnNames += \"repo_language_count_in_user_recent_repo_languages\"\n\nval alsModelPath = s\"${settings.dataDir}\/${settings.today}\/alsModel.parquet\"\nval alsModel = ALSModel.load(alsModelPath)\n  .setUserCol(\"user_id\")\n  .setItemCol(\"repo_id\")\n  .setPredictionCol(\"als_score\")\n  .setColdStartStrategy(\"drop\")\n\ncontinuousColumnNames += \"als_score\"\n\nval categoricalTransformers = categoricalColumnNames.flatMap((columnName: String) =&gt; {\n  val stringIndexer = new StringIndexer()\n    .setInputCol(columnName)\n    .setOutputCol(s\"${columnName}__idx\")\n    .setHandleInvalid(\"keep\")\n\n  val oneHotEncoder = new OneHotEncoder()\n    .setInputCol(s\"${columnName}__idx\")\n    .setOutputCol(s\"${columnName}__ohe\")\n    .setDropLast(false)\n\n  Array(stringIndexer, oneHotEncoder)\n})\n\nval listTransformers = listColumnNames.flatMap((columnName: String) =&gt; {\n  val countVectorizerModel = new CountVectorizer()\n    .setInputCol(columnName)\n    .setOutputCol(s\"${columnName}__cv\")\n    .setMinDF(10)\n    .setMinTF(1)\n\n  Array(countVectorizerModel)\n})\n\nval textTransformers = textColumnNames.flatMap((columnName: String) =&gt; {\n  val hanLPTokenizer = new HanLPTokenizer()\n    .setInputCol(columnName)\n    .setOutputCol(s\"${columnName}__words\")\n    .setShouldRemoveStopWords(true)\n\n  val stopWordsRemover = new StopWordsRemover()\n    .setInputCol(s\"${columnName}__words\")\n    .setOutputCol(s\"${columnName}__filtered_words\")\n    .setStopWords(StopWordsRemover.loadDefaultStopWords(\"english\"))\n  val word2VecModelPath = s\"${settings.dataDir}\/${settings.today}\/word2VecModel.parquet\"\n  val word2VecModel = Word2VecModel.load(word2VecModelPath)\n    .setInputCol(s\"${columnName}__filtered_words\")\n    .setOutputCol(s\"${columnName}__w2v\")\n\n  Array(hanLPTokenizer, stopWordsRemover, word2VecModel)\n})\n\nval finalBooleanColumnNames = booleanColumnNames.toArray\nval finalContinuousColumnNames = continuousColumnNames.toArray\nval finalCategoricalColumnNames = categoricalColumnNames.map(columnName =&gt; s\"${columnName}__ohe\").toArray\nval finalListColumnNames = listColumnNames.map(columnName =&gt; s\"${columnName}__cv\").toArray\nval finalTextColumnNames = textColumnNames.map(columnName =&gt; s\"${columnName}__w2v\").toArray\nval vectorAssembler = new SimpleVectorAssembler()\n  .setInputCols(finalBooleanColumnNames ++ finalContinuousColumnNames ++ finalCategoricalColumnNames ++ finalListColumnNames ++ finalTextColumnNames)\n  .setOutputCol(\"features\")\n\nval featureStages = mutable.ArrayBuffer.empty[PipelineStage]\nfeatureStages += userRepoTransformer\nfeatureStages += alsModel\nfeatureStages ++= categoricalTransformers\nfeatureStages ++= listTransformers\nfeatureStages ++= textTransformers\nfeatureStages += vectorAssembler\n\nval featurePipeline = new Pipeline().setStages(featureStages.toArray)\nval featurePipelineModel = featurePipeline.fit(profileStarringDF)<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"https:\/\/spark.apache.org\/docs\/latest\/ml-pipeline.html\">https:\/\/spark.apache.org\/docs\/latest\/ml-pipeline.html<\/a><br \/>\n<a href=\"https:\/\/spark.apache.org\/docs\/latest\/ml-features.html\">https:\/\/spark.apache.org\/docs\/latest\/ml-features.html<\/a><\/p>\n<h2>Handle Imbalanced Data<\/h2>\n<p>\u56e0\u70ba\u6211\u5011\u8981\u8a13\u7df4\u4e00\u500b Binary Classification \u4e8c\u5143\u5206\u985e\u6a21\u578b\uff0c\u6703\u540c\u6642\u9700\u8981 positive\uff08\u6b63\u6a23\u672c\uff09\u548c negative\uff08\u8ca0\u6a23\u672c\uff09\u3002\u4f46\u662f\u6211\u5011\u7684\u539f\u59cb\u6578\u64da <code>rawStarringDS<\/code> \u90fd\u662f\u6b63\u6a23\u672c\uff0c\u4e5f\u5c31\u662f\u8aaa\u6211\u5011\u53ea\u6709\u300c\u7528\u6236\u6709\u5c0d\u54ea\u4e9b repo \u6253\u661f\u7684\u8cc7\u6599\u300d\uff08\u6b63\u6a23\u672c\uff09\uff0c\u537b\u6c92\u6709\u300c\u7528\u6236\u6c92\u6709\u5c0d\u54ea\u4e9b repo \u6253\u661f\u7684\u8cc7\u6599\u300d\uff08\u8ca0\u6a23\u672c\uff09\u3002\u6211\u5011\u7576\u7136\u662f\u53ef\u4ee5\u7528\u300c\u6240\u6709\u7528\u6236\u6c92\u6709\u6253\u661f\u7684 repo \u505a\u70ba\u8ca0\u6a23\u672c\u300d\uff0c\u4f46\u662f\u8003\u616e\u5230\u9019\u7a2e\u505a\u6cd5\u7522\u751f\u7684\u8ca0\u6a23\u672c\u7684\u6578\u91cf\u5be6\u5728\u592a\u5927\uff0c\u800c\u4e14\u4e5f\u4e0d\u592a\u5408\u7406\uff0c\u56e0\u70ba\u90a3\u4e9b\u7528\u6236\u6c92\u6709\u6253\u661f\u7684 repo \u4e0d\u898b\u5f97\u662f\u56e0\u70ba\u4ed6\u4e0d\u559c\u6b61\uff0c\u53ef\u80fd\u53ea\u662f\u56e0\u70ba\u4ed6\u4e0d\u77e5\u9053\u6709\u90a3\u500b repo \u5b58\u5728\u3002<\/p>\n<p>\u6211\u5011\u5f8c\u4f86\u63a1\u7528\u7684\u505a\u6cd5\u662f\u300c\u7528\u71b1\u9580\u4f46\u662f\u7528\u6236\u6c92\u6709\u6253\u661f\u7684 repo \u505a\u70ba\u8ca0\u6a23\u672c\u300d\uff0c\u6211\u5011\u5beb\u4e86\u4e00\u500b Spark Transformer \u4f86\u505a\u9019\u4ef6\u4e8b\uff1a<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import ws.vinta.albedo.transformers.NegativeBalancer\n\nimport scala.collection.mutable\n\nval sc = spark.sparkContext\n\nval popularReposDS = loadPopularRepoDF()\nval popularRepos = popularReposDS\n  .select($\"repo_id\".as[Int])\n  .collect()\n  .to[mutable.LinkedHashSet]\nval bcPopularRepos = sc.broadcast(popularRepos)\n\nval negativeBalancer = new NegativeBalancer(bcPopularRepos)\n  .setUserCol(\"user_id\")\n  .setItemCol(\"repo_id\")\n  .setTimeCol(\"starred_at\")\n  .setLabelCol(\"starring\")\n  .setNegativeValue(0.0)\n  .setNegativePositiveRatio(2.0)\nval balancedStarringDF = negativeBalancer.transform(reducedStarringDF)<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"https:\/\/github.com\/vinta\/albedo\/blob\/master\/src\/main\/scala\/ws\/vinta\/albedo\/evaluators\/RankingEvaluator.scala\">https:\/\/github.com\/vinta\/albedo\/blob\/master\/src\/main\/scala\/ws\/vinta\/albedo\/evaluators\/RankingEvaluator.scala<\/a><br \/>\n<a href=\"http:\/\/www.kdnuggets.com\/2017\/06\/7-techniques-handle-imbalanced-data.html\">http:\/\/www.kdnuggets.com\/2017\/06\/7-techniques-handle-imbalanced-data.html<\/a><\/p>\n<h2>Split Data<\/h2>\n<p>\u76f4\u63a5\u4f7f\u7528 holdout \u7684\u65b9\u5f0f\uff0c\u96a8\u6a5f\u5206\u914d\u4e0d\u540c\u7684 row \u5230 training set \u548c test set\u3002\u5176\u4ed6\u7684\u505a\u6cd5\u53ef\u80fd\u662f\u6839\u64da\u6642\u9593\u4f86\u62c6\u5206\uff0c\u7528\u4ee5\u524d\u7684\u6578\u64da\u4f86\u9810\u6e2c\u4e4b\u5f8c\u7684\u884c\u70ba\u3002<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">val profileBalancedStarringDF = balancedStarringDF\n  .join(userProfileDF, Seq(\"user_id\"))\n  .join(repoProfileDF, Seq(\"repo_id\"))\n\nval tmpDF = featurePipelineModel.transform(profileBalancedStarringDF)\nval keepColumnName = tmpDF.columns.filter((columnName: String) =&gt; {\n  !columnName.endsWith(\"__idx\") &amp;&amp;\n  !columnName.endsWith(\"__ohe\") &amp;&amp;\n  !columnName.endsWith(\"__cv\") &amp;&amp;\n  !columnName.endsWith(\"__words\") &amp;&amp;\n  !columnName.endsWith(\"__filtered_words\") &amp;&amp;\n  !columnName.endsWith(\"__w2v\")\n})\nval featuredBalancedStarringDF = tmpDF.select(keepColumnName.map(col): _*)\n\nval Array(trainingFeaturedDF, testFeaturedDF) = featuredBalancedStarringDF.randomSplit(Array(0.9, 0.1))<\/code><\/pre>\n<h2>Build the Model Pipeline<\/h2>\n<p>\u70ba\u4e86\u65b9\u4fbf\u4e4b\u5f8c\u7684\u64f4\u5145\u6027\uff0c\u9019\u88e1\u4e5f\u4f7f\u7528 Spark ML Pipeline \u7684\u5beb\u6cd5\u3002Spark ML \u7684 <code>LogisticRegression<\/code> \u53ef\u4ee5\u984d\u5916\u8a2d\u7f6e\u4e00\u500b <code>weightCol<\/code> \u4f86\u8abf\u6574\u4e0d\u540c row \u7684\u6b0a\u91cd\u3002<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.ml.classification.LogisticRegression\nimport org.apache.spark.ml.{Pipeline, PipelineStage}\n\nimport scala.collection.mutable\n\nval weightSQL = \"\"\"\nSELECT *,\n       1.0 AS default_weight,\n       IF (starring = 1.0, 0.9, 0.1) AS positive_weight,\n       IF (starring = 1.0 AND datediff(current_date(), starred_at) &lt;= 365, 0.9, 0.1) AS recent_starred_weight\nFROM __THIS__\n\"\"\".stripMargin\nval weightTransformer = new SQLTransformer()\n  .setStatement(weightSQL)\n\nval lr = new LogisticRegression()\n  .setMaxIter(200)\n  .setRegParam(0.7)\n  .setElasticNetParam(0.0)\n  .setStandardization(true)\n  .setLabelCol(\"starring\")\n  .setFeaturesCol(\"standard_features\")\n  .setWeightCol(\"recent_starred_weight\")\n\nval modelStages = mutable.ArrayBuffer.empty[PipelineStage]\nmodelStages += weightTransformer\nmodelStages += lr\n\nval modelPipeline = new Pipeline().setStages(modelStages.toArray)\nval modelPipelineModel = modelPipeline.fit(trainingFeaturedDF)<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"https:\/\/spark.apache.org\/docs\/latest\/ml-classification-regression.html\">https:\/\/spark.apache.org\/docs\/latest\/ml-classification-regression.html<\/a><\/p>\n<h2>Evaluate the Model: Classification<\/h2>\n<p>\u56e0\u70ba Logistic Regression \u662f\u4e8c\u5143\u5206\u985e\u6a21\u578b\uff0c\u6240\u4ee5\u6211\u5011\u53ef\u4ee5\u7528 Spark ML \u7684 <code>BinaryClassificationEvaluator<\/code> \u4f86\u8a55\u4f30\u7d50\u679c\u3002\u4e0d\u904e\u56e0\u70ba\u6211\u5011\u505a\u7684\u662f\u63a8\u85a6\u7cfb\u7d71\uff0c\u771f\u6b63\u5728\u4e4e\u7684\u662f Top N \u7684\u6392\u5e8f\u554f\u984c\uff0c\u6240\u4ee5\u9019\u88e1\u7684 AUC \u7684\u6578\u503c\u53c3\u8003\u4e00\u4e0b\u5c31\u597d\u3002<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.ml.evaluation.BinaryClassificationEvaluator\n\nval testRankedDF = modelPipelineModel.transform(testFeaturedDF)\n\nval binaryClassificationEvaluator = new BinaryClassificationEvaluator()\n  .setMetricName(\"areaUnderROC\")\n  .setRawPredictionCol(\"rawPrediction\")\n  .setLabelCol(\"starring\")\n\nval classificationMetric = binaryClassificationEvaluator.evaluate(testRankedDF)\nprintln(s\"${binaryClassificationEvaluator.getMetricName} = $classificationMetric\")\n\/\/ areaUnderROC = 0.9450631491281277<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"https:\/\/spark.apache.org\/docs\/latest\/api\/scala\/index.html#org.apache.spark.ml.evaluation.BinaryClassificationEvaluator\">https:\/\/spark.apache.org\/docs\/latest\/api\/scala\/index.html#org.apache.spark.ml.evaluation.BinaryClassificationEvaluator<\/a><br \/>\n<a href=\"https:\/\/docs.databricks.com\/spark\/latest\/mllib\/binary-classification-mllib-pipelines.html\">https:\/\/docs.databricks.com\/spark\/latest\/mllib\/binary-classification-mllib-pipelines.html<\/a><\/p>\n<h2>Generate Candidates<\/h2>\n<p>\u63a8\u85a6\u7cfb\u7d71\u7684\u53e6\u5916\u4e00\u500b\u91cd\u8981\u90e8\u5206\u5c31\u662f\u7522\u751f\u5019\u9078\u7269\u54c1\u96c6\uff0c\u9019\u88e1\u6211\u5011\u4f7f\u7528\u4ee5\u4e0b\u5e7e\u7a2e\u65b9\u5f0f\uff1a<\/p>\n<ul>\n<li>ALS: \u5354\u540c\u904e\u6ffe\u7684\u63a8\u85a6<\/li>\n<li>Content-based: \u57fa\u65bc\u5167\u5bb9\u7684\u63a8\u85a6<\/li>\n<li>Popularity: \u57fa\u65bc\u71b1\u9580\u7684\u63a8\u85a6<\/li>\n<\/ul>\n<p>\u4e0d\u904e\u56e0\u70ba\u9019\u7bc7\u6587\u7ae0\u7684\u4e3b\u984c\u662f\u6392\u5e8f\u548c\u7279\u5fb5\u5de5\u7a0b\u7684 Machine Learning Pipeline\uff0c\u6240\u4ee5\u7522\u751f\u5019\u9078\u7269\u54c1\u96c6\u7684\u90e8\u5206\u5c31\u4e0d\u591a\u8aaa\u4e86\uff0c\u6709\u8208\u8da3\u7684\u4eba\u53ef\u4ee5\u76f4\u63a5\u770b\u5e95\u4e0b\u9023\u7d50\u7684 source code \u6216\u662f\u9019\u500b\u7cfb\u5217\u7684\u5176\u4ed6\u6587\u7ae0\u3002<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import ws.vinta.albedo.recommenders.ALSRecommender\nimport ws.vinta.albedo.recommenders.ContentRecommender\nimport ws.vinta.albedo.recommenders.PopularityRecommender\n\nval topK = 30\n\nval alsRecommender = new ALSRecommender()\n  .setUserCol(\"user_id\")\n  .setItemCol(\"repo_id\")\n  .setTopK(topK)\n\nval contentRecommender = new ContentRecommender()\n  .setUserCol(\"user_id\")\n  .setItemCol(\"repo_id\")\n  .setTopK(topK)\n  .setEnableEvaluationMode(true)\n\nval popularityRecommender = new PopularityRecommender()\n  .setUserCol(\"user_id\")\n  .setItemCol(\"repo_id\")\n  .setTopK(topK)\n\nval recommenders = mutable.ArrayBuffer.empty[Recommender]\nrecommenders += alsRecommender\nrecommenders += contentRecommender\nrecommenders += popularityRecommender\n\nval candidateDF = recommenders\n  .map((recommender: Recommender) =&gt; recommender.recommendForUsers(testUserDF))\n  .reduce(_ union _)\n  .select($\"user_id\", $\"repo_id\")\n  .distinct()\n\n\/\/ \u6bcf\u500b Recommender \u7684\u7d50\u679c\u985e\u4f3c\u9019\u6a23\uff1a\n\/\/ +-------+-------+----------+------+\n\/\/ |user_id|repo_id|score     |source|\n\/\/ +-------+-------+----------+------+\n\/\/ |652070 |1239728|0.6731846 |als   |\n\/\/ |652070 |854078 |0.7187486 |als   |\n\/\/ |652070 |1502338|0.70165294|als   |\n\/\/ |652070 |1184678|0.7434903 |als   |\n\/\/ |652070 |547708 |0.7956538 |als   |\n\/\/ +-------+-------+----------+------+<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"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\/ALSRecommender.scala<\/a><br \/>\n<a href=\"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\/ContentRecommender.scala<\/a><br \/>\n<a href=\"https:\/\/github.com\/vinta\/albedo\/blob\/master\/src\/main\/scala\/ws\/vinta\/albedo\/recommenders\/PopularityRecommender.scala\">https:\/\/github.com\/vinta\/albedo\/blob\/master\/src\/main\/scala\/ws\/vinta\/albedo\/recommenders\/PopularityRecommender.scala<\/a><\/p>\n<h2>Predict the Ranking<\/h2>\n<p>\u628a\u9019\u4e9b\u5019\u9078\u7269\u54c1\u96c6\u4e1f\u7d66\u6211\u5011\u8a13\u7df4\u597d\u7684 Logistic Regression \u6a21\u578b\u4f86\u6392\u5e8f\u3002\u7d50\u679c\u4e2d\u7684 <code>probability<\/code> \u6b04\u4f4d\u7684\u7b2c 0 \u9805\u8868\u793a\u7d50\u679c\u70ba 0 \u7684\u6a5f\u7387\uff08negative\uff09\u3001\u7b2c 1 \u9805\u8868\u793a\u7d50\u679c\u70ba 1 \u7684\u6a5f\u7387\uff08positive\uff09\u3002<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">val profileCandidateDF = candidateDF\n  .join(userProfileDF, Seq(\"user_id\"))\n  .join(repoProfileDF, Seq(\"repo_id\"))\n\nval featuredCandidateDF = featurePipelineModel\n  .transform(profileCandidateDF)\n\nval rankedCandidateDF = modelPipelineModel\n  .transform(featuredCandidateDF)\n\n\/\/ rankedCandidateDF \u7684\u7d50\u679c\u985e\u4f3c\u9019\u6a23\uff1a\n\/\/ +-------+--------+----------+----------------------------------------+\n\/\/ |user_id|repo_id |prediction|probability                             |\n\/\/ +-------+--------+----------+----------------------------------------+\n\/\/ |652070 |83467664|1.0       |[0.12711894229094317,0.8728810577090568]|\n\/\/ |652070 |55099616|1.0       |[0.1422859437320775,0.8577140562679224] |\n\/\/ |652070 |42266235|1.0       |[0.1462014853157966,0.8537985146842034] |\n\/\/ |652070 |78012800|1.0       |[0.15576081067098502,0.844239189329015] |\n\/\/ |652070 |5928761 |1.0       |[0.16149848941925066,0.8385015105807493]|\n\/\/ +-------+--------+----------+----------------------------------------+<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"https:\/\/stackoverflow.com\/questions\/37903288\/what-do-colum-rawprediction-and-probability-of-dataframe-mean-in-spark-mllib\">https:\/\/stackoverflow.com\/questions\/37903288\/what-do-colum-rawprediction-and-probability-of-dataframe-mean-in-spark-mllib<\/a><\/p>\n<h2>Evaluate the Model: Ranking<\/h2>\n<p>\u6700\u5f8c\u6211\u5011\u4f7f\u7528 Information Retrieval \u9818\u57df\u4e2d\u7528\u4f86\u8a55\u50f9\u6392\u5e8f\u80fd\u529b\u7684\u6307\u6a19 NDCG (Normalized Discounted Cumulative Gain) \u4f86\u8a55\u4f30\u6392\u5e8f\u7684\u7d50\u679c\u3002Spark MLlib \u6709\u73fe\u6210\u7684 <code>RankingMetrics<\/code> \u53ef\u4ee5\u7528\uff0c\u4f46\u662f\u5b83\u53ea\u9069\u7528\u65bc RDD-based \u7684 API\uff0c\u6240\u4ee5\u6211\u5011\u6539\u5beb\u6210\u9069\u5408 DataFrame-based \u7684 <code>Evaluator<\/code>\u3002<\/p>\n<pre class=\"line-numbers\"><code class=\"language-scala\">import org.apache.spark.sql.expressions.Window\nimport org.apache.spark.sql.functions._\nimport ws.vinta.albedo.evaluators.RankingEvaluator\n\nval userActualItemsDF = reducedStarringDF\n  .withColumn(\"rank\", rank().over(Window.partitionBy($\"user_id\").orderBy($\"starred_at\".desc)))\n  .where($\"rank\" &lt;= topK)\n  .groupBy($\"user_id\")\n  .agg(collect_list($\"repo_id\").alias(\"items\"))\n\nval userPredictedItemsDF = rankedCandidateDF\n  .withColumn(\"rank\", rank().over(Window.partitionBy($\"user_id\").orderBy(toArrayUDF($\"probability\").getItem(1).desc)))\n  .where($\"rank\" &lt;= topK)\n  .groupBy($\"user_id\")\n  .agg(collect_list($\"repo_id\").alias(\"items\"))\n\nval rankingEvaluator = new RankingEvaluator(userActualItemsDF)\n  .setMetricName(\"NDCG@k\")\n  .setK(topK)\n  .setUserCol(\"user_id\")\n  .setItemsCol(\"items\")\nval rankingMetric = rankingEvaluator.evaluate(userPredictedItemsDF)\nprintln(s\"${rankingEvaluator.getFormattedMetricName} = $rankingMetric\")\n\/\/ NDCG@30 = 0.021114356461615493<\/code><\/pre>\n<p>ref:<br \/>\n<a href=\"https:\/\/github.com\/vinta\/albedo\/blob\/master\/src\/main\/scala\/ws\/vinta\/albedo\/evaluators\/RankingEvaluator.scala\">https:\/\/github.com\/vinta\/albedo\/blob\/master\/src\/main\/scala\/ws\/vinta\/albedo\/evaluators\/RankingEvaluator.scala<\/a><br \/>\n<a href=\"https:\/\/spark.apache.org\/docs\/latest\/mllib-evaluation-metrics.html#ranking-systems\">https:\/\/spark.apache.org\/docs\/latest\/mllib-evaluation-metrics.html#ranking-systems<\/a><br \/>\n<a href=\"https:\/\/weekly.codetengu.com\/issues\/83#kOxuVxW\">https:\/\/weekly.codetengu.com\/issues\/83#kOxuVxW<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5728\u672c\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u5011\u5c07\u4ee5 Ranking \u968e\u6bb5\u5e38\u7528\u7684\u65b9\u6cd5\u4e4b\u4e00\uff1aLogistic Regression \u908f\u8f2f\u8ff4\u6b78\u70ba\u4f8b\uff0c\u5229\u7528 Apache Spark \u7684 Logistic Regression \u6a21\u578b\u5efa\u7acb\u4e00\u500b GitHub repositories \u7684\u63a8\u85a6\u7cfb\u7d71\uff0c\u4ee5\u7528\u6236\u5c0d repo \u7684\u6253\u661f\u7d00\u9304\u548c\u7528\u6236\u8207 repo \u7684\u5404\u9805\u5c6c\u6027\u505a\u70ba\u7279\u5fb5\uff0c\u9810\u6e2c\u51fa\u7528\u6236\u6703\u4e0d\u6703\u6253\u661f\u67d0\u500b repo\uff08\u5206\u985e\u554f\u984c\uff09\u3002\u6700\u5f8c\u8a13\u7df4\u51fa\u4f86\u7684\u6a21\u578b\u5c31\u53ef\u4ee5\u505a\u70ba\u6211\u5011\u7684\u63a8\u85a6\u7cfb\u7d71\u7684 Ranking \u6a21\u7d44\u3002\u4e0d\u904e\u56e0\u70ba LR \u662f\u7dda\u6027\u6a21\u578b\uff0c\u6240\u4ee5\u901a\u5e38\u9700\u8981\u5927\u91cf\u7684 Feature Engineering \u4f86\u7fd2\u5f97\u975e\u7dda\u6027\u95dc\u4fc2\u3002\u6240\u4ee5\u9019\u7bc7\u6587\u7ae0\u7684\u91cd\u9ede\u662f Spark ML \u7684 Pipeline \u6a5f\u5236\u548c\u7279\u5fb5\u5de5\u7a0b\uff0c\u4e0d\u6703\u5728\u6f14\u7b97\u6cd5\u7684\u90e8\u5206\u8457\u58a8\u592a\u591a\u3002<\/p>\n","protected":false},"author":1,"featured_media":457,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[97,112],"tags":[108,111,98,104,109],"class_list":["post-456","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-about-ai","category-about-big-data","tag-apache-spark","tag-feature-engineering","tag-machine-learning","tag-recommender-system","tag-scala"],"_links":{"self":[{"href":"https:\/\/vinta.ws\/code\/wp-json\/wp\/v2\/posts\/456","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/vinta.ws\/code\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/vinta.ws\/code\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/vinta.ws\/code\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/vinta.ws\/code\/wp-json\/wp\/v2\/comments?post=456"}],"version-history":[{"count":0,"href":"https:\/\/vinta.ws\/code\/wp-json\/wp\/v2\/posts\/456\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/vinta.ws\/code\/wp-json\/wp\/v2\/media\/457"}],"wp:attachment":[{"href":"https:\/\/vinta.ws\/code\/wp-json\/wp\/v2\/media?parent=456"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/vinta.ws\/code\/wp-json\/wp\/v2\/categories?post=456"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/vinta.ws\/code\/wp-json\/wp\/v2\/tags?post=456"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}