![\"\"](\"https:\/\/www.growthhk.cn\/wp-content\/uploads\/mmexport1524568219004.png\")
<\/p>\n\u643a\u7a0b\u57fa\u7840\u4e1a\u52a1\u7814\u53d1\u90e8-\u6570\u636e\u4ea7\u54c1\u548c\u670d\u52a1\u7ec4\uff0c\u4e13\u6ce8\u4e8e\u4e2a\u6027\u5316\u63a8\u8350<\/a><\/span>\u3001\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u56fe\u50cf\u8bc6\u522b\u7b49\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u5148\u8fdb\u6280\u672f\u5728\u65c5\u6e38\u884c\u4e1a\u7684\u5e94\u7528\u7814\u7a76\u5e76\u843d\u5730\u4ea7\u751f\u4ef7\u503c\u3002<\/p>\n
\u4e1a\u5185\u6bd4\u8f83\u4f20\u7edf\u7684\u7b97\u6cd5\uff0c\u4e3b\u8981\u662fCF[1][2]\u3001\u57fa\u4e8e\u7edf\u8ba1\u7684 Contextual \u63a8\u8350\u548cLBS\uff0c\u4f46\u8fd1\u671f\u6765\u6df1\u5ea6\u5b66\u4e60\u88ab\u5e7f\u6cdb\u5f15\u5165\uff0c\u7b97\u6cd5\u6027\u53d6\u5f97\u8f83\u5927\u7684\u63d0\u5347\uff0c\u5982\uff1a2015\u5e74 Netflix \u548c Gravity R&D Inc \u63d0\u51fa\u7684\u5229\u7528 RNN\u7684Session-based \u63a8\u8350[5]\uff0c2016\u5e74Recsys\u4e0a\u63d0\u51fa\u7684\u7ed3\u5408 CNN \u548c PMF<\/a><\/span> \u5e94\u7528\u4e8e Context-aware \u63a8\u8350[10]\uff0c2016\u5e74 Google \u63d0\u51fa\u7684\u5c06 DNN \u4f5c\u4e3a MF \u7684\u63a8\u5e7f\uff0c\u53ef\u4ee5\u5f88\u5bb9\u6613\u5730\u5c06\u4efb\u610f\u8fde\u7eed\u548c\u5206\u7c7b\u7279\u5f81\u6dfb\u52a0\u5230\u6a21\u578b\u4e2d[9]\uff0c2017\u5e74IJCAI\u4f1a\u8bae\u4e2d\u63d0\u51fa\u7684\u5229\u7528LSTM\u8fdb\u884c\u5e8f\u5217\u63a8\u8350[6]\u3002<\/p>\n
\u4e00\u3001\u6570\u636e<\/p>\n
\u673a\u5668\u5b66\u4e60\uff1d\u6570\u636e\uff0b\u7279\u5f81\uff0b\u6a21\u578b<\/p>\n
\u4e8c\u3001\u53ec\u56de<\/p>\n
\u4e09\u3001\u6392\u5e8f<\/p>\n
\u4e8b\u5b9e\u4e0a\uff0c\u867d\u7136\u6df1\u5ea6\u5b66\u4e60\u7b49\u65b9\u6cd5\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u51cf\u5c11\u4e86\u7e41\u6742\u7684\u7279\u5f81\u5de5\u7a0b\u5de5\u4f5c\uff0c\u4f46\u6211\u4eec\u8ba4\u4e3a\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u7279\u5f81\u5de5\u7a0b\u4ecd\u65e7\u662f\u4e0d\u53ef\u6216\u7f3a\u7684, \u5176\u4e2d\u5982\u4f55\u8fdb\u884c\u7279\u5f81\u7ec4\u5408\u662f\u6211\u4eec\u5728\u5b9e\u8df5\u4e2d\u7740\u91cd\u8003\u8651\u7684\u95ee\u9898\u3002\u4e00\u822c\u7684\uff0c\u53ef\u4ee5\u5206\u4e3a\u663e\u5f0f\u7279\u5f81\u7ec4\u5408\u548c\u534a\u663e\u5f0f\u7279\u5f81\u7ec4\u5408\u3002<\/p>\n
\u663e\u5f0f\u7279\u5f81\u7ec4\u5408<\/p>\n
\u5bf9\u7279\u5f81\u8fdb\u884c\u79bb\u6563\u5316\u540e\u7136\u540e\u8fdb\u884c\u53c9\u4e58\uff0c\u91c7\u7528\u7b1b\u5361\u5c14\u79ef(cartesian product)\u3001\u5185\u79ef(inner product)\u7b49\u65b9\u5f0f\u3002<\/p>\n
\u5728\u6784\u9020\u4ea4\u53c9\u7279\u5f81\u7684\u8fc7\u7a0b\u4e2d\uff0c\u9700\u8981\u8fdb\u884c\u7279\u5f81\u79bb\u6563\u5316\uff1b\u9488\u5bf9\u4e0d\u540c\u7684\u7279\u5f81\u7c7b\u578b\uff0c\u6709\u4e0d\u540c\u7684\u5904\u7406\u65b9\u5f0f\u3002<\/p>\n
1. numerical feature<\/p>\n
\u65e0\u76d1\u7763\u79bb\u6563\u5316\uff1a\u6839\u636e\u7b80\u5355\u7edf\u8ba1\u91cf\u8fdb\u884c\u7b49\u9891\u3001\u7b49\u5bbd\u3001\u5206\u4f4d\u70b9\u7b49\u5212\u5206\u533a\u95f4<\/p>\n
\u6709\u76d1\u7763\u79bb\u6563\u5316\uff1a1R\u65b9\u6cd5\uff0cEntropy-BasedDiscretization (e.g. D2\uff0cMDLP)<\/p>\n
2. ordinal feature\uff08\u6709\u5e8f\u7279\u5f81\uff09<\/p>\n
\u7f16\u7801\u8868\u793a\u503c\u4e4b\u95f4\u7684\u987a\u5e8f\u5173\u7cfb\u3002\u6bd4\u5982\u5bf9\u4e8e\u536b\u751f\u6761\u4ef6\u8fd9\u4e00\u7279\u5f81\uff0c\u5206\u522b\u6709\u5dee\uff0c\u4e2d\uff0c\u597d\u4e09\u6863\uff0c\u90a3\u4e48\u53ef\u4ee5\u5206\u522b\u7f16\u7801\u4e3a(1,0,0),(1,1,0),(1,1,1)\u3002<\/p>\n
3. categorical feature (\u65e0\u5e8f\u7279\u5f81)<\/p>\n
![\"\"](\"https:\/\/www.growthhk.cn\/wp-content\/uploads\/Screenshot_2018-04-24-19-14-17-053_\u5fae\u4fe1.png\")
<\/p>\n
\u79bb\u6563\u5316\u4e3a\u54d1\u53d8\u91cf\uff0c\u5c06\u4e00\u7ef4\u4fe1\u606f\u5d4c\u5165\u6a21\u578b\u7684bias\u4e2d\uff0c\u8d77\u5230\u7b80\u5316\u903b\u8f91\u56de\u5f52\u6a21\u578b\u7684\u4f5c\u7528\uff0c\u964d\u4f4e\u4e86\u6a21\u578b\u8fc7\u62df\u5408\u7684\u98ce\u9669\u3002<\/p>\n
\u79bb\u6563\u7279\u5f81\u7ecf\u8fc7OHE\u540e\uff0c\u6bcf\u4e2a\u5206\u7c7b\u578b\u53d8\u91cf\u7684\u5404\u4e2a\u503c\u5728\u6a21\u578b\u4e2d\u90fd\u53ef\u4ee5\u770b\u4f5c\u72ec\u7acb\u53d8\u91cf\uff0c\u589e\u5f3a\u62df\u5408\u80fd\u529b\u3002\u4e00\u822c\u7684\uff0c\u5f53\u6a21\u578b\u52a0\u6b63\u5219\u5316\u7684\u60c5\u51b5\u4e0b\u7ea6\u675f\u6a21\u578b\u81ea\u7531\u5ea6\uff0c\u6211\u4eec\u8ba4\u4e3aOHE\u66f4\u597d\u3002<\/p>\n
\u5229\u7528 feature hash \u6280\u672f\u5c06\u9ad8\u7ef4\u7a00\u758f\u7279\u5f81\u6620\u5c04\u5230\u56fa\u5b9a\u7ef4\u5ea6\u7a7a\u95f4<\/p>\n
\u534a\u663e\u5f0f\u7279\u5f81\u7ec4\u5408<\/p>\n
\u533a\u522b\u4e8e\u663e\u5f0f\u7279\u5f81\u7ec4\u5408\u5177\u6709\u660e\u786e\u7684\u7ec4\u5408\u89e3\u91ca\u4fe1\u606f\uff0c\u534a\u663e\u5f0f\u7279\u5f81\u7ec4\u5408\u901a\u5e38\u7684\u505a\u6cd5\u662f\u57fa\u4e8e\u6811\u65b9\u6cd5\u5f62\u6210\u7279\u5f81\u5212\u5206\u5e76\u7ed9\u51fa\u76f8\u5e94\u7ec4\u5408\u8def\u5f84\u3002<\/p>\n
\u4e00\u822c\u505a\u6cd5\u662f\u5c06\u6837\u672c\u7684\u8fde\u7eed\u503c\u7279\u5f81\u8f93\u5165 ensemble tree\uff0c\u5206\u522b\u5728\u6bcf\u9897\u51b3\u7b56\u6811\u6cbf\u7740\u7279\u5b9a\u5206\u652f\u8def\u5f84\u6700\u7ec8\u843d\u5165\u67d0\u4e2a\u53f6\u5b50\u7ed3\u70b9\u5f97\u5230\u5176\u7f16\u53f7\uff0c\u672c\u8d28\u4e0a\u662f\u8fd9\u4e9b\u7279\u5f81\u5728\u7279\u5b9a\u53d6\u503c\u533a\u95f4\u5185\u7684\u7ec4\u5408\u3002<\/p>\n
ensemble tree \u53ef\u4ee5\u91c7\u7528 Gbdt \u6216\u8005 random forest \u5b9e\u73b0\u3002\u6bcf\u4e00\u8f6e\u8fed\u4ee3\uff0c\u4ea7\u751f\u4e00\u68f5\u65b0\u6811\uff0c\u6700\u7ec8\u901a\u8fc7 one-hotencoding \u8f6c\u5316\u4e3a binary vector\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n
![\"\"](\"https:\/\/www.growthhk.cn\/wp-content\/uploads\/mmexport1524568563955.jpg\")
<\/p>\n
\u4ee5\u4e0b\u51e0\u70b9\u662f\u6211\u4eec\u5728\u5b9e\u8df5\u4e2d\u7684\u4e00\u4e9b\u603b\u7ed3\u548c\u601d\u8003\u3002<\/p>\n
\u5728\u5b9e\u9a8c\u4e2d\u53d1\u73b0\u5982\u679c\u5c06\u8fde\u7eed\u503c\u7279\u5f81\u8fdb\u884c\u79bb\u6563\u5316\u540e\u5582\u5165 gbdt\uff0cgbdt \u7684\u6548\u679c\u4e0d\u4f73\uff0cAUC \u6bd4\u8f83\u4f4e\u3002<\/p>\n
\u8fd9\u662f\u56e0\u4e3a gbdt \u672c\u8eab\u80fd\u5f88\u597d\u7684\u5904\u7406\u975e\u7ebf\u6027\u7279\u5f81\uff0c\u4f7f\u7528\u79bb\u6563\u5316\u540e\u7684\u7279\u5f81\u53cd\u800c\u6ca1\u4ec0\u4e48\u6548\u679c\u3002<\/p>\n
xgboost \u7b49\u6811\u6a21\u578b\u65e0\u6cd5\u6709\u6548\u5904\u7406\u9ad8\u7ef4\u7a00\u758f\u7279\u5f81\u6bd4\u5982 user id \u7c7b\u7279\u5f81\uff0c\u53ef\u4ee5\u91c7\u7528\u7684\u66ff\u4ee3\u65b9\u5f0f\u662f\uff1a\u5c06\u8fd9\u7c7b id \u5229\u7528\u4e00\u79cd\u65b9\u5f0f\u8f6c\u6362\u4e3a\u4e00\u4e2a\u6216\u591a\u4e2a\u65b0\u7684\u8fde\u7eed\u578b\u7279\u5f81\uff0c\u7136\u540e\u7528\u4e8e\u6a21\u578b\u8bad\u7ec3\u3002<\/p>\n
\u9700\u8981\u6ce8\u610f\u7684\u662f\u5f53\u91c7\u7528\u53f6\u5b50\u7ed3\u70b9\u7684 index \u4f5c\u4e3a\u7279\u5f81\u8f93\u51fa\u9700\u8981\u8003\u8651\u6bcf\u68f5\u6811\u7684\u53f6\u5b50\u7ed3\u70b9\u5e76\u4e0d\u5b8c\u5168\u540c\u5904\u4e8e\u76f8\u540c\u6df1\u5ea6\u3002<\/p>\n
\u5b9e\u8df5\u4e2d\u91c7\u7528\u4e86 Monte Carlo Search \u5bf9 xgboost \u7684\u4f17\u591a\u53c2\u6570\u8fdb\u884c\u8d85\u53c2\u6570\u9009\u62e9\u3002<\/p>\n
\u5728\u79bb\u7ebf\u8bad\u7ec3\u9636\u6bb5\u91c7\u7528\u57fa\u4e8e Spark \u96c6\u7fa4\u7684 xgboost \u5206\u5e03\u5f0f\u8bad\u7ec3\uff0c\u800c\u5728\u7ebf\u9884\u6d4b\u65f6\u5219\u5bf9\u6a21\u578b\u6587\u4ef6\u76f4\u63a5\u8fdb\u884c\u89e3\u6790\uff0c\u80fd\u591f\u6ee1\u8db3\u7ebf\u4e0a\u5b9e\u65f6\u54cd\u5e94\u7684\u9700\u6c42\u3002<\/p>\n
\u6b64\u5916\uff0c\u5728\u5b9e\u8df5\u53d1\u73b0\u5355\u7eaf\u91c7\u7528 xgboost \u81ea\u52a8\u5b66\u5230\u7684\u9ad8\u9636\u7ec4\u5408\u7279\u5f81\u540e\u7eed\u8f93\u5165LR\u6a21\u578b\u5e76\u4e0d\u80fd\u5b8c\u5168\u66ff\u4ee3\u4eba\u5de5\u7279\u5f81\u5de5\u7a0b\u7684\u4f5c\u7528\uff1b<\/p>\n
\u53ef\u4ee5\u5c06\u539f\u59cb\u7279\u5f81\u4ee5\u53ca\u4e00\u4e9b\u4eba\u5de5\u7ec4\u5408\u7684\u9ad8\u9636\u4ea4\u53c9\u7279\u5f81\u540c xgboost \u5b66\u4e60\u5230\u7684\u7279\u5f81\u7ec4\u5408\u4e00\u8d77\u653e\u5165\u540e\u7eed\u7684\u6a21\u578b\uff0c\u83b7\u5f97\u66f4\u597d\u7684\u6548\u679c\u3002<\/p>\n
\u56db\u3001\u603b\u7ed3<\/p>\n
\u5b8c\u6574\u7684\u63a8\u8350\u7cfb\u7edf\u662f\u4e00\u4e2a\u5e9e\u5927\u7684\u7cfb\u7edf\uff0c\u6d89\u53ca\u591a\u4e2a\u65b9\u9762\uff0c\u9664\u4e86\u53ec\u56de\u3001\u6392\u5e8f\u3001\u5217\u8868\u751f\u4ea7\u7b49\u6b65\u9aa4\u5916\uff0c\u8fd8\u6709\u6570\u636e\u51c6\u5907\u4e0e\u5904\u7406\uff0c\u5de5\u7a0b\u67b6\u6784\u4e0e\u5b9e\u73b0\uff0c\u524d\u7aef\u5c55\u73b0\u7b49\u7b49\u3002<\/p>\n
\u5728\u5b9e\u9645\u4e2d\uff0c\u901a\u8fc7\u628a\u8fd9\u4e9b\u6a21\u5757\u96c6\u6210\u5728\u4e00\u8d77\uff0c\u6784\u6210\u4e86\u4e00\u4e2a\u96c6\u56e2\u901a\u7528\u63a8\u8350\u7cfb\u7edf\uff0c\u5bf9\u5916\u63d0\u4f9b\u63a8\u670d\u52a1\uff0c\u5e94\u7528\u572810\u591a\u4e2a\u680f\u4f4d\uff0c60\u591a\u4e2a\u573a\u666f\uff0c\u53d6\u5f97\u4e86\u5f88\u597d\u7684\u6548\u679c\u3002<\/p>\n
\u672c\u6587\u4fa7\u91cd\u4ecb\u7ecd\u4e86\u53ec\u56de\u4e0e\u6392\u5e8f\u7b97\u6cd5\u76f8\u5173\u7684\u76ee\u524d\u5df2\u6709\u7684\u4e00\u4e9b\u5de5\u4f5c\u4e0e\u5b9e\u8df5\uff0c\u4e0b\u4e00\u6b65\uff0c\u8ba1\u5212\u5f15\u5165\u66f4\u591a\u5730\u6df1\u5ea6\u6a21\u578b\u6765\u5904\u7406\u53ec\u56de\u4e0e\u6392\u5e8f\u95ee\u9898\uff0c\u5e76\u7ed3\u5408\u5728\u7ebf\u5b66\u4e60\u3001\u5f3a\u5316\u5b66\u4e60\u3001\u8fc1\u79fb\u5b66\u4e60\u7b49\u65b9\u9762\u7684\u8fdb\u5c55\uff0c\u4f18\u5316\u63a8\u8350\u7684\u6574\u4f53\u8d28\u91cf\u3002<\/p>\n
References<\/p>\n
[1] Koren, Yehuda,Robert Bell, and Chris Volinsky. “Matrix fact<\/a><\/span>orization techniques forrecommender systems.” Computer 42.8 (2009).
[2] Sedhain, Suvash,et al. “Autorec: Autoencoders meet collaborative filtering.” Proceedingsof the 24th International Conference on World Wide Web. ACM, 2015.
[3] Rendle, Steffen,Christoph Freudenthaler, and Lars Schmidt-Thieme. “Factorizingpersonalized markov chains for next-basket recommendation.” Proceedings ofthe 19th international conference on World wide web. ACM, 2010.
[4] Dong, Xin, etal. “A Hybrid Collaborative Filtering Model with Deep Structure forRecommender Systems.” AAAI. 2017.
[5] Hidasi, Bal\u00e1zs,et al. “Session-based recommendations with recurrent neuralnetworks.” arXiv preprint arXiv:1511.06939 (2015).
[6] Zhu, Yu, et al.”What to Do Next: Modeling User Behaviors by Time-LSTM.” Proceedingsof the Twenty-Sixth International Joint Conference on Artificial Intelligence,IJCAI-17. 2017.
[7] Barkan, Oren,and Noam Koenigstein. “Item2vec: neural item embedding for collaborativefiltering.” Machine Learning for Signal Processing (MLSP), 2016 IEEE 26thInternational Workshop on. IEEE, 2016.
[8] Wang, Hao,Naiyan Wang, and Dit-Yan Yeung. “Collaborative deep learning forrecommender systems.” Proceedings of the 21th ACM SIGKDD InternationalConference on Knowledge Discovery and Data Mining. ACM, 2015.
[9] Covington, Paul,Jay Adams, and Emre Sargin. “Deep neural networks for youtuberecommendations.” Proceedings of the 10th ACM Conference on RecommenderSystems. ACM, 2016.
[10] Kim, Donghyun, et al. “Convolutional matrix factorization fordocument context-aware recommendation.” Proceedings of the 10th ACMConference on Recommender Systems. ACM, 2016.[11] https:\/\/mli.github.io\/2013\/03\/24\/the-end-of-feature-engineering-and-linear-model\/[12] Richardson, Matthew, Ewa Dominowska, and Robert Ragno.”Predicting clicks: estimating the click-through rate for new ads.”Proceedings of the 16th international conference on World Wide Web. ACM, 2007[13] Andrew, Galen, and Jianfeng Gao. “Scalable training of L1-regularized log-linear models.” Proceedings of the 24th internationalconference on Machine learning. ACM, 2007.[14] Graepel, Thore, et al. “Web-scale bayesian click-through rateprediction for sponsored search advertising in microsoft’s bing searchengine.” Omnipress, 2010.
[15] McMahan, H. Brendan, et al. “Ad click prediction: a view fromthe trenches.” Proceedings of the 19th ACM SIGKDD international conferenceon Knowledge discovery and data mining. ACM, 2013.
[16] Chen, Tianqi, and Carlos Guestrin. “Xgboost: A scalable treeboosting system.” Proceedings of the 22nd acm sigkdd internationalconference on knowledge discovery and data mining. ACM, 2016.
[17] Rendle, Steffen. “Factorization machines.” Data Mining(ICDM), 2010 IEEE 10th International Conference on. IEEE, 2010.
[18] Juan, Yuchin, et al. “Field-aware factorization machines forCTR prediction.” Proceedings of the 10th ACM Conference on RecommenderSystems. ACM, 2016.
[19] Gai, Kun, et al. “Learning Piece-wise Linear Models fromLarge Scale Data for Ad Click Prediction.” arXiv preprint arXiv:1704.05194(2017).
[20] He, Xinran, et al. “Practical lessons from predicting clickson ads at facebook.” Proceedings of the Eighth International Workshop onData Mining for Online Advertising. ACM, 2014.
[21] Cheng, Heng-Tze, et al. “Wide & deep learning forrecommender systems.” Proceedings of the 1st Workshop on Deep Learning forRecommender Systems. ACM, 2016.
[22] Guo, Huifeng, et al. “DeepFM: A Factorization-Machine basedNeural Network for CTR Prediction.” arXiv preprint arXiv:1703.04247(2017).
[23] Zhou, Guorui, et al. “Deep Interest Network for Click-ThroughRate Prediction.” arXiv preprint arXiv:1706.06978 (2017).
[24] Blondel, Mathieu, et al. “Higher-orderfactorization machines.” Advances in Neural Information ProcessingSystems. 2016.
[25] http:\/\/breezedeus.github.io\/2014\/11\/20\/breezedeus-feature-hashing.html
[26] https:\/\/en.wikipedia.org\/wiki\/Categorical_variable
[27] https:\/\/www.zhihu.com\/question\/48674426
[28] \u591a\u9ad8\u7684AUC\u624d\u7b97\u9ad8\uff1fhttps:\/\/zhuanlan.zhihu.com\/p\/24217322<\/p>\n
\u6587\uff1a\u643a\u7a0b\u6280\u672f\u4e2d\u5fc3<\/a><\/span><\/p>\n
<\/span>\u76f8\u5173\u6587\u7ae0\u63a8\u8350\uff1a<\/strong><\/p>\n
\u300a\u300a\u589e\u957f\u9ed1\u5ba2\u300b| \u5bfb\u627e\u751f\u547d\u4e2d\u7684AHA\u65f6\u523b<\/a>\u300b
\u300a\u9999\u6c34\u4ea7\u54c1\u4e3a\u4f8b\uff0c\u9996\u5e2d\u589e\u957f\u5b98\u5e38\u7528\u6570\u636e\u91c7\u96c6\u3001\u5206\u6790\u3001\u6316\u6398\u7684\u8fc7\u7a0b\u53ca\u7b97\u6cd5\u6a21\u578b<\/a>\u300b
\u300a\u9996\u5e2d\u589e\u957f\u5b98\u7684\u5317\u6781\u661f\u6307\u6807\uff1a\u65b0\u7528\u6237\u6fc0\u6d3b<\/a>\u300b<\/p><\/blockquote>\n\u66f4\u591a\u7cbe\u5f69\uff0c\u5173\u6ce8\uff1a\u589e\u957f\u9ed1\u5ba2<\/a><\/span>\uff08GrowthHK.cn\uff09<\/p>\n
![\"\"](\"https:\/\/www.growthhk.cn\/wp-content\/uploads\/mmexport1524568300468.png\")
<\/p>\n![\"\"](\"https:\/\/www.growthhk.cn\/wp-content\/uploads\/mmexport1524568358371.jpg\")
<\/p>\n