【大数据公益大学201397聚会】王兴星排序中的机器学习.pptVIP

广告排序中的机器学习 王兴星 cquwxx@sina.cn 个人简介 姓名：王兴星 公司：搜狗 （副研究员） 部门：探索引擎 方向：排序、推荐 Agenda Prior Problem Solution Summary Prior Applications of CTR prediction : - Search - Recommendation - Sponsored Ads - Display Ads - Real Time Bidding Problem Bias position、color etc. Sparsity Data is sparse, tail is long； Model OWLQN Ad predictor LR FRTL-Proximal Position Bias Clicks and time percentage Under different positions: Bias Eye Tracking： Position Bias Different level of results under same positions： Position Model Examination Hypothesis 浏览 相关 点击 Position Model Cascade Model r1 r2 rN C C C Position Model Dependent Click Model(DCM) Position Model User Browsing Model(UBM) Bayesian Browsing Model(BBM) Other Bias Click Over Expected Click(COEC): Build two models: Bias Model : with bias factors; Pure Model : without bias factors. Sparsity For ad : IF , are good predictor for future performance. But: Data is sparse, tail is long; New querys, New ads. Solution: Aggregate Data Sparsity Aggregate By Account Sparsity Aggregate By Collective Intelligence Collaborative Filtering(CF) Random Walk Sparsity Aggregate By Reduction Sparsity Aggregate By Reduction Document Segment Clustering Topic Model Sparsity Other By Time: Discretization: One bin for missing value. Model OWLQN Objective function Sparsity: L1 better than L2 Derivative of L1 not smooth Solution Define pseudo-gradient Restricted to an orthant Model OWLQN Model OWLQN Model Adpredictor Model Adpredictor Factor Graph Nodes ： = Factors = Variables Edges: Dependencies of factors on variables. Model Adpredictor Model