Abstract: User portraits are a kind of tagged user model constructed from user’s social attributes,lifestyle and consu-mer behavior,etc.The core work of building user portraits is to “tag” the user.Based on the user’s query word history,this paper proposed a two-level stacking algorithm framework for predicting user’s multi-dimensional labels.For the first-level models,a variety of models are built on each tag prediction subtask.The SVM model and Trigram feature are used to extract the differences of user’s words habit.The doc2vec shallow neural network model is used to extract the semantic relation information of the query words,and the convolution neural network model is used to extract the deep semantic association information between the query words.Experiments show that doc2vec has relatively good predictive accuracy in dealing with short texts related tasks (such as user queries).For the second-level models,the XGBTree model and the Stacking method are used to extract the association information between the label’s attributes of the user,so that the average prediction accuracy is further improved by 2%.In the big data competition “Sougou User Portrait Mining For Precision Marketing” organizated by China Computer Federation in 2016,this two-level stacking algorithm framework won the championship from 894 teams.

Key words: User portraits,Tag prediction,Short text classification,Multi-model ensemble