Presentation is loading. Please wait.

Presentation is loading. Please wait.

The Combination of Supervised and Unsupervised Approach

Similar presentations


Presentation on theme: "The Combination of Supervised and Unsupervised Approach"— Presentation transcript:

1 The Combination of Supervised and Unsupervised Approach
Yingju Xia, Shuangyong Song, Qingliang Miao, Zhongguang Zheng Fujitsu Research and Development Center Copyright 2015 FUJITSU R&D CENTER CO., LTD.

2 Overview Training data Test data Feature Extraction
Unsupervised method Features Features Predictions Supervised Ensemble Learning Combination Model Final predictions Copyright 2015 FUJITSU R&D CENTER CO., LTD.

3 Main Features Time features Product and Category feature
Start time, End time, day, hour, weekday, … Product and Category feature Single level : such as ‘A00001’, ‘B00001’ , ‘C00001’, ‘D00001’ Combinations: such as ‘A00001/B00001’, ‘A00001/B00001/C00012’, ‘/B00001/C00012’ Transferring features The transferring from one record to other record in the same session for example: ‘A00002/B00003/C00014/D11017/’ ;‘A00010/B00055/C00135/D11018/’ has the feature: ‘A00002-A00010’, ‘B00003-B00055’, … Product ID Prefix For example: ‘D09233’ has the prefix feature ‘D0923’, ‘D092’, ‘‘D09’’ ‘D09232’ also has the prefix feature ‘D0923’, ‘D092’, ‘‘D09’’ Copyright 2015 FUJITSU R&D CENTER CO., LTD.

4 Supervised Ensemble Learning
Dynamic Classifier Selection using competence smoothness We adopt the DCS framework[1] and use competence to measure the classifier behavior The competence is defined according to the BAC evaluation Classifiers competence is learning on Training data Graph method is used for smoothness Experimental results Dataset: the training set of PAKDD’15 (15000 samples) Training set: 11000, Valid set: 2000, Test set 2000 Models: random forest, naïve bayes, decision tree, knn, boost, neural networks About 3% enhancement by using the model fusion [1] Giacinto G, Roli F, Dynamic classifier selection based on multiple classifier behavior, Pattern Recognition 34 (2001) 1879–1881. [2] Woloszynski T, Kurzynski M, Podsiadlo P, et al. A measure of competence based on random classification for dynamic ensemble selection[J]. Information Fusion, 2012, 13(3): Copyright 2015 FUJITSU R&D CENTER CO., LTD.

5 Combination of Supervised and Unsupervised method
The classification is usually under the assumption of independent and identical distribution of objects The internal structure information among the objects are good complement to classification. We follow the idea of maximizing the consensus among both supervised predictions and unsupervised constraints[1]. We tried several unsupervised approaches to put the data into groups The most efficient grouping method for this data set is using the time interval of adjacent sessions About 2% enhancement by using the combination of supervised and unsupervised method [1] Gao J, Liang F, Fan W, Sun Y, and Han J. Graph-based consensus maximization among multiple supervised and unsupervised models. Advances in Neural Information Processing Systems (NIPS), 22:585–593, 2009. Copyright 2015 FUJITSU R&D CENTER CO., LTD.

6 Remarks The task of PAKDD 2015 is wonderful platform for evaluating machine learning methods We adopt the Dynamic Classifier Selection method for model fusion Task oriented classifiers competence and graph based smooth method is explored in model fusion The combination of supervised and unsupervised approach is explored in this contest. Future work: More general fusion method should be explored Optimal target need to explored for finding the tradeoff between supervised and unsupervised approach Copyright 2015 FUJITSU R&D CENTER CO., LTD.

7 Copyright 2010 FUJITSU LIMITED


Download ppt "The Combination of Supervised and Unsupervised Approach"

Similar presentations


Ads by Google