I am currently researching on Ensemble Multiclass Classification and Outlier detection in data mining. I intend to use Tukey's method to detect outliers. To facilitate the ranking, I intend to compute LOF for each of the instances and then sort them using the feature outliers as the primary key and LOF as the second key. I will create classifiers using Random-forest. Using the predictive values of root mean square error (RMSE), i will generate a learning curve by removing the outliers and recording their RMSE and ROC values starting with the top most outlier. The process will continue until all the outliers detected by the Tukey method have been removed from the dataset.
To establish which outliers to retain in the dataset after the experiments the threshold shall be all the outliers below the least value for RMSE value depicted in the experiment.
Hi Dalton,
Your method seems to be fine but you have to be careful and clear about the features and the target class labels.
However, a pretty straightforward method is to apply IQR that directly produces information about whether the instance is outlier or not, so you'll deal with a binary classification. After that, classify your data with Random Forest algorithm.
HTH.
Samer
Hi Dalton ,
Are you familiar with Isolation Forest? It should be interesting to compare it against your method.
https://cs.nju.edu.cn/zhouzh/zhouzh.files/publication/icdm08b.pdf
It also has an easy to use implementation on scikit-learn:
http://scikit-learn.org/stable/modules/generated/sklearn.ensemble.IsolationForest.html
Best Regards,
Heitor
This thought is so interestin. Try to make a loss function based on adaboost methodolgy that can help in improving your outlier methods with lower processing time.
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining