anyone suggest me to improve classification accuracy when using xgboost method applied to the problem of forecasting the link in the network's co-authors. Here result of xgboost is not better than the J48 and SVM, because this problem is the label positive often less than negative grade.
According to my experience. classification accuracy depends aupon several factors and can be improved at different stages:
1) what are features used? features should complement each other and be not "noisy" (non- or poorly informative);
2) what is a chosen classifier (in this sense, each author dealing with a particular classifier either likes it or has experince of working with it ot has some reasons for using this classifier). It's difficult to say what is the best classifiier for Your particular case;
3) how learning is done? You have already advices from people above;
4) is some postprocessing applied to improve classification (correct or remove misclassification)? This can help in some cases.
5) is some pre-processing done? Someting signal/image denoising before feature extraction can be helpful.
If you have an adversely unbalanced dataset, it could mean trouble for any gradient computing classifiers -- this includes XGBoost. If you are doing binary classification, see of you could approach your challenge as an anomaly detection problem.
If you able to get the balancing sorted out, you could try referring the following link to see if you missed anything in the process of applying XGBoost in R.
https://www.analyticsvidhya.com/blog/2016/01/xgboost-algorithm-easy-steps/
Please check the data whether it is an imbalance or not. If it is the imbalance, you can use oversampling, undersampling, or SMOTE techniques to balance instances for output. Therefore, accuracy will increase...
The classification accuracy depends on the features of the pattern that you focus on (i mean the dataset), this is in case that you want to apply the classification on a particular dataset. but if you want to build a special dataset, you should ensure a good way to generate features with more differences.
I recommend this chapter:
http://www.jars1974.net/pdf/12_Chapter11.pdf
According to my experience. classification accuracy depends aupon several factors and can be improved at different stages:
1) what are features used? features should complement each other and be not "noisy" (non- or poorly informative);
2) what is a chosen classifier (in this sense, each author dealing with a particular classifier either likes it or has experince of working with it ot has some reasons for using this classifier). It's difficult to say what is the best classifiier for Your particular case;
3) how learning is done? You have already advices from people above;
4) is some postprocessing applied to improve classification (correct or remove misclassification)? This can help in some cases.
5) is some pre-processing done? Someting signal/image denoising before feature extraction can be helpful.
If you problem is unbalanced data, you could try to use some of the already mentioned techniques or use cost-sensitive learning.
Use principal components analysis to reduce dimension of data set. In this way,only significant patterns were used by classification.
hiiii
i think use features, it is best if you got it best answers
https://www.analyticsvidhya.com/blog/2016/01/xgboost-algorithm-easy-steps/
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification