For example, a given dataset is having 500 samples, 12 attributes and 2 classes. On applying various classification techniques like LDA, QDA, Decision trees, ANN, SVM etc. the obtained accuracy rates are not significant because of any unobserved reason like high variance in attribute values etc. Kindly suggest and share the methods for improving the results in such cases.
Hi Aman,
You can try different ways, for example Normalization, (i.e., normalize values between 0 and 1, or -1 to 1.
Another method you can use is Principle Component Analysis PCA for dimensional reduction to get rid of unwanted attributes for better results.
Finally, you can use one of feature selection methods also to reduce the number of attributes for better learning.
Try one of the above methods and I am quite sure you will see a difference in the results.
Kind regards
Ahmed
Dear Maureen,
One by one we are trying every approach for classification, the point here is that if let suppose the data set is containing so dynamic values of attributes that classification techniques face difficulty in separating it into different spaces. or may any other reason. Is their any way to normalize the data or under-sample or over-sample the data so that good accuracy can be achieved.
Dear Ahmed,
Thanks for answering.
Kindly explain how to perform normalization in detail.
Dear Aman,
I could say the following:
1. ANN is the best method of classification.
2. If the ANN gives bad results, then you have bad database.
3. The bad database is not possible to get a good classification model by any method.
What a bad database?
The database is bad, if for example:
1. The database contains a lot of mistakes (ie contain anomalous values and contradictions).
2. The database does not contain the factors that strongly influence the objective function.
3. The database contains a lot of factors, and few examples (table rows).
Send me your table in MS Excel format, and I'll check, Your base is bad or not.
My mail: [email protected]
Respected V. S. Abrukov,
That it totally agree that the database can be bad. now what are the solution for correcting the database.
Dear Aman,
The ways are:
1. to exclude anomalous values and contradictions.
2. to include the factors that strongly influence the objective function.
3. to add examples (table rows).
4. etc
What is the accuracy rate ? The classes may not be well separated it may be imbalanced there might be some noisy features. Feature selection is a well documented and researched area , which can be applied to many problems.
Since your data-set is small, you can explore and experiment using Weka. Weka is a free software with a friendly GUI and you can perform exploratory data analysis. It allows user to statistically and visually analyse each attribute, correlations. It also offers a wide variety of classifiers. Using decision tree or rule based can yield valuable insights. Several ensemble methods are also available. You can choose your pick.
All the best.
- Badges
- Science topic