Using SVM, I got the result, the model predict in the test data was not fit, Could you suggest to me some parameters I should try to select the best model, Thank you so much! Have a nice day. I use the e1071 packages tune function to adjust the parameters(cost and gamma)
The first pic is a selection of some variables, the second pic I loaded all variables
I assume you are trying to perform a binary classification.
What kind of data do you use? How many features? How big is your training and test data set?
The predicted results using SVM are good.i suggest u also use codes to compute model sensitivity ,precision and accuracy...using the confusion matrix table.
Also u may compare the performance of SVM in terms of precision n accuracy,with logistic regression and naive Bayes,or even decision tree algorithms ...
Depending on ur data structures,one algorithm may prove more reliable than the other ...reliability varies with data.
Cheers.
Meshack Owira Amimo: Why do you think the results are good? For me it looks quite random in the test set (precision/recall roughly 0.5), the training set is overfitted.
As Random Forest has even more a tendency to overfit, I don't think this will be useful. Maybe you could reduce the dimensionality first with PCA, UMAP?
Christian Winkler Thank you so much, but i do a confusionmatrix it shows the ACC in trainset around 0.8, in the test as you said around 0.5. Could you recommend some packages to handle my data (PCA, UMAP). Thanks for your kind advice. My data is not big just 401, I slipped the trainset and testset by 3/4 and 1/4 respectively.
Pei Li: Good split between test and training set.
How many dimensions does your data have? If this number is too high, you will get overfitting with so few samples.
I am not too familiar with R, but `prcomp` seems to be a package for PCA, `uwot` provides a UMAP algorithm.
Pei Li: These are too many dimensions. Try to reduce it with PCA first, only taking the largest eigenvalues. If you can bring it down to two or three dimensions, this will probably help a lot.
UMAP is much more powerful, you could also give it a try.
Maybe it is worth to normalize the features first.
The caret package in R will perform classification for u in R...and will show u the variable importance...
That way you will know which variables play key roles in ur predictions.
Please down the 'caret' package and read it's usage...just type 'caret in CRAN' in Google.
Play with the examples highlighted first before customising the codes to ur specific problem .
Good Luck.
Dr.Amimo
Sorry ..I meant to say you download the caret package from the CRAN.. the R website
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training