I have a dataset with 120 features and 5000 instances. The dataset is combination of categorical and numerical values. It is a tabular dataset. My problem is a binary classification problem. I trained my dataset with all classic classification algorithms like Naive Bayes, Bayesian net, SVM, MLP, Random forest, Logistic regression etc. I would like to know is there any algorithms available in the machine learning field which is new and not the classic one and can be implemented using a tabular dataset.

I heard about Convolution neural network, deep neural network etc but I believe they are used in image data not tabular data.


1 Answer 1


You can try Generalized Additive Models (GAM). It models the response variable, $y$, as a sum of functions of individual features $f_i(x_i)$: $y = \sum\limits_if_i(x_i)$. You don't need to provide $f_i$, the algorithm finds learns them from the data. By analyzing the $f_i$, you can see, which feature contributes significantly, and in which way. You can even fit the $f_i$ to simple analytic functions and get an analytic dependence of $y$ on $x_i$.

Links for GAM: GAM in wikipedia, GAM in R, GAM in Python, ClassificationGAM in Matlab, as well as here, here, here, and here.

A related method is Alternating Conditional Expectations (ACE) from this paper: link. I wrote a blog post about it here.

These methods are not new, but I have a feeling that you didn't try them.

I also recommend browsing through methods in weka. It has some interesting, non-mainstream algorithms, such as classifier for learning functional trees, HotSpot, alternating model trees, alternating decision trees, and many other.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.