I finished a course on machine learning last semester and came across kaggle some time around then. I immediately pushed it in my winter-break-to-do list and lucky kaggle got selected as the only thing that actually happened from that list.
Anyhow, I have been working on the easiest problem they have on kaggle and my performance has been abysmal, at best.
I wanted to get a result out there right away so implemented a model using what I knew already - MATLAB. Having done an assignment for the aforementioned course, I knew how to implement SVM in MATLAB and submitted an answer based on ten minutes of work. I was placed in the bottom twenty of 150+ teams (accuracy of ~80%). Like I said, abysmal.
A little reading on the forums and quora suggested that MATLAB might not be a logical choice for solving problems on Kaggle and hence I shifted to R thereafter. The implementation of SVM in R (in the e0171 package) improved my accuracy by roughly 3%.
The dataset has some 40 attributes and so, it looks like some dimensionality reduction technique should improve the results. However, I want to try different basic implementations first. I am trying the k nearest neighbours approach next.
I am appending the R commands for implementing SVM on the given data set, for future reference
setwd("C:/Users/Siddhant/Kaggle/Data Science london + Scikit-learn") #reading the datasets. The first row in the csv files should be the titles for the columns test <- read.csv("test 1.csv") train <- read.csv("train1.csv") trainLabels <- read.csv("trainLabels1.csv") #building the model. Since the trainLabels is not of type factor, we need to explicitly mention the type to be C-Classification to avoid regression model <- model <- svm(train,trainLabels, type = 'C-classification') pred <- predict(model,test) write.csv(pred,"answer.csv")