My solution in Python for the kaggle problem StumbleUpon Evergreen Classification Challenge
This is the second project which I worked on NLP. Like the previous one, this has the same story - that I started Kaggle in december end, and that time no competition was running which interested me. Hence although this competition had ended but I was deeply interested in doing this and thus did. Here's the project link: http://www.kaggle.com/c/stumbleupon Also,Since this was one of my first attempts I missed some key techniques back then. But nevertheless it provided decent results, if you put the submission file.
The problem dealt with webpage classification, the performace was dominated by the features of webpage text, as it gave max relevant features. We had to use NLP methods - owing to the brilliance of sklearn package, we could use the TFID techniques to vectorize our inputs. Since a problem of NLP at its best, The TFID approach was tuned with different parameters like different n-grams,norm, df, token patterns, customized tokenizers. We also used Latent Semantic Analysis (Truncated SVM in this case), as we got amazing results when we reduced the components correctly (BUT, This brings a crazy Variance in our model!) Quite a lot of classifiers were tried but we sticked with Logsistic Regression and moreoever used an ensemble of logistic regression models (the data was divided using k-fold cross validation. GridSearch was used to search for the best matching parameters as well. Surprisingly, SVMs didnt work in this one(all kernels were tried in hope) to that extent and moreover it took plenty of time to train.
stumble.py - The core file containing the implementation.
stumble-51.csv - This was one of the setting when I got the best rank of 61. This is the submission file prepared from code.
kaggle-51.jpg - The screenshot showing the same (total 625 teams), But crazy variance has slightly different results each time (LSA!)
numpy (For arrays mostly), pandas , sklearn (scikit-learn), nltk (for stemmers)