Walmart Trip-Type Classification

---

Walmart, one of the leading retailers in US market continually strives on their core mission of better understanding and serving the customers. One way Walmart is able to improve customer’s shopping experience is by segmenting their store visits into different trip types, which enables Walmart to create the best shopping experience for every customer. The problem here is to improve the science behind trip type classification, which will help Walmart to refine their segmentation process.

We converted the walmart data to LIBSVM format, built training models on classification algorithms such as Naive Bayes, K means, Random Forest and SVM, compared them to find accuracy, precision and recall. Finally, we concluded that SVM classifier outperform others.

This project was a part of Kaggle competition. With logarithmic loss of 1.27 for the classsification, I secured overall 432nd rank in the competition

• Client: Walmart
• Date: November 2015
• Type: Classification Problem

Close

Large Scale Text Classification

---

Hierarchies are becoming ever more popular for the organization of text documents, particularly on the Web. Wikipedia is an example. Along with their widespread use, comes the need for automated classification of new documents to the categories in the hierarchy. As the size of the hierarchy grows and the number of documents to be classified increases, a number of interesting machine learning problems arise. In particular, it is one of the rare situations where data sparsity remains an issue despite the vastness of available data.

The large scale of data called for the use of advanced classification algorithms instead of the traditional ones that could not scale.Parallelized k-NN algorithm with the help of Hadoop Map-Reduce paradigm was implemented to handle space and time complexities of large data. Cosine- vector similarity was used to compare distances between two points instead of the euclidean distance with absolute distance in the traditional algorithm. Tie- breaking with the nearest neighbour instead of random tie breaking was used. Multinomial Naïve Bayes algorithm was used to handle the weighted features in the web document. Further, Transformed Complement Weight Naïve Bayes Classification algorithm was used to handle skewed data. Additionally, we also implemented a hierarchical algorithm using scikit-learn python API.

• Client: Wikipedia
• Date: May 2014

Close

Movie Recommender System

---

The idea of this project is to recommend movies which users are likely to find interesting based on the ratings of similar users. To start with, we collected a list of users and their ratings for the movies they have already watched. To predict the rating for an unrated movie for a given user, we used User Based Collaborative filtering approach. The rating for the movie was predicted using the similarity of its k nearest neighbhours determined through their Pearson Coefficent.

• Date: October 2015

Close