Machine Learning Projects

Recommender System based on Collaborative Filtering

The project focuses on recommending a user, Jobs related to what he has applied in distant past and jobs applied by similar users based on recent past. Each job is then scored based on the parameters like, users location, job location, users education and job requirement in addition to how many similar users applied to that particular job.

The project is made in Python uses an OO approach and spans up to 6 classes. Since the data files were pretty huge, garbage collection was used extensively to flush memory leaks and unused portions of the memory.

Several Machine learning and Collaborative filtering techniques are used in the project.

Supervised Learning

Data was provided to predict the Gender of the Incoming Test data, the Data had three attributes (Height, Age, Weight) based on the training data provided the algorithms were trained and performance of all the above mentioned algorithms.

Gaussian Naive Bayes

GNB actually looks at the entire N-Dimensional Distribution and the continuous attributes are evaluated by a Gaussian Created around their data with an assumption of independence between the attributes.It was developed in Python 2.7

KNN

KNN just sees the least Cartesian Distance and makes its judgment depending on the K value, the Tuples that are selected. It was developed in Python 2.7

Logistic Regression

Logistic Regression learns by applying gradient descent on the data. The Input was given randomly to the algorithm to fine tune itself, making sure that we don't over-fit the training data.

It was built in Matlab.

Supervised Learning - Decision Tree

An algorithm was developed to make a decision tree, based on the Entropy of the attributes from the training data, attributes were selected with the highest getting selected first and then based on it's each branch subsequent attributes were selected.

The Tree was then tested on the Test data, and it's performance was evaluated.

The Code was written in Python, by having two classes A Tree class and a Node Class.

Neural Networks for Image Classification

A set of images (*.pgm of 16x15) of Balls and Trees were provided to train as well are to test the network. The networks were built using Forwards feed and Back Propagation algorithm, having 1 Input Layer, 1 Hidden Layer and 1 Output Layer.

Hidden Layer was to be constructed with 4, 8 and 12 hidden nodes choosing one at a time and performance of all the three networks was assessed, based on the learning time and classification accuracy.

Network used gradient descent to learn the weights of the networks, training data was randomly provided to the network repeatedly and after sufficient training test data was provided for classification.

User - Job Predictions

Done as part of Data Mining Class Project, where we were supposed to predict the jobs users will apply based on the previous application of the users or similar users and job history of the users, the Data Provided was quite hefty a simple Cartesian product would have taken 2 billion cycles, hence Data Mining Techniques such as Vector Space Model was used to find cosine similarity and Naive Bayes was used to calculate metrics on more than 10 dimensions and predict the probability of the jobs, and order the Top 150 predictions in decreasing order of the likelihood of a user-job pair.