Anuj Katiyal

(Skills : Python, Keras, Scikit-Learn, Pandas, Numpy, Matplotlib, Seaborn)

Datasets used: MNIST, http://ufldl.stanford.edu/housenumbers/

1. Trained a multilayer perceptron (feed forward neural network) with two hidden layers and rectified linear nonlinearities on the MNIST dataset using the Keras Sequential interface. Compared the baseline model with a model using drop-out resulting in an improved accuracy.

2. Trained a convolutional neural network on the SVHN dataset… Show more

(Skills : Python, Keras, Scikit-Learn, Pandas, Numpy, Matplotlib, Seaborn)

Datasets used: MNIST, http://ufldl.stanford.edu/housenumbers/

1. Trained a multilayer perceptron (feed forward neural network) with two hidden layers and rectified linear nonlinearities on the MNIST dataset using the Keras Sequential interface. Compared the baseline model with a model using drop-out resulting in an improved accuracy.

2. Trained a convolutional neural network on the SVHN dataset. Achieved an accuracy of 92.7% on the test-set with a base model. Also build a model using batch normalization and dropout which led to an increased accuracy of 95.3%.

3. Imported the weights of a pre-trained convolutional neural network, VGG, and used it as feature extraction method to train a multi-layered perceptron on the pets dataset. We achieved an accuracy of 73.2% on the 37-class classification task. Show less

(Skills Used: Python, Numpy, Scikit-Learn, Pandas, Matplotlib, Seaborn, Git, Travis)
Dataset Used : https://archive.ics.uci.edu/ml/datasets/bank+marketing
Task was to predict whether someone will subscribe to the term deposit or not based on the direct marketing campaign ran by a Banking Institution based on phone calls.

Stood 2nd amongst 100 teams participating in the In-Class Kaggle Classification Problem. The best ROC-AUC score of 0.798 on the test set was obtained using a… Show more

(Skills Used: Python, Numpy, Scikit-Learn, Pandas, Matplotlib, Seaborn, Git, Travis)
Dataset Used : https://archive.ics.uci.edu/ml/datasets/bank+marketing
Task was to predict whether someone will subscribe to the term deposit or not based on the direct marketing campaign ran by a Banking Institution based on phone calls.

Stood 2nd amongst 100 teams participating in the In-Class Kaggle Classification Problem. The best ROC-AUC score of 0.798 on the test set was obtained using a poor-man stacking ensemble model with Logistic Regression, SVM and Random Forest. The major modules in the project included:

1. Data Cleaning and Pre-processing to exclude outliers, remove redundant independent variables and impute missing values.
2. Created a pipeline to perform feature engineering, feature selection and model validation on the dataset.
3. Applied classification models such as Logistic Regression, SVM followed by tree-based models such as Random Forest, Gradient boosted trees, XgBoost etc. Tuned the hyper-parameters for selecting the best model for each algorithm.
4. Applied various Ensemble Methods including Voting Classifiers, Poor Man's Stacking, Weighted Ensembles. The best ROC-AUC score of 0.798 on the test set was obtained using a poor-man stacking ensemble model with Logistic Regression, SVM and Random Forest.
5. Applied Resampling techniques like RandomUnderSampler, RandomOverSampler with various Classification Algorithms followed by Ensembles using them. Resampling was also applied using techniques like Edited Nearest Neighbors, SMOTE, SMOTEENN and SMOTETOMEK.

Show less

(Skills Used: Python, Scikit-learn, Matplotlib, Travis, Git, Numpy)

Datasets used are available at: https://www.census.gov/housing/nychvs/data/2014/userinfo2.html
(Data can be accessed at https://www.census.gov/housing/nychvs/data/2014/uf_14_occ_web_b.txt)

We utilized the linear models (Regression approaches) for predicting the monthly rent prices of an apartment in NYC based on the collected 2014 Census data. Feature engineering was done to obtain features that directly affect… Show more

(Skills Used: Python, Scikit-learn, Matplotlib, Travis, Git, Numpy)

Datasets used are available at: https://www.census.gov/housing/nychvs/data/2014/userinfo2.html
(Data can be accessed at https://www.census.gov/housing/nychvs/data/2014/uf_14_occ_web_b.txt)

We utilized the linear models (Regression approaches) for predicting the monthly rent prices of an apartment in NYC based on the collected 2014 Census data. Feature engineering was done to obtain features that directly affect the pricing of an apartment like the number of rooms, presence of elevator, floor number etc. The model was tested using a separate test data that was extracted from the complete dataset, based on metrics like the R^2 score (obtained as 0.62)

The project involved the followed major steps:
1. Loading and Initial Analysis of Data. This steps involved cleaning data to drop the rows with missing rent, dropping columns with leaked information and additional pre-processing like One Hot Encoding and Feature Selection/Engineering.
2. We were able to obtain an R^2 value of around 0.6, which is at par with most of the analysis work done on this dataset using linear models like Linear Regression, Ridge Regression, Lasso Regression and Polynomial Regression.
3. Important observations from the dataset included dropping rows with topcoded rent and also dropping rows which are rent controlled, which lead to an increase in the R^2 score from 0.45 to 0.62 Show less

The project was done as a part of the course in Computer Vision. The aim of the project was to obtain similar images to the trained Oxford Building Dataset using SIFT and SURF features based on the tf-idf score.
Tools Used: C, SQL

(Skills: Python, Numpy, Scikit-learn, Matplotlib, Seaborn, Pandas, Nltk, Spacy, Gensim)

Major modules:

1. Data Cleaning - It involved the initial loading of data along with visualization of the class distribution. The duplicated labels were consolidated into one and the duplicated data points were removed.

2. Baseline Multi-Classification Algorithms - The baseline multi-classification models using bag of words approach were applied to the dataset, namely, Logistic… Show more

(Skills: Python, Numpy, Scikit-learn, Matplotlib, Seaborn, Pandas, Nltk, Spacy, Gensim)

Major modules:

1. Data Cleaning - It involved the initial loading of data along with visualization of the class distribution. The duplicated labels were consolidated into one and the duplicated data points were removed.

2. Baseline Multi-Classification Algorithms - The baseline multi-classification models using bag of words approach were applied to the dataset, namely, Logistic Regression, SGDClassifier, Multinomial Naive Bayes etc. The logistic regression with Tf-Idf Vectorizer results in the best Macro F1 score of 0.536.

3. Improved the Macro F1 score obtained using Logistic Regression with Tf-Idf Vectorize to 0.557 using complex text features including n-grams, character n-grams and domain specific features like length of the complaint, number of punctuation marks and uppercase letters in the complaint while putting a limit on the maximum number of features, ignoring stop words and ignoring infrequent words.
The results of the tuned model were analyzed using the confusion matrices, analysis of important feature with their coefficient weights and printing out sample mistakes by the model.

5. Clustering Analysis was done using the dataset by using methodologies like LDA, NMF and K-Means. The best topics obtained using LDA and NMF were visualized with the most important features belonging to every topic. The clustering methodologies were compared based on the calculated ARI score and NMF was shown to have the highest ARI score value.

6. Improved Classification Models using the classes obtained by the Clustering Analysis done in the previous module using NMF, LDA and K-Means. We reassigned the classes based on the clustering analysis done using the techniques and they resulted in improved F1 scores, 0.79 for NMF, 0.64 using LDA and 0.84 using K-Means. Further, word2vec vector embedding was used to improve classification models with improved features extracted from text. Show less

(Skills Used: Python, R, D3.js, Matplotlib, ggplot2)

I have been an avid runner and fitness enthusiast for the past 5 years, and what has been a fun part of all those is how much we can learn about ourselves from exploring the data we capture.

I started capturing data about myself around the year 2012, which was when we saw the advent of wearable devices, and have always wanted to carry out an analysis to gain insights about my running patterns, my strength training schedules, my… Show more

(Skills Used: Python, R, D3.js, Matplotlib, ggplot2)

I have been an avid runner and fitness enthusiast for the past 5 years, and what has been a fun part of all those is how much we can learn about ourselves from exploring the data we capture.

I started capturing data about myself around the year 2012, which was when we saw the advent of wearable devices, and have always wanted to carry out an analysis to gain insights about my running patterns, my strength training schedules, my diet logs and any other metadata which I can collect about myself like the Heart Rate, Cadence for runs and also the step counts.

Show less

(Skills: Numpy, Scipy, Python, Matplotlib)

Project 1: Linear and Ridge Regression
Implemented Ridge Regression to predict mileage per gallon for various car models in an automobile dataset. Also, visualized the results of how the importance of various features vary with the change in regularization parameter. Further, implemented pth-order polynomial regression to observe how it affects the RMSE as a function of regularization parameter for various orders of p = 1, 2… Show more

(Skills: Numpy, Scipy, Python, Matplotlib)

Project 1: Linear and Ridge Regression
Implemented Ridge Regression to predict mileage per gallon for various car models in an automobile dataset. Also, visualized the results of how the importance of various features vary with the change in regularization parameter. Further, implemented pth-order polynomial regression to observe how it affects the RMSE as a function of regularization parameter for various orders of p = 1, 2, 3.

Project 2: Naive Bayes, K-Nearest Neighbors and Logistic Regression Algorithms
Implemented Naive Bayes algorithm to classify a dataset of emails into spam/ham. Secondly, implemented K-Nearest Neighbors algorithm using the same dataset and observed the changes in accuracy with varying values of K. Finally, used the dataset and implemented Logistic Regression to observe the changes in the objective function with the increase in the iterations.

Project 3: Gaussian Process and Ada-boost Algorithm
Implemented Gaussian Process and calculated RMSE by varying the values of the parameters of the model to observe the variation. Also, implemented boosting for the Least Squares Classifier on the dataset found at https://archive.ics.uci.edu/ml/datasets/Occupancy+Detection+

Project 4: K-Means Algorithm and Probabilistic Matrix Factorization
Implemented the K-means algorithm to cluster data generated from a weighted mixture of Gaussian 
distributions. Also, predicted movie ratings and found ‘similar’ movies in the MovieLens dataset by implementing Maximum A Posteriori (MAP) inference Probabilistic Matrix Factorization collaborative filtering approach.

Project 5: First Order Markov Chain Model and Non-Negative Matrix Factorization
Implemented a First Order Markov Chain Model to rank 760 college football teams using game scores from the 2016 college football season. Also, implemented Non-Negative Matrix Factorization algorithm to do topic modeling on a New York Times data set. Show less

(Skills Used: HTML, Javascript, D3.js, Jquery)

Since the 1930s, Marvel and DC Comics have had a friendly rivalry, stemming from the comic book pages and now transcending onto the big screen. Many of them have deep and well-established backgrounds that go beyond just pages and screens, but where do they come from in the first place? How are they different? Are Marvel characters stronger than their DC counterparts?

We use visualizations created using D3.js to compare their origins,… Show more

(Skills Used: HTML, Javascript, D3.js, Jquery)

Since the 1930s, Marvel and DC Comics have had a friendly rivalry, stemming from the comic book pages and now transcending onto the big screen. Many of them have deep and well-established backgrounds that go beyond just pages and screens, but where do they come from in the first place? How are they different? Are Marvel characters stronger than their DC counterparts?

We use visualizations created using D3.js to compare their origins, superpowers, and other attribute differences between them! Show less

The work involved analyzing the the code data of several JAVA experts to discern their skill-set, tools and techniques used most often by them. Further, we looked into ways to tag and classify these skill-sets into specific subfields for the ease of visualisation.
(Tools Used: Numpy, Pandas, Scikit-Learn, NetworkX, D3.js, Flask, AWS Hosting)

The project was done as a honors project for the Bachelors in Technology Degree. We created a visual experience for important historical monuments like the Golconda fort in Hyderabad. The visual experience was created by stitching together a lot of captured images in the form of a 3D image synths embedded within video captured at places of high interest. The work was later extended to involve automatic annotation of Images taken at Golconda Fort via a mobile device.

Tools Used: C… Show more

The project was done as a honors project for the Bachelors in Technology Degree. We created a visual experience for important historical monuments like the Golconda fort in Hyderabad. The visual experience was created by stitching together a lot of captured images in the form of a 3D image synths embedded within video captured at places of high interest. The work was later extended to involve automatic annotation of Images taken at Golconda Fort via a mobile device.

Tools Used: C, Bundler, Flash, SQL, Python, Symbian
Show less