Machine Learning and Analytics Portfolio

It's time for spring cleaning and I have been compiling some of my portfolio in Machine Learning and Analytics. I wrote up some of my work which you may find useful if you are learning how to approach a data analysis task and have little experience.

Along with the reproducible code examples, I explain my 4-step iterative approach to machine learning and predictive analytics :

1. Data cleaning and Exploratory data analysis: Always develop expectations from the data before starting. This includes data cleaning, Natural Language Processing (NLP), for text features, fitting simple regression models for missing value imputation, scaling and transformations, dimension reduction and feature engineering.

2. Always start linear: start fitting linear models to 'clean' data, including regularized models like Ridge, Lasso and Elastic Net. My hope is to build an intuition about the performance of these linear approaches in different problems and data sets, whether it is a classification or regression question. Understanding how much signal we can extract from a noise basket using linear models is essential, because linear models are the ones with the best interpretability.

3. Train Machine Learning Algorithms: train both kernel-based (such as Support Vector Machines (SVM) and tree-based learners (such as RandomForest or Boosted trees), develop an intuition about their performance for a particular problem. Performing cross-validation and trying to understand for when the models start overfitting. Developing an idea of the computational cost of building a particular classifier or regression model, for a given set of features and hyperparameters.

4. Hyperparameter tuning (optimization): once there is an understanding of the performance of different models and computational costs associated with them, I perform hyperparameter optimization. This involves considerations involving some intuition about how hyperparameters might effect the performance of the model, how much time we can effort to wait, what would be a sensible hyperparameter space to scan, among others. Use Bayesian Optimization, GridSearch or RandomSearch with cross-validation. Whenever possible, build a pipeline and perform hyperparameter optimization using the entire pipeline.

Implementation: There are many computer languages you can use, and many libraries are available to install for a given language. Python and R are probably the most popular languages by which you can handle almost all data analysis tasks today. Both of them are open source, and they are backed by a large community. I believe using both R and Python makes a powerful combination, also depending on preferences of your team. I use R to explore data sets, which has an unmatched data visualization capacity thanks to user friendly libraries like ggplot2. Data cleaning process is a blast thanks to libraries like dplyr. Machine learning tasks are greatly simplified using the caret package. I use Python in parallel to R, to perform NLP, and to train advanced machine learning algorithms. Sklearn API and its integration with many built in Python libraries, ease of configuration for distributed computation, makes Python my choice when it comes to building machine learning pipelines and hyperparameter optimization.

A framework based on Natural Language Processing, Gradient Boosting and Deep Learning to predict retail prices.

Retail Price Suggestion

Predictive modeling of human activity recognition: explore the data and see the performance of a RandomForest classifier.

Quantified Self

Neural network optimization using dropout regularization

Deep Learning Optimization

Predicting housing market in Iowa: extreme gradient boosting solution to a traditional regression problem.

Housing Market

Bayesian optimization for Hyperparameter Tuning of XGboost classifier

Bayesian Optimization

Geospatial analysis of severe weather events from U.S. National Oceanic and Atmospheric Administration(NOAA) storm database.

Storm Events and Public Health

Streaming data and ensemble classifier for incremental learning

Incremental learning

Unsupervised learning and pattern extraction using time-series data

Dynamic Time Warping

Pipeline-based feature selection strategies to monitor classifier performance

Feature selection screening

JurkatQC Scraper: analytics solution for real-time monitoring of streaming Mass Spectrometer data.

Real-time MS quality analytics

GSEA taskforce: a standalone software for user-friendly Gene Set Enrichment Analysis.

Gene Set Enrichment Analysis

Companion: monitor real-time anti-cancer drug adverse events using openFDA API

Real-time drug adverse event monitoring

Anomaly Detective: a tool to quantify anomalies in time-series data

Anomaly detection in time-series

Web scraping using selenium and python: simulating browser behavior while scraping data

Web scraping using selenium

A deep dive into the historical data set for feature engineering and building classifiers.

Predicting Survival in Titanic