Welcome to My Data Science Portfolio!

The portfolio showcases a variety of projects that include data wrangling, exploratory analysis, data mining, advanced statistical techniques, psychometric analysis, predictive modeling with machine learning algorithms, and data visualization. The datasets used for these projects come from various domains, such as human behavior, marketing, finance, sports analytics, the entertainment industry, demographics, and more.
You can take a walk through the forest of my work below if you are interested.

Of course I am interested, that is why I am here. In fact, I can't wait to check it out.

About me

Hello! I'm a data scientist specializing in applied statistics, psychometrics, and behavioral research methodology. My background in psychology and psychometrics provides a unique advantage in analyzing and interpreting data related to human behavior. I have over 10 years of overall experience in data analysis, with over 5 years dedicated to measurement and advanced statistical techniques. Additionally, I possess a strong proficiency in machine learning algorithms, allowing me to effectively tackle complex data problems. I'm also nearing the completion of my Ph.D. in psychometrics. As a person, I think outside the box and am a continuous learner, always seeking to expand my knowledge. I love reading whenever possible and thrive on both large-scale projects and minor assignments. I especially enjoy tasks that require integrating abstract concepts to achieve simple, yet impactful solutions. My personal favorites are projects that involve advanced statistical techniques, clustering, predictive modeling, or forecasting.

Creative Outlook on Data

Domain expertise in psychometrics and psychology assists me in reasoning about findings, being aware of cognitive biases, and communicating insights to diverse audiences, ranging from basic to advanced levels of understanding. Over time, I have acquired keen divergent and lateral thinking skills, which aid with creating new variables, asking questions, and deciding how to solve the problem and obtain insights. At the same time, my thinking can be analytical and logical, so I adjust the approach based on the task at hand. My problem-solving abilities have been further sharpened through practical experience with data analysis, scientific research, and even in leisure activities like playing chess.

Systematic Problem-solver

I love solving problems of mathematical and/or psychological nature. Discovering solutions to complex problems and presenting them in a straightforward way give me buzz. Over time, I have become foolproof to frustration and discouragement. I enjoy devoting days, even weeks, working on a particular problem and learning. As a result, I have gained technical skills, such as statistical programming, machine learning, and RDBMS. My primary tool is R, though I am able to work in Python when required. Additionally, I'm familiar with various software packages such as SPSS, Tableau, and Orange, and I actively seek opportunities to expand my technological toolkit. For instance, I built this website myself.

Portfolio Projects

Predictive Modeling for Bank Customer Churn

This project, conducted in R, uses the dataset about 10,000 customers of a particular bank with branches in France, Germany, and Spain. I did feature engineering, conducted EDA, and selected predictors of customer churn using random forest algorithm. To address the class imbalance in the outcome variable, I employed a combination of undersampling and oversampling, using the ROSE algorithm. Then I trained a blending ensemble comprising random forest, C5.0 tree, and support vector machine with Gaussian radial basis kernel function for binary classification. To expedite computation, I parallelized the procedure. The ensemble performed excellently in churn prediction, as indicated by accuracy, sensitivity, specificity, and AUC in excess of .95 on test dataset.

A Comprehensive Analysis of Automobile Extinction

This project, conducted in R, uses a dataset with multiple tables about American car models production from 1896 to 2014. I joined and cleaned the tables, did feature engineering, conducted EDA, and used hierarchical multiple zero-inflated negative binomial regression to examine the predictors of the number of years a car model is in production. Car manufacturers were segmented based on car model production policies in terms of car model life span using k-medoids cluster analysis. After that, historical time series analysis with Poisson model was used to check how economic conditions influence car model production for each segment. Some useful insights about economic conditions, car model competition, and car manufacturer segments were obtained.

R Shiny App: Visualizing Probability Distributions

This project, developed in R, represents a deployed Shiny web application. This interactive application covers various statistical distributions including Gaussian, t, F, chi-square, uniform, Bernoulli, binomial, Poisson, exponential, gamma, inverse Gaussian, Gompertz, beta, log-normal, Weibull, Pareto, geometric, hypergeometric, Cauchy, log-Cauchy, Rayleigh, Laplace, etc. Users can adjust parameters for each distribution and see real-time updates, allowing them to observe the impact on the distribution's shape. This functionality makes the application particularly valuable for teaching and learning purposes, helping users gain a deeper understanding of particular statistical distribution through hands-on interaction.

Data Mining on Black Friday Transactions Data

This project, conducted in R, uses the Black Friday dataset on 537,577 transactions that was previously used in a predictive analytics contest organized by Analytics Vidhya. In this project, association rules on demographics and purchasing behavior were used instead of predictive modeling. After inspecting variable distributions, I discretized numeric variables using k-means clustering to make the discretization objective. Afterward, apriori algorithm was used to discover the rules with above regular amounts of purchases for each product category on the right-hand side. Useful rules with high confidence and lift that are applicable for a marketing campaign were found and visualized using a network plot.

Fraud detection in Ethereum Transactions Data

This project, conducted in R, uses the dataset on 9,841 known fraudulent and non-fraudulent transactions provided by Ethereum. Initially, data preparation and feature engineering were conducted, revealing that patterns of missing data could serve as an effective means of identifying potential fraud. In the following analysis, I used random forest algorithm for predictor selection. Then I used a blending ensemble of random forest, C5.0 tree, gradient boosting machine, and k-nearest neighbors for the binary classification problem. To address the class imbalance in the outcome variable, I employed the SMOTE algorithm. The ensemble performed excellently in fraud detection, as indicated by accuracy and specificity in excess of .95, sensitivity approaching .90, and AUC in excess of .90.

Clustering USA Communities Based on Crime

This project, conducted in R, uses the dataset that combines socio-economic data from the 1990 US Census, law enforcement data from the 1990 US LEMAS survey, and crime data from the 1995 FBI UCR about 2,215 communities. After joining the datasets, cleaning them up, and handling missing values through dropping variables and single imputation using the random forest algorithm, principal component analysis was used for dimension reduction before the cluster analysis. Afteraward, I employed hierarchical DBCAN algorithm to cluster the communities based on crime-related variables and found six meaningful clusters in terms of crime rate profiles, which can be useful for policymakers, law enforcement, and urban planners to tailor interventions to specific community crime profiles.

EDA and Customer Segmentation for Online Retail

This project, conducted in R, analyzes a dataset on 541,909 transactions from a UK-based online retail platform over a one-year period. I cleaned the dataset, performed feature engineering, and EDA. Insights were gained on cancellations, the most profitable products, customers, areas, and periods. Subsequently, I used exploratory factor analysis and winsorization to prepare the data for segmentation. Afterward, divisive hierarchical cluster analysis and fuzzy c-means cluster analysis were employed to segment customers based on transaction behavior. The discovered segments were described in terms of transaction behavior metrics, such as total spending, average spending per transaction, and the quantity of products purchased, which can be used for marketing purposes.

Lahman's Baseball 1871-2020 - EDA and Clustering

This project, conducted in R, uses Lahman's dataset on baseball that covers the period from 1871 to 2020. I joined several tables and analyzed player induction into the Hall of Fame based on performance, origin, and voting groups. EDA was conducted, followed by feature engineering with Bayesian rates representing player performance. Random forest algorithm was used to discover importances of particular predictors of entrance into Hall of Fame. Moreover, exploratory factor analysis combined with hierarchical agglomerative cluster analysis and visualization were used to gains insights into player performance aspects, the origin of players, and voter group representation for the Hall of Fame entrance over time. Some insights were gained into the Hall of Fame induction.

Video Games Clustering and CA

This project, conducted in R, uses RAWG Video Games database that consists of variables related to 474,417 video games. After EDA, games were clustered based on their features using hierarchical DBSCAN algorithm and based on RAWG user behavior using k-means clustering. Flexible discriminant analysis with polynomial and ridge regression, respectively, was used to get additional insights on cluster separation. Correspondence analysis (CA) was used to explore the relationship between the two clusterings. It was found which game features correspond to specific categories of popularity among RAWG users, providing insights into video game types and user preferences. Such findings have practical implications for game development, marketing, and research.

HR analytics - Analysis of Absence & Churn

This project, conducted in R, is based on a synthetic dataset about 311 employees, so the results are not generalizable to any real situations. However, the insights would be useful for the fictitious company. I was able to heavily rely on domain expertise in psychology for this analysis. Following an EDA, Poisson regression and multilevel binary logistic regression were used to gain insights into the processes related to employee absence and churn. This is an example of situations in which negative findings are also informative for insights about organizational processes and potential improvement within the organization as a whole and within particular departments. Recommendations were provided on further actions that could be taken if similar insights were found in a real-world setting.

Animal census in Nepal - EDA, CA, and Clustering

This project, conducted in R, uses a dataset created from the results of a livestock census of 2,128 smallholder farmers within two districts of Nepal (Nuwakot and Dhading). I did feature engineering and conducted EDA. Correspondence analysis (CA) followed by agglomerative hierarchical cluster analysis combined with k-means clustering to consolidate the solution were conducted on villages based on specific and general livestock category counts. It was discovered which livestock categories contribute the most to village differentiation, which villages are associated with which livestock category, and how the villages are clustered in terms of livestock category prevalence and approaches to raising animals.

Personality across the World - Multigroup CFA

This project, conducted in R, uses the dataset with IPIP-50 personality items on 19,719 participants. Multigroup confirmatory factor analysis (CFA) was used to assess measurement invariance across English as a native language and handedness, including configural, metric, scalar, strict, latent variances and covariances, and latent means levels. To enhance the computation speed, I parallelized the process. Partial measurement invariance was achieved for English as a native language up to the strict invariance level and for handedness up to the metric invariance level, suggesting that IPIP-50 items can be used confidently across different language and handedness groups, while acknowledging that full invariance across the groups was not achieved.

FAANG - Forecasting Stock Prices

This project, conducted in R, uses data on daily stock prices of each FAANG company (Facebook/Meta, Amazon, Apple, Netflix, and Google/Alphabet) from their inception up to the beginning of February 2023. After performing univariate time series analysis and decomposition into seasonal, trend, and noise components, I fitted a vector error-correction model (VECM) to explore the long-term relationships among the FAANG stock prices. It was found that FAANG companies' stock prices are cointegrated, suggesting they share a common long-term equilibrium despite short-term fluctuations. Granger causality tests within the VECM framework highlighted mutually predictive relationships, such as Facebook/Meta influencing Amazon and Netflix influencing Facebook/Meta.

Mushrooms - Predictive Modeling of Edibility

This project, conducted in R, a dataset comprising 8,123 hypothetical mushrooms representing 23 species of gilled mushrooms was used. This analysis aimed to address a typical binary classification problem, focusing on determining whether the mushrooms are poisonous. After discovering the missing value pattern was not random using the apriori algorithm, multiple imputation with random forest was conducted. Multiple imputation was parallelized for faster computation. Random forest was used for predictor selection. A blending ensemble of random forest, C5.0 tree, and neural network was trained. The ensemble performed exceptionally well on the test dataset, with perfect accuracy, sensitivity, specificity, and ROC, which is similar to some previous attempts on this dataset.

Zambia Elections - Comparative Analysis

This project, conducted in R, uses a dataset on 1,764 participants related to elections in Zambia. After data preparation, I applied the random forest algorithm to select the predictors of voting choices before and after the elections. Afterward, decision tree algorithm with voting choices before and after the election as outcome variables was used for the comparative analysis and to provide more insights about processes behind the choices. This analysis reveals a shift in voter preferences from ideological considerations in terms of democraticness before the elections to performance-based evaluations of party cadres afterward. It also emphasizes the critical influence of perceived representation of interests of voter's ethnic group on voting choices.

I offer my expertise in the following areas for freelance projects:

  • Data wrangling/Cleaning/Preparation/Munging
  • Data visualization
  • Exploratory data analysis
  • Data analysis on datasets from industry or academia
  • Psychometric analysis
  • Experiment design
  • Data mining
  • Clustering
  • Predictive modeling
  • Time series forecasting
  • R Shiny app development
  • Pre-submission peer-review of your scientific paper in social sciences

Techniques in my toolkit:

  • Regression analysis (linear, polynomial, logistic, Poisson, quasi-Poisson, negative binomial, quantile, robust, partial least squares, principal components regression)
  • Non-parametric statistical tests (Sign Test, Wilcoxon Signed-Rank Test, Mood's Median Test, Mann-Whitney U Test, Friedman Test, Kruskal-Wallis Test, McNemar's Test, Fisher's Exact Test, Cochran's Q Test, Chi-square Test, Barnard's Test)
  • (M)AN(C)OVA
  • Monte Carlo Simulation
  • Cluster analysis (k-means, k-medoids, Gaussian mixture modeling, fuzzy, hierarchical, (H)DBSCAN, OPTICS)
  • Latent variable modeling (FA, path analysis, full SEM, LCA, LPA)
  • Multilevel modeling (regression, FA, SEM)
  • Correspondence analysis (multiple, canonical, detrended)
  • Discriminant analysis (linear, quadratic, flexible, shrinkage)
  • Survival analysis (Kaplan-Meier, Cox proportional Hazards)
  • Time series analysis and forecasting (univariate-(S)AR(F)(I)(MA)(X) & (G)ARCH, multivariate-VAR, VECM)
  • Machine learning (shrinkage regression, decision trees, random forests, GBM, XGBoost, k-NN, SVM, naive Bayes classifier, neural networks, market basket analysis, and more)

Reach out if you want to hire me, outsource an assignment, or team up for something: