Press "Enter" to skip to content

Category: Data Science

Repeated Cross-Validation in R

Published 2019-05-07 by Kevin Feasel

Ludvig Olsen walks us through a couple of nice R packages:

The benefits of using groupdata2 to create the folds are 1) that it allows us to balance the ratios of our output classes (or simply a categorical column, if we are working with linear regression instead of classification), and 2) that it allows us to keep all observations with a specific ID (e.g. participant/user ID) in the same fold to avoid leakage between the folds.

The benefit of cvms is that it trains all the models and outputs a tibble (data frame) with results, predictions, model coefficients, and other sweet stuff, which is easy to add to a report or do further analyses on. It even allows us to cross-validate multiple model formulas at once to quickly compare them and select the best model.

Ludvig also gives us some examples of how both packages can help you out. H/T R-Bloggers

Comments closed

Exploratory Data Analysis on Categorical Variables

Published 2019-05-03 by Kevin Feasel

Giorgio Garziano continues digging into earthquake data:

To understand relationship or dependencies among categorical variables, we take advantage of various types of tables and graphical methods. Also stratifying variables can be encompassed in order to highlight if the relationship between two primary variables is the same or different for all levels of the stratifying variable under consideration.

The contingency table are said to be of one-way flavor when involving just one categorical variable. They are said two-way when involving two categorical variables, and so on (N-way).

Read on for various techniques for data analysis against categorical variables.

Comments closed

Techniques for Determining Cluster Counts

Published 2019-04-30 by Kevin Feasel

Stephanie Glen has an image which describes different techniques for figuring out how many clusters you have:

If you want to determine the optimal number of clusters in your analysis, you’re faced with an overwhelming number of (mostly subjective) choices. Note that there’s no “best” method, no “correct” k, and there isn’t even a consensus as to the definition of what a “cluster” is. With that said, this picture focuses on three popular methods that should fit almost every need: Silhouette, Elbow, and Gap Statistic.

Click through for the picture and references.

Comments closed

Predicting Database Growth

Published 2019-04-30 by Kevin Feasel

James Livingston uses linear regression to plot database growth over time:

Utilizing the equation for a line, instead of solving for y we will solve for x, where:
– x corresponds to the day we will hit capacity based on current growth rate
– y corresponds to drive capacity in GB
– m is the slope of our regression line, provided by the model via lm.coef_
– b is the intercept of the regression line, also provided by the model via lm.intercept_

Click through for an example. This is one of the areas where DBAs can gain a lot by learning a bit of data science.

Comments closed

Exploratory Analysis of Earthquake Data

Published 2019-04-29 by Kevin Feasel

Giorgio Garziano walks us through an earthquake data set:

Boxplots for each quantitative variables are shown. We take advantage of the quantitative variable names (quantitative_vars) determined before to apply a ggplot2 package based boxplot function. The Y axis labeling and title are determined by the variable to be plot. Further, legend is not displayed and we adopt the coordinate flip option for improved readability.

Check it out to get an idea of how to do exploratory data analysis.

Comments closed

Sentiment Analysis with Spark on Qubole

Published 2019-04-24 by Kevin Feasel

Jonathan Day, et al, have a tutorial on using Qubole to build a sentiment analysis model:

This post covers the use of Qubole, Zeppelin, PySpark, and H2O PySparkling to develop a sentiment analysis model capable of providing real-time alerts on customer product reviews. In particular, this model allows users to monitor any natural language text (such as social media posts or Amazon reviews) and receive alerts when customers post extremely nice (high sentiment) or extremely negative (low sentiment) comments about their products.

In addition to introducing the frameworks used, we will also discuss the concepts of embedding spaces, sentiment analysis, deep neural networks, grid search, stop words, data visualization, and data preparation.

Click through for the demo.

Comments closed

Running Spark MLlib to Feed Power BI

Published 2019-04-24 by Kevin Feasel

Brad Llewellyn shows how you can take Spark MLlib results and feed them into Power BI:

MLlib is one of the primary extensions of Spark, along with Spark SQL, Spark Streaming and GraphX.  It is a machine learning framework built from the ground up to be massively scalable and operate within Spark.  This makes it an excellent choice for machine learning applications that need to crunch extremely large amounts of data.  You can read more about Spark MLlib here.

In order to leverage Spark MLlib, we obviously need a way to execute Spark code.  In our minds, there’s no better tool for this than Azure Databricks.  In the previous post, we covered the creation of an Azure Databricks environment.  We’re going to reuse that environment for this post as well.  We’ll also use the same dataset that we’ve been using, which contains information about individual customers.  This dataset was originally designed to predict Income based on a number of factors.  However, we left the income out of this dataset a few posts back for reasons that were important then.  So, we’re actually going to use this dataset to predict “Hours Per Week” instead.

Check it out. And Brad’s not joking when he says the resulting model is terrible. But that’s okay, because it was never about the model.

Comments closed

Time Series Analysis with Forensic Accounting

Published 2019-04-24 by Kevin Feasel

I have another part up in my series on forensic accounting techniques:

This post will be a pretty short one. In my talk, I don’t have any demos, mostly because much of cohort analysis has secretly been time series analysis at the same time. Instead, I’ll lob out a few points and call it a day.

Time series analysis, at its core, is all about how your data changes over time. The grain for time series analysis is important: as we saw in the last post, we were able to get an excellent result at the yearly level when regressing number of active buses versus number of line items.

Spoilers: it’s not as short as I thought it would be.

Comments closed

Naive Bays in R

Published 2019-04-23 by Kevin Feasel

Zulaikha Lateef takes us through the Naive Bayes algorithm and implementations in R:

Naive Bayes is a Supervised Machine Learning algorithm based on the Bayes Theorem that is used to solve classification problems by following a probabilistic approach. It is based on the idea that the predictor variables in a Machine Learning model are independent of each other. Meaning that the outcome of a model depends on a set of independent variables that have nothing to do with each other. 

Naive Bayes is one of the simplest algorithms available and yet it works pretty well most of the time. It’s almost never the best solution but it’s typically good enough to give you an idea of whether you can get a job done.

Comments closed

Forensic Accounting: Cohort Analysis

Published 2019-04-19 by Kevin Feasel

I continue my series on forensic accounting techniques with cohort analysis:

In the last post, we focused on high-level aggregates to gain a basic understanding of our data. We saw some suspicious results but couldn’t say much more than “This looks weird” due to our level of aggregation. In this post, I want to dig into data at a lower level of detail. My working conception is the cohort, a broad-based comparison of data sliced by some business-relevant or analysis-relevant component.

Those familiar with Kimball-style data warehousing already understand where I’m going with this. In the basic analysis, we essentially look at fact data with a little bit of disaggregation, such as looking at data by year. In this analysis, we introduce dimensions (sort of) and slice our data by dimensions.

Click through for some fraud-finding fun.

Comments closed