Press "Enter" to skip to content

Category: Data Science

Fun with Residual Plots

Published 2019-09-30 by Kevin Feasel

Nina Zumel explains why, when plotting residuals, you always put predictions on the X axis and residuals on the Y axis:

One reason that the proper residual graph (for a well fit model) should smooth out to the line y=0 is known as reversion to mediocrity, or regression to the mean.

Imagine that you have an ideal process that always produces a single value y. You don’t actually observe this “true value”; instead, what you observe is y plus (IID, zero mean) noise. You can build a “model” for this process that predicts the mean of the observations, in this case the value 0.1033149. Then you can calculate the residuals of your “model” in the usual way.

This post went in a direction I wasn’t expecting, and it was all the better for it.

Comments closed

Topic Modeling

Published 2019-09-30 by Kevin Feasel

Federico Pascual has an article on topic modeling and topic classification:

Topic modeling is an unsupervised machine learning technique that’s capable of scanning a set of documents, detecting word and phrase patterns within them, and automatically clustering word groups and similar expressions that best characterize a set of documents. It’s known as ‘unsupervised’ machine learning because it doesn’t require a predefined list of tags or training data that’s been previously classified by humans.

Since topic modeling doesn’t require training, it’s a quick and easy way to start analyzing your data. However, you can’t guarantee you’ll receive accurate results, which is why many businesses opt to invest time training a topic classification model.

The article is long but worth the read, with examples in Python and additional notes for R.

Comments closed

A Primer on Survey Analysis

Published 2019-09-23 by Kevin Feasel

Federico Pascual has a long primer on survey analysis:

When it comes to customer feedback, you’ll find that not all the information you get is useful to your company. This feedback can be categorized into non-insightful and insightful data. The former refers to data you had already spotted as problematic, while insightful information either helps you confirm your hypotheses or notice new issues or opportunities. 

Let’s imagine your company carries out a customer satisfaction survey, and 60% of the respondents claim that the pricing of your product or service is too high. You can use that valuable data to make a decision. That’s why this data is also called actionable insights because they either lead to action, validation, or rethinking of a specific strategy you have already implemented. 

Survey design and implementation can be pretty difficult. This article does a good job pushing you away from some of the pitfalls around it.

Comments closed

Linear Regression in Power BI

Published 2019-09-23 by Kevin Feasel

Joseph Yeates shows how to implement linear regression in Power BI:

The goal of a simple linear model is to fit a line onto this plot to summarize the shape of the data using the equation above.

The “a” value is the slope of the fitted line (rise over run) and the “b” value is the intercept on the y-axis (when x is equal to zero).

In the gapminder example, the life expectancy column was assigned as the “y” variable, as it is the outcome that we are interested in predicting or understanding. The year1950 column was assigned as the “x” variable, as it is what we are using to try and measure the change in life expectancy.

This is a little more complicated than adding a regression line to a scatterplot (the “normal” way to do linear regression with Power BI) but this method lets you work with the outputs in a way that the normal method doesn’t.

Comments closed

Python and R Data Reshaping

Published 2019-09-10 by Kevin Feasel

John Mount takes us through a couple of data shaping packages:

The advantages of data_algebra and cdata are:

– The user specifies their desired transform declaratively by example and in data. What one does is: work an example, and then write down what you want (we have a tutorial on this here).
– The transform systems can print what a transform is going to do. This makes reasoning about data transforms much easier.
– The transforms, as they themselves are written as data, can be easily shared between systems (such as R and Python).

Let’s re-work a small R cdata example, using the Python package data_algebra.

Click through for the example.

Comments closed

When to Use Different ML Algorithms

Published 2019-09-09 by Kevin Feasel

Stefan Franczuk explains the different categories of machine learning algorithms available in Talend:

Clustering is the task of grouping together a set of objects in such a way, that objects in the same group are more similar to each other than to those in other groups. Clustering is really useful for identify separate groups and therefore is used to solve use cases such as “who are my premium customers?”.

Understanding when to use which algorithm is important. You don’t want to build out the world’s best regression if your benefactors are asking for a classifier.

Comments closed

Exploratory Data Analysis with ExPanDaR

Published 2019-09-04 by Kevin Feasel

Joachim Gassen walks us through the ExPanDaR package in R:

The ‘ExPanDaR’ package offers a toolbox for interactive exploratory data analysis (EDA). You can read more about it here. The ‘ExPanD’ shiny app allows you to customize your analysis to some extent but often you might want to continue and extend your analysis with additional models and visualizations that are not part of the ‘ExPanDaR’ package.

Thus, I am currently developing an option to export the ‘ExPanD’ data and analysis to an R Notebook. While it is not ready for CRAN yet, it seems to work reasonably well and I would love to see some people trying it and letting me know about any bugs or other issues that they encounter. Hence, this blog post.

Looks like an interesting package. H/T R-bloggers

Comments closed

Calculating Consistency of Ratings

Published 2019-09-03 by Kevin Feasel

Sebastian Sauer looks at computing reliability between raters:

Computing inter-rater reliability is a well-known, albeit maybe not very frequent task in data analysis. If there’s only one criteria and two raters, the proceeding is straigt forward; Cohen’s Kappa is the most widely used coefficient for that purpose. It is more challenging to compare multiple raters on one criterion; Fleiss’ Kappa is one way to get a coefficient. If there are multiple criteria, one way is to compute the mean of multiple Fleiss’ coefficients.

However, a different way, and the way presented in this post, consists of checking of all raters agree on one given item (and repeating that for all items). If rater A assigns two tags/criteria (tag1, tag2) to item A, then the other raters may not assign different tags (eg tag3, tag4) to that item, if a match should be scored. Note that this proceeding allows for different numbers of tags/criteria for the items (eg., item 1 with only 1 tag, but item 2 with 3 tags etc.). However, our grading should give some points, if, say, rater1 assigns tag1 and tag2, but raters 2 and 3 only assign tag1.

Read the whole thing.

Comments closed

Sampling and Estimating Rare Events

Published 2019-08-28 by Kevin Feasel

Yi Liu takes us through a process to estimate rare events:

Naturally, we get an unbiased estimate of the overall prevalence of violation if we sample the videos uniformly from the population and have them reviewed by human raters to estimate the proportion of violating videos. We also get an unbiased estimate of the violation rate in each policy vertical. But given the low probability of violation and wanting to use our rater capacity wisely, this is not an adequate solution — we typically have too few positive labels in uniform samples to achieve an accurate estimate of the prevalence, especially for those sensitive policy verticals. To obtain a relative error of no more than 20%, we need roughly 100 positive labels, and more often than not, we have zero violation videos in the uniform samples for rarer policies.

This is similar in nature to testing for rare diseases, where a random sample of N people in the population is likely to turn up 0 cases of it.

Comments closed