How do we build relatable machine learning models that regular people can understand? This is a presentation about how design principles apply to the development of machine learning systems. Too often in data science, machine learning software is not built with regular people who will interact with it in mind.
I argue that in order to make machine learning software relatable, we need to use design thinking to intentionally build in mechanisms for users to form their own mental models of how the machine learning software works. Failing to include theses helps cultivate the common sense that machine learning is a black box for users.
I gave three different versions of this talk at Quant UX Con on June 8th, 2022, the Royal Institute of Anthropology’s annual conference on June 10th, 2022, and Google’s AI + Design Tooling Research Symposium on August 5th, 2022.
I hope you find it interesting and feel free to share any thoughts you might have.
Thank you for the conference and talk organizers for making this happen, and I appreciate all the insightful conversations I had about the role of design thinking in building relatable machine learning.
Hello, my name is Stephen Paff. I am a data scientist and an ethnographer. The goal of this blog is to explore the integration of data science and ethnography as an exciting and innovative way to understand people, whether consumers, users, fellow employees, or anyone else.
I want to think publicly. Ideas worth having develop in
conversation, and through this blog, I hope to present my integrative vision so
that others can potentially use it to develop their own visions and in turn
help shape mine.
Please Note: Because my blog straddles two technical areas, I will split my posts based on how in-depth they go into each technical expertise. Many posts I will write for a general audience. I will write some posts, though, for data scientists discussing technical matters within that field, and other posts will focus on technical topics withn ethnography for anthropologists and other ethnographers. At the top of each post, I will provide the following disclaimers:
As a data scientist and ethnographer, I have worked on many types of research projects. In professional and business settings, I am excited by the enormous growth in both data science and ethnography but have been frustrated by how, despite recent developments that make them more similar, their respective teams seem to be growing apart and competitively against each other.
Within academia, quantitative and qualitative research methods have developed historically as distinct and competing approaches as if one has to choose which direction to take when doing research: departments or individual researchers specialize in one or the other and fight over scarce research funding. One major justification for this division has been the perception that quantitative approaches tend to be prescriptive and top-down compared with qualitative approaches which tend to be to descriptive and bottom-up. That many professional research contexts have inherited this division is unfortunate.
Recent developments in data science draw parallels with qualitative research and if anything, could be a starting point for collaborative intermingling. What has developed as “traditional” statistics taught in introductory statistics courses is generally top-down, assuming that data follows a prescribed, ideal model and asking regimented questions based on that ideal model. Within the development of machine learning been a shift towards models uniquely tailored to the data and context in question, developed and refined iteratively.[i] These trends may show signs of breaking down the top-down nature of traditional statistics work.
If there was ever a time to integrate quantitative data science and qualitative ethnographic research, it is now. In the increasingly important “data economy,” understanding users/consumers is vital to developing strategic business practices. In the business world, both socially-oriented data scientists and ethnographers are experts in understanding users/consumers, but separating them into competing groups only prevents true synthesis of their insights. Integrating the two should not just include combining the respective research teams and their projects but also encouraging researchers to develop expertise in both instead of simply specializing in one or the other. New creative energy could burst forth when we no longer treat these as distinct methodologies or specialties.
[i] Nafus, D., & Knox, H. (2018). Ethnography for a Data-Saturated World. Manchester: Manchester University Press, 11-12.
As an anthropologist and data scientist, I often feel caught in the middle two distinct warring factions. Anthropologists and data scientists inherited a historic debate between quantitative and qualitative methodologies in social research within modern Western societies. At its core, this debate has centered on the difference between objective, prescriptive, top-downtechniques and subjective, sitautional, flexible, descritpive bottom-up approaches.[i] In this ensuing conflict, quantative research has been demarcated into the top-down faction and qualitative research within the bottom-up faction to the detriment of understanding both properly.
In my experience on both “sides,” I have seen a tendency among anthropologists to lump all quantitative social research as proscriptive and top-down and thus miss the important subtleties within data science and other quantitative techniques. Machine learning techniques within the field are a partial shift towards bottom-up, situational and iterative quantitative analysis, and business anthropologists should explore what data scientists do as a chance to redevelop their relationship with quantitative analysis.
Shifts in Machine Learning
Shifts within machine learning algorithm development give impetus for incorporating quantitative techniques that are local and interpretive. The debate between top-down vs. bottom-up knowledge production does not need – or at least may no longer need– to divide quantitative and qualitative techniques. Machine learning algorithms “leave open the possibility of situated knowledge production, entangled with narrative,” a clear parallel to qualitative ethnographic techniques.[ii]
At the same time, this shift towards iterative and flexible machine learning techniques is not total within data science: aspects of top-down frameworks remain, in terms of personnel, objectives, habits, strategies, and evaluation criteria. But, seeds of bottom-up thinking definitely exist prominently within data science, with the potential to significantly reshape data science and possibly quantitative analysis in general.
As a discipline, data science is in a uniquely formative and adolescent period, developing into its “standard” practices. This leads to significant fluctuations as the data scientist community defines its methodology. The set of standard practices that we now typically call “traditional” or “standard” statistics, generally taught in introductory statistics courses, developed over a several decade period in the late nineteenth and early twentieth century, especially in Britain.[iii] Connected with recent computer technology, data science is in a similarly formative period right now – developing its standard techniques and ways of thinking. This formative period is a strategic time for anthropologists to encourage bottom-up quantative techniques.
Conclusion
Business
anthropologists could and should be instrumental in helping to develop and
innovatively utilize these situational and iterative machine learning
techniques. This is a strategic time for business anthropologists to do the
following:
Immerse themselves into data science and encourage and cultivate bottom-up quantative machine learning techniques within data science
Cultivate and incorporate (when applicable) situational and iterative machine learning approaches in its ethnographies
For both, anthropologists should use the strengths of ethnographic and anthropological thinking to help develop bottom-up machine learning that is grounded in flexible to specific local contexts. Each requires business anthropologists to reexplore their relationship with data science and machine learning instead of treating it as part of an opposing “methodological clan.” [iv]
[i] Nafus, D.,
& Knox, H. (2018). Ethnography for a Data-Saturated World. Manchester:
Manchester University Press, 11-12
I suspect everyone has seen a bad graph, a mess of bars, lines, pie slices, or what have you that you dreaded having to look at. Maybe you have even made one, which you look at today and wonder what on earth you were thinking.
These graphs violate the most basic graph-making rule in data visualization:
A graph is like a sentence, expressing one idea.
This rule applies to all uses of graphs, whether you are a data scientist, data analyst, statistician, or just making graphs for your friends for fun.
In grade school, your grammar teachers likely explained that a sentence, at its most basic, expresses on thought or idea. Graphs are visual sentences: they should state one and only one thought or idea about the data.
When you look at a graph, you should be able to say, in one sentence, what the graph is saying: such as “Group A is greater than Group B,” or “Y at first improved but is now declining.” If you cannot, then you have yourself a run-on graph.
For example, the above graph is trying to say too many statements: trying to depict the immigration patterns of twenty-two different countries over the course of nearly a century. There are likely useful statements in this data, but the representation as one graph prevents a viewer/reader from being able to easily decipher them.
Likewise, this graph shows way too many lens sizes to meaningfully express a single, coherent idea, leaving the reader/viewer struggling to determine which fields to focus on.
Potential Objection #1: But I have more to say about the data than a single statement.
Great! Then provide more than one graph. Say everything you need to say about the data; just use one graph for each of your statements.
Don’t fall into the One-Graph-to-Rule-Them-All Fallacy: trying to use one graph to express all your statements about the data that ends up a visual mess of incomprehensibility. Create multiple easy-to-read graphs where each graph demonstrates one of your points at a time. Condensing everything into one graph just prevents your viewers from determining what you have to say at all.
One-Graph-to-Rule-Them-All Fallacy: Trying to use one graph to express all your thoughts about the data that ends up a visual mess of incomprehensibility
Instead, use one graph for each of your points
Potential
Objection #2: I want the viewers to interpret the findings for themselves, not
just impart my own ideas/conclusions.
Fair point. When presenting/communicating data, there is a time for showing your own insights and a time to open-endedly display the information for your viewers/readers to interpret for themselves. Graphs are tools for the former, and for the latter, use tables. Tables, among other potential uses, convey a wide scope of information for the reader/viewer to interpret on their own.
Remember that first example above about U.S. immigration from various parts of Europe? A table (see below) would convey that information much more easily and allow readers to track whatever places, patterns, or questions they would to learn about. Are you in a situation where you would like to report a large amount of information that your readers can use for their own purposes? Then tables are a much better starting point than graphs.
Some situations require that I lean towards sharing my insights/analysis and others towards encouraging my readers/viewers to form their own conclusions, but since most situations require a combination of the two, I generally combine graphs and tables. I try, when I can, to put smaller tables in the document or slides themselves and, when I cannot, include full tables in an Appendix.
Potential Objection #3: My main idea/point has multiple subpoints.
Many sentences have multiple subpoints needed to express the single idea as well, which does not prevent the sentence structure from meaningfully capturing those ideas. The fancy grammar word for such a subpoint is a claus. Even though some sentences are simple and straightforward with only one subject and predicate, many (like this very sentence) require multiple sets of subjects and predicates to express its thought.
Likewise, some graphical ideas require multiple subordinate or compounded subpoints, and there are types of graphs that allow this. Consider Joint Plots, like the one below. To present the relationships and combined distribution between the two variables adequately, they also display each variable’s individual distributions above and to the right. That way, the viewer can see how both distributions might be influencing the combined distribution. Thus, it displays each variable’s distribution on the side like a subordinate clause.
The darker colors in this graph signify a higher density of data points, showing the combined joint distribution of the variables.
These are advanced graphs to make, since like with multi-part sentences, one must present the subpoints carefully to make clear what the main point is. Multi-part sentences, likewise, require carefulness in how to organize multiple clauses cohesively. I intend to write a post later describing how to develop these multi-part graphs in more detail.
The general rule still applies for these more complicated graphs:
Can you summarize what the graph is saying in one coherent sentence?
If you cannot, do not use/show that graph. Our brains are very good at intuiting whether a sentence carries one thought, so use this to determine whether your graph is effective.
Here is a list of resources about integrating data science and ethnography. Even though it is an up and coming field without a consistent list of publications, several fascinating and insightful resources do exist.
If there are any resources about integrating data science and ethnography that you have found useful, feel free to share them as well.
Nick Seaver. “The nice thing about context is that everyone has it.” Media, Culture & Society (2015).
Books:
Nafus, Dawn and Hannah Knox. Ethnography for a Data-Saturated World. Manchester: Manchester Univeristy Press, 2018.
Boellstorff, Tom and Bill Maurer. Data, Now Bigger and Better! Chicago: Prickly Paradigm Press, 2015.
Mackenzie, Adrian. Machine Learners: Archaeology of a Data Practice. Cambridge: The MIT Press, 2017.
Examples and Case Studies:
“Autonomous Drive: Teaching Cars Human Behaviour” by Melissa Cefkin on the Youtube Channel DrivingTheNation: https://www.youtube.com/watch?v=6koKuDegHAM
Eslami, Motahhare, et al. “First I “like” it, then I hide it: Folk Theories of Social Feeds.” Curation and Algorithms (2016).
Giaccardi, Elisa, Chris Speed and Neil Rubens. “Things Making Things: An Ethnography of the Impossible.” (2014).
Elish, M. “The Stakes of Uncertainty: Developing and Integrating Machine Learning in Clinical Care.” EPIC (2018).
Madsen, Matte My, Anders Blok and Morten Axel Pedersen. “Transversal collaboration: an ethnography in/of computational social science.” Nafus, Dawn. Ethnography for a Data-saturated World. Manchester: Manchester Univeristy Press, 2018.
Thomas, Suzanne, Dawn Nafus and Jamie Sherman. “Algorithms as fetish: Faith and possibility in algorithmic work.” Big Data & Society (2018): 1-11.
Elish, Madeleine. “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction.” Engaging Science, Technology, and Society (2019).
boyd, danah and Kate Crawford. “Critical Questions for Big Data: Provocations for a Cultural, Technological, and Scholarly Phenomenon.” Information, Communication, & Society (2012): 662-679.
Data science’s popularity has grown in the last few years, and many have confused it with its older, more familiar relative: statistics. As someone who has worked both as a data scientist and as a statistician, I frequently encounter such confusion. This post seeks to clarify some of the key differences between them.
Before I get into their differences, though, let’s define them. Statistics as a discipline refers to the mathematical processes of collecting, organizing, analyzing, and communicating data. Within statistics, I generally define “traditional” statistics as the the statistical processes taught in introductory statistics courses like basic descriptive statistics, hypothesis testing, confidence intervals, and so on: generally what people outside of statistics, especially in the business world, think of when they hear the word “statistics.”
Data science in its most broad sense is the multi-disciplinary science of organizing, processing, and analyzing computational data to solve problems. Although they are similar, data science differs from both statistics and “traditional” statistics:
Difference
Statistics
Data Science
#1
Field of Mathematics
Interdisciplinary
#2
Sampled Data
Comprehensive Data
#3
Confirming Hypothesis
Exploratory Hypotheses
Difference
#1: Data Science Is More than a Field of Mathematics
Statistics is a field of mathematics; whereas, data science refers to more than just math. At its simplest, data science centers around the use of computational data to solve problems,[i] which means it includes the mathematics/statistics needed to break down the computational data but also the computer science and engineering thinking necessary to code those algorithms efficiently and effectively, and the business, policy, or other subject-specific “smarts” to develop strategic decision-making based on that analysis.
Thus, statistics forms a crucial component of data science, but data science includes more than just statistics. Statistics, as a field of mathematics, just includes the mathematical processes of analyzing and interpreting data; whereas, data science also includes the algorithmic problem-solving to do the analysis computationally and the art of utilizing that analysis to make decisions to meet the practical needs in the context. Statistics clearly forms a crucial part of the process of data science, but data science generally refers to the entire process of analyzing computational data. On a practical level, many data scientists do not come from a pure statistics background but from a computer science or engineering, leveraging their coding expertise to develop efficient algorithmic systems.
Difference
#2: Comprehensive vs Sample Data
In statistical studies, researchers are often unable to analyze the entire population, that is the whole group they are analyzing, so instead they create a smaller, more manageable sample of individuals that they hope represents the population as a whole. Data science projects, however, often involves analyzing big, summative data, encapsulating the entire population.
The tools of traditional statistics work well for scientific studies, where one must go out and collect data on the topic in question. Because this is generally very expensive and time-consuming, researchers can only collect data on a subset of the wider population most of the time.
Recent developments in computation, including the ability to gather, store, transfer, and process greater computational data, have expanded the type of quantitative research now possible, and data science has developed to address these new types of research. Instead of gathering a carefully chosen sample of the population based on a heavily scrutinized set of variables, many data science projects require finding meaningful insights from the myriads of data already collected about the entire population.
Difference
#3: Exploratory vs Confirming
Data scientists often seek to build models that do something with the data; whereas, statisticians through their analysis seek to learn something from the data. Data scientists thus often assess their machine learning models based on how effectively they perform a given task, like how well it optimizes a variable, determines the best course of action, correctly identifies features of an image, provides a good recommendation for the user, and so on. To do this, data scientists often compare the effectiveness or accuracy of the many models based on a chosen performance metric(s).
In traditional statistics, the questions often center around using data to understand the research topic based on the findings from a sample. Questions then center around what the sample can say about the wider population and how likely its results would represent or apply to that wider population.
In contrast, machine learning models generally do not seek to explain the research topic but to do something, which can lead to very different research strategy. Data scientists generally try to determine/produce the algorithm with the best performance (given whatever criteria they use to assess how a performance is “better”), testing many models in the process. Statisticians often employ a single model they think represents the context accurately and then draw conclusions based on it.
Thus, data science is often a form of exploratory analysis, experimenting with several models to determine the best one for a task, and statistics confirmatory analysis, seeking to confirm how reasonable it is to conclude a given hypothesis or hypotheses to be true for the wider population.
A lot of scientific research has been theory confirming: a scientist has a model or theory of the world; they design and conduct an experiment to assess this model; then use hypothesis testing to confirm or negate that model based on the results of the experiment. With changes in data availability and computing, the value of exploratory analysis, data mining, and using data to generate hypotheses has increased dramatically (Carmichael 126).
Data science as a discipline has been at the
forefront of utilizing increased computing abilities to conduct exploratory work.
Conclusion
A data scientist friend of mine once quipped to me that data science simply is applied computational statistics (c.f. this). There is some truth in this: the mathematics of data science work falls within statistics, since it involves collecting, analyzing, and communicating data, and, with its emphasis and utilization of computational data, would definitely be a part of computational statistics. The mathematics of data science is also very clearly applied: geared towards solving practical problems/needs. Hence, data science and statistics interrelate.
They differ, however, both in their formal definitions and practical understandings. Modern computation and big data technologies have had a major influence on data science. Within statistics, computational statistics also seeks to leverage these resources, but what has become “traditional” statistics does not (yet) incorporate these. I suspect in the next few years or decades, developments in modern computing, data science, and computational statistics will reshape what people consider “traditional” or “standard” statistics to be a bit closer to the data science of today.
For more details, see the following useful resources:
I recently integrated ethnography and data science to develop a Show Rate Predictor for an (anonymous) hospital system. Many readers have asked for real-world examples of this integration, and this project demonstrates how ethnography and data science can join to build machine learning-based software that makes sense to users and meets their needs.
Part 1: Scoping out the Project
A particular clinic in the hospital system was experiencing a large number of appointment no-shows, which produced wasted time, frustration, and confusion for both its patients and employees. I was asked to use data science and machine learning to better understand and improve their scheduling.
I started the project by conducting ethnographic research into the clinic to learn more about how scheduling occurs normally, what effect it was having on the clinic, and what driving problems employees saw. In particular, I observed and interviewed scheduling assistants to understand their day-to-day work and their perspectives on no-shows.
One major lesson I learned through all this was that when scheduling an appointment, schedulers are constantly trying to determine how many people to schedule on a given doctor’s shift to ensure the right number of people show up. For example, say 12-14 patients is a good number of patients for Dr. Rodriguez’s (made up name) Wednesday morning shift. When deciding whether to schedule an appointment for the given patient with Dr. Rodriguez on an upcoming Wednesday, the scheduling assistants try to determine, given the appointments currently scheduled then, whether they can expect 12-14 patients to show up. This was often an inexact science. They would often have to schedule 20-25 patients on a particular doctor’s shift to ensure their ideal window of 12-14 patients would actually come that day. This could create the potential for chaos, however, where too many patients arriving on some days and too few on others.
This question – how many appointments can we expect or predict to occur on a given doctor’s shift – became my driving question to answer with machine learning. After checking in with the various stakeholders at the clinic to make sure this was in fact an important and useful question to answer with machine learning, I started building.
Part 2: Building the Model
Now that I had a driving, answerable question, I decided to break it down into two sequential machine learning models:
The first model learned to predict the probability that a given appointment would occur, learning from the history of occurring or no-show appointments.
The second model, using the appointment probabilities from the first model, estimated how many appointments might occur for every doctors’ shift.
The first model combined three streams of data to assess the no-show probability: appointment data (such as how long ago it was scheduled, type of appointment, etc.); patient information, especially past appointment history; and doctor information. I performed extensive feature selection to determine the best subset of variables to use and tested several types of machine learning models before settling on gradient boosting.
The second model used the probabilities in the first model as input data to predict how many patients to expect to come on each doctors’ shift. I settled on a neural network for the model.
Part 3: Building an App
Next, I worked with the software engineers on my team to develop an app to employ these models in real time and communicate the information to schedulers as they scheduled appointments. My ethnographic research was invaluable for developing how to construct the app.
On the back end, the app calculated the probability that all future appointments would occur, updating with new calculations for newly scheduled or edited appointments. Once a week, it would incorporate that week’s new appointment data and shift attendance to each model’s training data and update those models accordingly.
Through my ethnographic research, I observed how schedulers approached scheduling appointments, including what software they used in the process and how they used each. I used that to determine the best ways to communicate that information, periodically showing my ideas to the schedulers to make sure my strategy would be helpful.
I constructed an interface to communicate the information that would complement the current software they used. In addition to displaying the number of patients expected to arrive, if the machine learning algorithm was predicting that a particular shift was underbooked, it would mark the shift in green on the calendar interface; yellow if the shift was projected to have the ideal number of patients, and red if already expected have too many patients. The color-coding allowed easy visualization of the information in the moment: when trying to find an appointment time for a patient, they could easily look for the green shifts or yellow if they had to, but steer clear of the red. When zooming in on a specific shift, each appointment would be color-coded (likely, unlikely, and in the middle) as well based on the probability that it would occur.
Conclusion
This is one example of a projects that integrates data science and ethnography to build a machine learning app. I used ethnography to construct the app’s parameters and framework. It tethered the app in the needs of the schedulers, ensuring that the machine learning modeling I developed was useful to those who would use it. Frequent check-ins before each step in their development also helped confirm that my proposed concept would in fact help meet their needs.
My data science and machine learning expertise helped guide me in the ethnographic process as well. Being an expert in how machine learning worked and what sorts of questions it could answer allowed me to easily synthesize the insights from my ethnographic inquiries into buildable machine learning models. I understood what machine learning was capable (and not capable) of doing, and I could intuitively develop strategic ways to employ machine learning to address issues they were having.
Hence, my dual role as an ethnography and data scientist benefitted the project greatly. My listening skills from ethnography enabled me to uncover the underlying questions/issues schedulers faced, and my data science expertise gave me the technical skills to develop a viable machine learning solution. Without listening patiently through extensive ethnography, I would not have understood the problem sufficiently, but without my data science expertise, I would have been unable to decipher which questions(s) or issue(s) machine learning could realistically address and how.
This exemplifies why a joint expertise in data science and ethnography is invaluable in developing machine learning software. Two different individuals or teams could complete each separately – an ethnographer(s) analyze the users’ needs and a data scientist(s) then determine whether machine learning modeling could help. But this seems unnecessarily disjointed, potentially producing misunderstanding, confusion, and chaos. By adding an additional layer of people, it can easily lead to either the ethnographer(s) uncovering needs way too broad or complex for a machine learning-based solution to help or the data scientist(s) trying to impose their machine learning “solution” to a problem the users do not have.
Developing expertise in both makes it much easier to simultaneously understand the problems or questions in a particular context and build a doable data science solution.
Aspiring data scientists will frequently ask me for recommendations about the best way to learn data science. Should they try a bootcamp or enroll in an online data science course, or any of the myriad options out there?
In the last several years, we have seen the development of many different types of educational programs that teach data science, ranging from free online tutorials to bootcamps to advanced degrees at universities, and the pandemic has seemed to have fostered the establishment of even more programs to meet the increased demand for remote learning. Although probably overall a good thing, having more options increases the complexity of deciding which one to do and the potential noise of programs upselling their services.
This article is a high-level survey of the four basic types of data science education programs to help you think about which might work best for you. Without already knowing data science, it can be difficult to assess how effective a program is at teaching it. Hopefully, this article will help break that chicken-and-the-egg conundrum.
These are the four basic ways to learn data science:
Do-it-yourself learning
Online courses
Bootcamps
Master’s degree or other university degree in data science (or related field)
I will discuss them in order from the cheapest to most expensive. I also included two hybrid strategies that combine a few of these that are worth considering as well. This table provides a quick, high-level synopsis of each one:
Option 1: Do-It-Yourself Online
There are tons of free, online data science resources that can either teach data science from scratch or explain just about any data science content you could possibly want to know. These range from tutorials for those who learn by doing like W3Schools, videos on YouTube and other sites for audio learners like Andrew Ng’s YouTube series, articles for visual learners who enjoy reading like Towards Data Science. You could scour the internet and teach yourself. It has the pros of being free and perfectly flexible to tailor to your schedule.
But as a former teacher, I have found independent learning is not for everyone. You must be entirely self-motivated and self-structured to teach yourself like this. So, know yourself: are you the type of person who could learn well completely independently like this?
Education programs tend to provide these resources that you might lack if you went it alone:
1) Curriculum Oversight: Data science experts in any education program generally establish some kind of data science curriculum for you that includes the necessary topics in the field. Many people who are new to data science do not know yet what data science concepts and skills are most important to learn about. This can create a chicken and egg problem for self-learners who must learn the field at least a little to know the most important items to learn in the first place. Data science programs help circumvent this by giving you an initial curriculum to started with.
2) Guidance of the Norms of the Field: In addition to the teaching the material, education programs implicitly introduce students to data science norms and ways of thinking. Even though there are times to deviate from the established custom, they are important when first working on teams with fellow data scientists. Sometimes self-learners learn the literal material but do not gather the implicit perspectives that enables their incorporation into the data science community.
3) External Social Accountability: Education programs provide a form of social accountability that subtly encourages you to get the work done. Self-learners must rely almost exclusively on their own self-motivation and self-accountability, which, in my experience, works for some people but not others.
4) Social Resources: Education programs (especially ones that meet either in-person or virtually) provide various people – teachers, students, and in some cases mentees/underlings – with whom one can talk through problems with, help you discover your weaknesses and shortcomings, and determine ways to address them. Minute programming details that are easily overlooked by beginners, but experts might easily spot can cause your entire program to fail. To learn independently, you will have to either solve all of these yourself or find data science friends or family who are willing to help you.
5) Certification of Skills: Education programs bestow degrees, grades, and other certifications as external proof that you do, in fact, possess the requisite skills in a data science role. Learning on your own, you must prove that you have these skills to employers by yourself. Developing a portfolio of thought-provoking projects, you have done is the best way to demonstrate this.
6) Guidance in Forming Projects: An impressive project works wonders for showcasing your data science skills. In my experience, beginners to data science often do not yet possess the skills to create, complete, and market a thought-provoking yet doable project, and one of the most important roles data science educators can have is helping students think through how to develop one. You must do this yourself when learning alone.
One can overcome each of these deficits. I have found that for people who learn well independently, its cost and flexibility advantages easily outweigh these cons. Thus, the crucial question is, Would this form of independent learning work for you? In my experience, it works for a comparatively small percent of people, but for those it works for, it is a great option.
If you do decide to teach yourself, I would recommend considering the following:
1) Be conscientious about your learning style when crafting your material. For example, if you are a visual learner, then reading online material resources would be best, but if you are more of an auditory learner, then I would recommend watching video tutorials/lectures on say YouTube.
2) If you have data science friends willing to help you, they can be a great asset, particularly in determining what data science materials to learn, troubleshooting any coding issues you might have, and/or developing a good project(s).
3) People in general learn data science best by doing data science. Avoid the common trap of only reading about data science without getting your hands dirty and experimenting yourself (preferably with unclean, annoying, real-world data, not already trimmed, “textbook perfect” data). Using pristine data to first learn the concepts is fine, but make sure you graduate yourself to practicing with real-life dirty data.
Option 2: Online Course
A variety of online courses exist. Most of them are relatively cheap (usually around $20-$50 a month or $100-$200 per course). For example, at the time of writing this, Udemy has an introductory data science course for a flat rate of $94.99, and Coursera a course for $19.99 a month (both with prices varying based on discounts and other special deals). Online courses are generally the cheapest of the courses you can enroll in, and because of the length of most, you will probably have to take several levels of courses (introductory to advanced) to learn the field.
Another advantage is that they are flexible: You can learn at your own pace, based on the needs of your schedule. This is really valuable for people who also working a job and studying on the side, with family commitments, and/or other obligations complicating their schedules. Keep in mind, though, that because you often pay per month, how many months you take often dictates the final cost. At the end of the day, spending an extra $100 or so to take a few more months to complete the course is still much cheaper than the other course options.
On the other hand, however, like doing it yourself, they tend to lack the social benefits of classroom learning: instructors to ask questions to and provide external social accountability, and fellow students to work alongside. In my experience, this makes it a very challenging for some learners, but others are not as comparatively affected by it.
In addition, many online courses provide more of a cursory summary of data science and lack the complex projects that are both necessary to learn data science and to market yourself to others. Even though there are exceptions, online courses are often good at introducing data science concepts rather than an in-depth exploration. Many focus on canned problems with already cleaned, ready-to-do data instead of letting you practice on the messy, complex, and often just plain silly data most data scientists actually have to use at their jobs. They also often lack the personnel for one-on-one coaching to mentor each student through portfolio-building projects with complex data.
Thus, online courses tend to provide good, cost-effective introductions to data science, helpful to see whether you like the field (see Hybrid #1 below), but do not generally provide the refined training necessary to become a data scientist. Now, some programs are evolving their courses. Especially as the pandemic increases demand for remote learning, online learning platforms are developing more robust online data science courses. If you choose to learn by taking online courses, I recommend supplementing it with your own projects to get experience practicing data science work and showcase in job interviews.
Hybrid #1: Use an Online course to Introduce Data Science (or Programming)
If you are completely new to data science, an online course can provide a low-cost, structured space to get a sense for what the field entails and determine whether it is a good fit for you. I have seen many people enroll in several thousand-dollar bootcamps or university degree programs only to learn there that they do not like doing data science work. An online course is a much cheaper space to discern that.
You could always explore data science yourself for free to decide whether you like it (see Option 1) instead of taking an online course, but I have found that many people who have never seen data science before do not know what to look up in the field to get started. An introductory online course is not that expensive, and the initial orientation into the major topic areas can be well worth the cost.
There are three basic versions of this approach:
1) If you do not already know a programming language, take an online programming course. I explained in this article why I would recommend Python as the language to learn (with Julia as a close second). If you do not like programming, then you have learned the lesson that you should not become a data scientist, and even if you do not end up in data science, programming is such a valuable skill that having some training in it will only help your occupational prospects in most other related fields.
2) If you do know a programming language, take an introductory data science course. These often provide a high-level overview of data science, especially helpful for people who need to work with data scientists and understand what they are talking about. If you need a math refresher, this is a great option as well.
3) I have seen prospective data scientists take online data analytics courses to prepare them for and determine their potential interest in data science. I would not recommend this, however. Even though data scientists will sometimes treat data analytics as a “diet” or “basic” version of data science, data analytics is different field requiring different skills. For example, data analytics courses typically do not include the rigorous programming. They generally focus on R and SQL if they teach programming at all, which are fine languages for data analytics and statistics but not enough for data science (for which you would want a language like Python). Data analytics and data science also generally emphasize different fields of math: data analytics tends to rely on statistics while data science on linear algebra, for example. Thus, what you would learn in those courses would not apply to data science as much as you would think. Now, if you are unsure of whether you would like to become a data scientist or data analyst, then a data analytics course might help you understand and get a feel for data analytics, but I would not use them to assess whether data science is a good fit for you.
Once you complete the online course, if you still think you would enjoy doing data science work, then you can choose any of the options to learn the field in more depth. This may seem like just getting you back to square one, but by taking an introductory programming or data science course, you have levelled yourself up so to speak and are more ready to face the “boss battle” of becoming a data scientist.
Option 3: Data Science Bootcamps
Data science bootcamps have also become popular. They tend to be several weeks long (in my experience often ranging from 2 to 6 months) intensive training programs. The traditional pre-pandemic bootcamp was in-person and would often cost around $10,000 to $15,000. Metis’s bootcamp is a good example of what they often look like.
Their biggest pros are that they offer the advantages of classroom education far more cheaply and in much less time than getting a university degree. They are a significant step-up cost-wise than the previous options (see Con 2 below), but they seek to provide a comparable (but less academically advanced and in-depth) scope of knowledge as a master’s degree in data science for a significantly lower price and in a fraction of the time. Even though it can often make their pace feel intense, the good bootcamps tend to mostly succeed at providing this. This makes them a great option for anyone who knows they want to become a data scientist. Finally, unlike the previous options, you get a teacher(s) to ask questions to and motivate you, and a set of fellow students to struggle through concepts with. The best programs offer the occupational coaching and build strong networks in data science communities to help their students find jobs afterwards.
They have some major cons, however:
1) They can feel fast-paced, unloading complex concepts in a short amount of time. Many of my friends who have done bootcamps have reported feeling cognitive whiplash. Expect those weeks/months to be mentally intense and to subsume your life. Data science bootcamps are often 9-5 full-time jobs during that time, and you will likely be too mentally exhausted to work on other things in the evenings or weekend (plus in some cases you will have homework to complete then anyways). A few weeks or months is not terribly long for such an ordeal, but it makes them much less flexible than the previous options. For example, this forces many students to take time from their current jobs to complete the bootcamp and to limit their social, familial, and other obligations as much as they can during their bootcamp. This makes it difficult for anyone unable to take time off work, with busy social or familial lives, or otherwise with a lot going on.
2) At several thousand dollars, they are clearly noticeably more expensive of the than the previous options (but still much cheaper than universities). Some offer scholarships and other services on a need-basis, but even then, the opportunity cost of having to put a job on hold can still be expensive. Given their general high salaries, landing a data science job would likely make the money back, but it takes a hefty initial investment.
This makes it an especially poor option for anyone thinking about data science but not sure whether they want to do it. $10,000 is a lot to spend to simply learn you do not like the field, and there are many cheaper ways to initially explore the field (see especially Hybrid #1). The cost still might be worth it, however, for anyone who really wants to become a data scientist but does not yet possess key skills and knowledge.
3) At the time of writing this, the Covid-19 pandemic has forced most data science bootcamps to meet remotely anyways, making their services far more similar to the much cheaper online courses. That said, many have sought to simulate the classroom environment virtually, trying to provide some type of social environment, but the classroom environment was a major advantage that made their significant increase in costs over the previous options worthwhile.
4) They tend to exist in large cities (especially tech centers). For example, bootcamps in the United States tend to concentrate in New York City, Los Angeles, Chicago, San Francisco, etc. Prior to the pandemic, anyone not living in those places would have to travel and temporarily reside in wherever their chosen bootcamp was, an additional expense.
5) They are often difficult for people who do not know programming and for those who do not know college-level mathematics like linear algebra, calculus, and statistics. If you do not know programming, I would recommend learning a programming language like Python (for more see this article I wrote explaining why to learn Python of all languages) through either a cheap online course and/or online tutorials first. Some data science bootcamps offer a preparatory introduction online course that teaches the prerequisite coding and math skills for those who do not understand it. They are worth consideration as well, but keep in mind the equivalent online course might be cheaper with roughly the same educational value.
If you decide to do a bootcamp, these criteria are important when researching which bootcamp to choose:
1) Project Orientation: How well do they enable you to practice data science through portfolio-building projects, and how impressive are the projects its alum did? The best data science bootcamps are generally teach in a project-oriented fashion.
2) Job-Finding Resources and/or Job Guarantee: What resources or coaching do they give to help you find a job afterwards? Help networking, presenting yourself, and interviewing, for example, are important skills to finding a job as a data scientist, and in addition to teaching you technical curriculum, the best programs tend to find occupational coaches to help specifically with the job-finding process. Also, some programs give a job guarantee: if you do not find a data science job after a certain number of months after graduating then they refund tuition. This generally shows they take job finding important enough to risk their own money on it (although do check at the fine print on the guarantee to see the exact terms they are agreeing to).
3) Alum Resources: A surprisingly import detail to consider is how much resources a bootcamp invests in cultivating alumni networks. I was surprised by how receptive to meeting/networking alum of the online bootcamp I did, and how satisfied alum tend to be with the bootcamp. The effort a bootcamp makes to work with and maintain relationships with its alum impact this significantly. Connectedness with alum can be difficult to assess when researching programs from afar, but asking whether you can speak with alum(s) to learn about their experiences with the program, checking a bootcamp’s alum activity on LinkedIn and other social media websites, and asking about what kind of networking opportunities with alum they facilitate can be great ways to assess how intentional a program is about cultivating relationships.
4) Scholarship Options: Some programs offer full or at least partial scholarships based on need. Clearly, ways to knock down the cost of the bootcamp would be great, especially if a bootcamp seems like an ideal option for you, but the cost seems too daunting.
Hybrid #2: Online Bootcamp
Online bootcamps tend to possess the schedule flexibility of online courses but offer more rigorous, personal (albeit remote) learning, allowing you to combine the best of aspects of data science bootcamps and online programs. They are also generally cheaper than traditional bootcamps (yet also more expensive than an online course). Finally, they tend to be a much better option for those who do not live in a major city that happens to have a local data science bootcamp program. The pandemic, if anything, has probably helped produce even more online bootcamp programs, since it has forced data science bootcamps to teach virtually.
I enrolled in Springboard’s online data science bootcamp in 2017, a great example of an online bootcamp. At the time, they cost roughly $1,000 a month (at the time of writing their standard rate is $1,490 a month and state their program generally takes six months). This is cheaper than traditional bootcamps but still a few totaling around $10,000 for six months. They had online curriculum typical of online courses but also provided weekly virtual meetings with an instructor to discuss the material and any issues you are having. Now they seem to include virtual lessons online. This individualized training and remote classroom environment are the main value adds over an online course, and you must assess whether, for you, they would be worth the additional cost. They are self-paced, providing much greater flexibility on when and how often you work than typical bootcamps. They also refunded your money if you did not find a job in six months after completion.
If you choose this option, be aware of the potential pitfalls of both online courses and traditional bootcamps. Just like with online programs, you will need to evaluate whether you are comfortable learning the curriculum by yourself (even you can meet with a mentor for major issues once a week, you would be doing the bulk learning by yourself throughout the week). Like with traditional bootcamps, expect the learning to be mentally intense and make sure they help you develop portfolio-building projects and provide job-finding resources and training.
Option 4: Master’s Degree or Other University Degree
The final option is to go back to school to get a degree in data science. This is the most expensive and time-consuming option: a master’s degree (a logical choice if you already have a bachelor’s) is generally the shortest, taking two years. But they cost upwards of $100,000. Even if partial or full scholarships decrease that cost, the opportunity cost of spending several years of your life in school is still higher than any of the other options. It can give a resume boost, however, if you know how to leverage it properly, which will likely increase your salary to make up for the initial cost. I would only recommend getting a master’s degree if you already know you love data science (say because you have already been working in the field, preferably if you also have already figured out the specific area of data science you want to do) but want to take your skills, technique, and/or theoretical knowledge of how the models work to the next level.
The best way to refine your data science skills is by doing data science: finding or creating contexts to push you as you practice data science. Graduate schools are not the only potential environment to refine one’s data science skills (e.g., all the previous options could involve that if done well), and even though graduate schools can be great at providing rigor, these other options can be a lot cheaper and more flexible. Finally, at the time of writing this, at least, the demand for data scientists exceeds the number of actual people in the field, and so getting a data science job without an “official” university degree in data science is pretty realistic.
University data science degree programs are relatively new – generally only a few years old. Thus, not all universities have literal data science degrees or departments but instead require that you enroll in a related program like computer science, statistics, or engineering to learn data science. This does not always mean these other programs are bad or unhelpful, but it often means you will have to perform extraneous or semi-extraneous tasks to data science proper in order to complete your degree (in some cases with minimal help from faculty from other fields).
When considering a program, you should make sure they are proactive about teaching professional and not just academic data science skillsets. These are the specific questions I would research to assess how well they might prepare you for non-academic data science jobs:
1) What proportion of their faculty currently work or at least have worked in the industry as a data scientist (or other similar job title)?
2) How well connected is the department with local organizations, and might they be able to leverage these relationships to help you work with these organizations through a work-study program or internship during the program and/or employment afterwards?
3) Will they help you build – or at least give you the flexibility to build – one’s thesis into an applied data science project that would boost your resume to future employers?
If your chosen program lacks these, I would strongly recommend building resume/portfolio-boosting projects and networking with local data scientists on the side while completing the program. This takes considerable time and energy, so ideally your department would actively help you in this work, instead of requiring that you do it on your own while also completing all their work.
Funding options is something else to consider. Are they willing to fund your degree fully or at least partly? Work-study programs where you work while getting your master’s can be a great way to graduate with no debt and gain resume-building work experiences (although they can make you busy). I benefitted greatly from working as a data scientist while completing my master’s, both because I graduated with no debt and because it allowed me to practice and refine my skills.
Finally, most universities require that you live nearby and attend physically (at least before and likely after the pandemic). Thus, you might have to find a place near you or be willing to relocate for a few years if there is not a data science degree program nearby. If so, you should factor moving expenses into the cost of doing the program.
Conclusion
Learning data science can be an awesome yet daunting prospect, and finding the right strategy for you is complicated, particularly given all the pedagogical, logistical, and financial considerations. Hopefully, this article has helped you think through how to journey forward.
In a
previous post about data visualization in data
science and statistics,
I discussed what I consider the single most important rule of graphing data. In
this post, I am following up to discuss the most important rules for making
data tables. I will focus on data tables in reporting/communicating findings to
others, as opposed to the many other uses of tables in data science say to
store, organize, and mine data.
To
summarize, graphs are like sentences, conveying one clear thought to the
viewer/reader. Tables, on the other hand, can function more like paragraphs,
conveying multiple sentences or thoughts to get an overall idea. Unlike graphs,
which often provide one thought, tables can be more exploratory, providing
information for the viewer/reader to analyze and draw his or her own
conclusions from.
Table Rule #1: Don’t be afraid to
provide as much or as little information as you need.
Paragraphs
can use multiple sentences to convey a series of thoughts/statements, and tables
are no different. One can convey multiple pieces of information that
viewers/readers can look through and analyze at his or her own leisure, using
the data to answer their own questions, so feel free to take up the space as
you need. Several page long tables are fair game and, in many cases, absolutely
necessary (although often end up in appendices for readers/viewers needing a
more in-depth take).
In my
previous data visualization post, I gave this bar chart as an
example of trying to say too many statements for a graph:
This is a
paragraphs-worth of information, and a table would represent it much better.[i]
In a table, the reader/viewer can explore the table values by country and year
themselves and answer whatever questions he or she might have. For example, if
someone wanted to analyze how a specific country changed overtime, he or she
could do so easily with a table, and/or if he or want to analyze compare the
immigration ratios between countries of a specific decade, that is possible as
well. In the graph above, each country’s subsegment starts in a different place
vertically for each decade column, making it hard to compare the sizes visually,
and since each decade has dozens of values, that the latter analysis is
visually difficult to decipher as well.
But, at the
same time, do not be afraid to convey a sentence- or graphs-worth of data into
a table, especially when such data is central for what you are saying. Sometimes
writers include one-sentence paragraphs when that single thought is crucial,
and likewise, a single statement table can have a similar effect. For example,
writing a table for a single variable does helps convey that that variable is
important:
Gender
Some Crucial Result
Male
36%
Female
84%
Now,
sometimes in these single statement instances, you might want to use a graph
instead of a table (or both), which I discuss in more detail in Rule #3.
Table Rule #2: Keep columns
consistent for easy scanning.
I have found
that when viewers/readers scan tables, they generally subconsciously assume
that all variables in a column are the same: same units and type of value.
Changing values of a column between rows can throw off your viewer/reader when
he or she looks at it. For example, consider this made-up study data:
Control Group (n = 100)
Experimental Group (n = 100)
Mean Age
45
44
Median Age
43
42
Male No. (%)
45 (45%)
36 (36%)
Female No. (%)
55 (55%)
64 (64%)
In this
table, the rows each mean different values and/or units. So, for example, going
down the control column, the first column is mean age measured in years. The
second column switches to median age, a different type of value than mean (although
the same unit of years). The final two rows convey the number and percentages
of males and females of each: both a different type of value and a different
unit (number and percent unlike years). This can be jarring for viewers/readers
who often expect columns to be of the same values and units and naturally
compare them as if they are similar types of values.
I would
recommend transposing it like this, so that the columns represent the similar
variables and the rows the two groups:
Mean Age
Median Age (IQR)
Male No. (%)
Female No. (%)
Control Group (n =
100)
45
43 (25, 65)
45 (45%)
55 (55%)
Experimental Group (n
= 100)
44
42 (27, 63)
36 (36%)
64 (64%)
Table Rule #3: Don’t be afraid to
also use a graph to convey magnitude, proportion, or scale
A table like
the gender table in Rule #1 conveys pertinent information numerically, but
numbers themselves do not visually show the difference between the values.
Gender
Some Crucial Result
Male
36%
Female
84%
Graphs excel at visually depicting the magnitude, proportion, and/or scale of data, so, if in this example, it is important to convey how much greater the “Some Crucial Result” is for females than males, then a basic bar graph allows the reader/viewer to see that the percent is more than double for the females than for the males.
Now, to convey this visual clarity, the graph loses the ability to precisely relate the exact numbers. For example, looking at only this graph, a reader/viewer might be unsure whether the males are at 36%, 37%, or 38%. People have developed many graphing strategies to deal with this (ranging from making the grid lines sharper, writing the exact numbers on top of, next to, or around the segment, among others), but combining the graph and table in instances where one both needs to convey the exact numbers and to convey a sense of their magnitude, proportion, or scale can also work well:
Finally, given
that tables can convey multiple statements, feel free to use several graphs to depict
the magnitude, proportion, or scale of one table. Do not try to overload a
multi-statement table into a single, incomprehensible graph. Break down each
statement you are trying to relate with that table and depict each separately
in a single graph.
Conclusion
If graphs are sentences, then tables can function more like paragraphs, conveying a large amount of information that make more than one thought or statement. This gives space for your reader/viewer to explore the data and interpret it on their own to answer whatever questions they have.
Loading…
Reload document 
Open in new tab 
