I Failed to Defend My First Research Topic. It Might Have Been the Most Important Lesson of My Data Science Journey.
There is something particularly uncomfortable about being asked to explain an idea you thought was brilliant, only to realize that you haven’t thought it through nearly as much as you believed.
That was me a few days ago. Week 4 and 5 of my data science journey didnt end without its own hassle, we were required to pick a topic and submit a research proposal. I knew this wouldnt be easy but i didnt expect this level of brainstorming and running back and forth. Here’s a quick glimpse into how it went.
I had an idea for a research project. It sounded relevant, addressed a problem I cared about, and seemed like the kind of thing that could benefit from machine learning. I was excited about it, so I submitted it.
Then my mentor asked me to explain it.
What exactly is the problem? What will the project achieve that existing research hasn’t already addressed? Where will the data come from? How will I incorporate machine learning? And how do I know the data will be good enough to support the work?
I went blank and struggled to give any convincing answer. Not because I hadn’t done any research, but because I had overlooked one of the most important parts of building a research project: establishing from the beginning that the available data could actually answer the question I was asking. I had done some preliminary research, but I hadn’t connected all the dots.
My proposed topic wasn’t approved, and I had to go back to the drawing board.
It was disappointing, but looking back, that experience may have taught me more about research than simply having my original idea approved would have. To understand why this particular project matters so much to me, I need to explain how I got here in the first place.
My Career Path Has Been Anything but Straightforward
I studied Educational Technology at university. It is one of those courses you study and your lecturers give you the impression that you could work almost anywhere. At the time, that sounded like an advantage. Looking at my career so far, I would say it is true to a considerable extent, not in the way you would expect though.
I’ve worked across different fields, but not in the educational technology field itself.
Sometimes, I wonder what my career might have looked like under different circumstances. Perhaps I would have worked as an educational technologist in a large organisation, helping to design training programmes, onboard new employees, and make complex information easier to understand through instructional design and technology.
That would have been a perfectly reasonable application of what I studied.
Instead, my professional experience has taken me elsewhere, and I’ve had to figure things out along the way.
I’m not suggesting that studying one course should automatically determine the rest of your career. People change interests, discover new opportunities, and sometimes deliberately choose a different direction. There is nothing inherently wrong with that.
What bothers me is this: How do you make informed career decisions when you don’t know where your degree can take you, what the labour market expects, or what additional skills you need to compete?
A degree tells you what you studied, but not necessarily which industries need your skills, what else employers expect, or how those expectations are changing. Some graduates figure this out early; others learn through internships, job applications, rejection emails, and years of trial and error.
I keep wondering what it would mean if students had access to this information before graduating. Not a system that dictates what they should become, but one that helps them explore career possibilities, understand industry requirements, and make more informed decisions.
That question is where my interest in this research began.
The Research Idea I Couldn’t Defend
My first proposed topic was centred on education–employment mismatch: why do graduates study one course and end up working in a different field?
It felt personal, relevant, and worth investigating. More importantly, it connected directly to a problem I wanted to understand.
I imagined a project that could examine the relationship between what people study and what they eventually do for a living. With enough good data, perhaps it could reveal patterns in the kinds of graduates who enter different industries, the skills they need beyond their qualifications, and the factors associated with working outside their field of study.
The bigger idea was even more ambitious: could machine learning eventually help graduates anticipate the career paths available to them and identify the additional skills they should acquire before leaving university?
I liked the idea so much that I initially underestimated the work required to turn it into a defensible research project.
When my mentor asked me to walk her through the proposal, I realised I had a broad problem I cared about, but I hadn’t sufficiently defined the research problem I could realistically investigate.
I couldn’t clearly explain what my project would contribute beyond what other researchers had already done. I hadn’t established a credible data source. I hadn’t demonstrated how machine learning would fit the problem rather than simply make the title sound more technical.
My interest was genuine. My research design was not yet strong enough.
That distinction matters.
It is possible to identify a real-world problem without having a workable research question. It is also possible to have an interesting research question that cannot be answered with the data available to you.
I had focused heavily on the first part and not enough on the second.
A Research Topic Is More Than an Impressive Title
One of the biggest lessons I learned is that a research topic shouldn’t be judged by how impressive it sounds. Adding “machine learning” or “artificial intelligence” to a title doesn’t automatically make a project innovative. A good research project needs a clearly defined problem, a defensible contribution, a suitable method, and data capable of answering the research question.
Take education–employment mismatch, for example. Before investigating it, I would need to define what mismatch means, identify what existing studies have covered, and find data containing the relevant information about graduates’ qualifications, jobs, and skills. If I wanted to build a predictive model, I would also need to establish exactly what I intended to predict and which information would be available at the time of prediction.
No algorithm can compensate for data that doesn’t contain the information needed to answer the question. I had to step back and ask myself what I could realistically investigate with the data, time, and skills available to me.
Finding a Problem I Could Actually Investigate
I didn’t abandon the bigger idea; I narrowed my immediate focus. I wanted a topic connected to graduate employability in Nigeria, with accessible data, measurable outcomes, and a realistic machine-learning approach.
My search led me to a longitudinal dataset published by Olofinyehun and colleagues in 2023: Pooled Longitudinal Dataset on the Assessment of an Apprenticeship-based Entrepreneurship Intervention in Nigeria.
The dataset follows participants across multiple survey waves, from NYSC orientation to an endline survey conducted approximately six months after they exited the programme.
What interested me was the ability to connect information recorded earlier with outcomes observed later. Since my goal was to predict post-NYSC employment outcomes, I could use baseline characteristics as predictors rather than information collected after those outcomes occurred. This would help prevent data leakage and make the prediction task more meaningful.
Finding the dataset was a promising start, but I still needed to understand its variables, assess its limitations, and determine whether it could genuinely support my research question.
One of the less glamorous parts of this process was understanding the dataset. It contained several survey waves, so I had to identify the baseline and endline records and match respondents across both. I started with 9,304 unique baseline respondents and 1,300 endline respondents, but only 1,271 appeared in both waves.
That raised an important concern: attrition bias. Respondents who completed the endline survey may differ from those who didn’t, potentially limiting how far I can generalise my findings. I would need to acknowledge this rather than assume my sample represents all Nigerian graduates.
Next, I had to decide what the model would actually predict. The endline data recorded whether respondents had a wage job and whether they operated a business. Some reported both, while others reported neither, so I couldn’t simply divide everyone into employed, self-employed, and unemployed.
Instead, I defined three mutually exclusive categories:
- Wage-employed: respondents with a wage job, including those who also owned a business.
- Business-owning without wage employment: respondents who owned a business but had no wage job.
- Neither wage -employed or owning business: respondents who reported neither a wage job nor a business.
I also avoided labelling the third group unemployed. That classification would require additional information, such as whether respondents were available for work and actively seeking it, which these variables don’t establish.
Get Mustaphasodiq’s stories in your inbox
Join Medium for free to get updates from this writer.
I could have used the word unemployed because it sounds simpler and more familiar. But doing so would have claimed more than my variables could support. That is a small example of a much bigger lesson I am beginning to appreciate: the way you define a problem determines what your analysis can legitimately claim to have discovered.
The data does not always give you the answer in the exact form you want. Sometimes, you have to make a defensible decision about how to represent what it actually measures. And you have to be honest about the limitations of that decision.
From a Personal Question to a Researchable Topic
After examining the dataset and refining the outcome categories, I settled on the following working title:
Predicting Post-NYSC Employment Outcomes among Nigerian Graduates Using Machine Learning
The objective is to investigate whether characteristics recorded during NYSC orientation can help predict which of the three outcome categories respondents report approximately six months after completing national service.
This is narrower than my original idea, but that is precisely why it is more workable.
I am no longer trying to explain the entire relationship between education and employment in Nigeria. I am asking a specific predictive question using an existing longitudinal dataset.
I plan to use baseline characteristics to predict the three post-NYSC outcome categories, comparing multinomial logistic regression and Random Forest against a simple baseline classifier. Rather than assume one model will perform better, I will evaluate their performance using metrics such as macro-F1, balanced accuracy, precision, and recall to assess how well they identify each category, especially given the differences in class sizes.
I don't even know how to build the models yet, and their performance remains to be seen. However, I now have a clear research question, a dataset to work with, and a methodology for investigating it. That is a significant step from where I started.
The Research Gap Is Not Something You Simply Announce
Another thing I learned while preparing the proposal is that identifying a research gap requires more than searching for papers with similar titles and declaring that nobody has done your project before.
I found related work on entrepreneurship among young Nigerian graduates, research examining the impact of apprenticeship-based entrepreneurship interventions, and research applying machine learning to employment-related outcomes among Nigerian corps members.
That literature is useful because it shows that my broad area of interest is not unexplored.
It also means I need to be precise about my own contribution.
My proposed study focuses on predicting three observed post-NYSC outcome categories from characteristics recorded at baseline, using the matched respondents in this particular longitudinal dataset. Its contribution will depend on how rigorously I develop the models, compare their performance, and interpret the findings.
I cannot honestly claim that machine learning has never been applied to this area, or that nobody has studied the dataset before. Related studies exist, and my proposal needs to acknowledge them.
The more defensible question is whether this particular predictive task, using these defined outcomes and baseline predictors, can provide useful evidence beyond what the existing studies have established.
That is a question I can investigate.
I am also learning that prediction is not the same as explanation, and neither is the same as causation. Even if a characteristic helps a model predict an employment outcome, that does not mean the characteristic caused the outcome. The model may identify an association that is useful for prediction without telling us why it exists.
Keeping those distinctions clear will matter throughout the project.
The Bigger Idea I Haven’t Let Go Of
Although I am satisfied with the direction of this proposal, it is not the full extent of what I hope to explore.
My long-term ambition is still connected to the problem that inspired the original idea: the gap between what graduates study and the work they eventually do.
I would like to investigate whether it is possible to build a learning system that helps students understand the career paths associated with their chosen courses, the skills commonly required for particular roles, and the additional competencies they might need to develop before graduation.
Imagine a student who is two years into a degree programme. Rather than waiting until graduation to discover how competitive the labour market is, the student could explore evidence about relevant occupations, the skills employers request, and the range of career options available to people with similar educational backgrounds.
Such a system would not tell the student that a degree guarantees a particular job. It would not predict the future with certainty, and it should not limit someone to the career paths followed by previous graduates.
Instead, it could help make career planning more informed by bringing together reliable labour-market information, educational backgrounds, occupational requirements, and skills data.
There are substantial challenges to solving that problem properly. It would require suitable data linking qualifications to actual occupations and skills, careful definitions of mismatch, and a way to account for changing labour-market demand. A credible system would also need to communicate uncertainty rather than present its predictions as guarantees.
My current dataset does not contain everything needed to investigate that larger question. It cannot, on its own, establish whether someone’s degree matches their occupation or identify every additional skill that person needs to become competitive.
I am not going to pretend otherwise.
But I do not think the fact that I cannot answer the bigger question yet means I should abandon it.
A research project can be a starting point rather than a complete solution. This one will help me develop the skills to work with real data, formulate a predictive problem, evaluate machine-learning models, and understand the limitations of what they produce.
Perhaps, with more experience and access to better-suited data, I can return to the original question and investigate it more directly.
For now, I have found a first step I can defend.
What This Experience Has Taught Me About Starting
I began this process thinking I needed to find the right research topic.
I now think I needed to learn how to recognize a researchable problem.
Those are not quite the same thing.
My first idea mattered to me, but personal relevance alone could not make it a sound research proposal. I needed to establish what was already known, find a gap in the research and design a method that could answer the question without making claims the evidence could not support.
I also had to accept the distance between what I wanted to build and what I currently know how to build.
That distance is intimidating, especially when you are learning data science without an extensive technical background. I understand the demands of academic research well enough to recognize how much I still need to learn about machine learning. Sometimes, knowing the depth of a task makes starting it more frightening, not less.
But waiting until I understand everything before beginning would mean never beginning at all.
I don’t know how well the models will perform. I don’t know whether the results will be as useful as I hope. And I certainly don’t know whether this project will eventually lead to the larger career-guidance system I have imagined.
Those are things the work itself will have to reveal.
What I do know is that I am no longer choosing a topic simply because it sounds impressive or because the problem matters to me personally. I am learning to connect the problem I care about with evidence I can use and a question I can realistically investigate.
And that, for me, is progress.
The proposal is taking shape. The next stage is to start the main research.
There is a long way to go.
This is not the research question I started with. It is the first one I have learned how to defend.
And perhaps that is a better place to begin.
