KleioKleio

Thesis & research · Ideas

AI thesis ideas, by course.

Concrete, defensible machine learning project shapes for business, psychology, education, engineering, marketing and agriculture students — with the data each one actually needs.

Kleio Software13 min read

Most lists of AI thesis topics are useless because they stop at the title. “Sentiment analysis of customer reviews” is not a research project; it is a phrase. What makes a topic workable is whether you can get the data, whether the question is specific enough to answer, and whether you can defend the method.

So every idea below comes with the three things that actually decide feasibility: what you are predicting, what data you need, and what a panel will push on. Read the data requirement before you fall in love with the topic.

A note that applies to all of them: the best thesis topic is usually one where you already have access to data nobody else has. A modest question on data you can actually get beats an ambitious question on data you cannot.

Business, management and accountancy

Business research has an advantage most other fields do not: the data already exists as a by-product of operating. Sales records, inventory movements, customer transactions and payment histories are all recorded because the business needs them, which means you are not starting from scratch.

The constraint is access. A cooperating business has to be willing to share real records, usually under a confidentiality arrangement. Secure that before you commit to the topic.

Customer churn prediction
Predict which customers stop buying, from transaction history and demographics. Needs at least a year of customer-level records with a definable lapse point. Panels will ask how you defined churn — have a defensible answer.
Demand forecasting for inventory
Forecast product demand from historical sales, seasonality and promotions. Needs two or more years of dated sales records. Must be split chronologically, never randomly.
Credit or payment risk scoring
Predict late payment from customer and transaction attributes, useful for cooperatives and lending institutions. Needs records with known outcomes. Expect ethics questions about fairness across borrower groups, and prepare for them.
Employee attrition modeling
Predict resignations from HR records. Needs several years of employee data including leavers. Anonymization is mandatory and the ethics clearance is real work.

Psychology and social sciences

Psychology research usually generates its own data through instruments and surveys, which means you control the structure — a genuine advantage if you design the collection with the analysis in mind.

The constraint is sample size. Machine learning on 80 survey responses is not defensible, and a panel will say so. If your sample is small, using AI as an instrument rather than as the contribution is usually the stronger project.

Text classification of open-ended responses
Automatically code free-text survey answers into themes, validated against human coding. Needs a few hundred responses with a portion manually coded as ground truth. Inter-rater reliability with the human coding is the number your panel will want.
Sentiment analysis on Filipino and Taglish text
Genuinely novel territory, because most off-the-shelf sentiment tools handle code-switching badly. That weakness is itself a defensible research contribution.
Screening-instrument scoring support
Model the relationship between instrument responses and an assessed outcome. Requires ethics clearance and careful framing: this supports screening, it does not diagnose, and your write-up must say so.
Predicting engagement or dropout
Identify who disengages from a program, using participation records. Often more feasible than clinical topics because the data already exists administratively.

Education

Education research sits on top of institutional records that are already collected and reasonably structured: grades, attendance, enrolment, assessment scores. If your institution will grant access, the data problem is largely solved.

The ethics work is the real overhead. Student data is sensitive, anonymization is non-negotiable, and clearance takes time. Start that process early.

Early identification of at-risk students
Predict failure or dropout risk from early-semester grades and attendance. Needs several cohorts with known outcomes. The interesting defense question is what an institution should ethically do with the prediction.
Automated short-answer scoring
Model agreement between machine and human scoring of short written answers. Needs a corpus of answers with teacher-assigned scores. Report agreement with human raters, not accuracy.
Predicting licensure-exam performance
Predict board-exam outcomes from academic records. Highly relevant to Philippine institutions and usually well received, because the practical use is obvious.
Curriculum pathway analysis
Identify which course sequences predict better outcomes, using historical enrolment and grade data. Descriptive and predictive at once.

Engineering and architecture

Engineering is the field where machine learning most often maps cleanly onto an existing decision, which is why it produces some of the most defensible theses. Both of the research projects in our own portfolio are engineering ones.

The constraint is that datasets tend to be small, because each record often represents a physical test or a real project. That pushes you toward simpler models, and knowing that in advance is an advantage.

Material or component selection
Predict an appropriate material or component from site and design parameters. This is exactly what our geotextile and steel-connection projects do, and the shape generalizes widely.
Structural or material property prediction
Predict strength, durability or performance from mix design and test conditions. Concrete compressive strength is the classic, and there is published data to benchmark against.
Defect detection from images
Detect cracks, corrosion or surface defects from photographs. Feasible but the labeling effort is the whole project — budget for it honestly.
Energy or water demand forecasting
Forecast consumption from historical usage and weather. Good data availability, and time-aware validation is the methodological point to get right.

Marketing and communication

Marketing research benefits from public data in a way most fields do not. Social platforms, review sites and e-commerce listings are all sources you can collect from without institutional access.

The constraint is that platform terms of service and data-privacy rules govern what you may collect and store. Check them, and document that you checked — a panel may ask.

Review sentiment and theme extraction
Classify customer reviews by sentiment and topic to identify recurring complaints. Straightforward to collect, and doing it on Filipino-language reviews makes it genuinely novel.
Campaign response prediction
Predict which customer segments respond to which campaign type. Needs historical campaign data from a cooperating business.
Content-performance modeling
Predict engagement from post attributes such as timing, format and length. Data is accessible through platform analytics for an account you have access to.
Customer segmentation
Clustering rather than prediction. A legitimate approach, but be ready to justify how many segments you chose and why — that is the whole defense.

Agriculture and environmental science

Agricultural research in the Philippines has strong practical relevance and reasonable data availability through government agencies, research institutes and cooperatives.

The constraint is that field data is seasonal. If your model needs multiple growing seasons and you have one, that shapes what you can claim.

Yield prediction
Predict crop yield from weather, soil and input data. Public agency data exists, which is a real advantage over primary collection.
Pest or disease identification from images
Classify plant diseases from leaf photographs. There are public datasets to build on, though the honest limitation is how well they transfer to local conditions and local varieties.
Water-level or flood forecasting
Forecast levels from rainfall and upstream measurements. High practical value, and time-series methodology is the thing to get right.
Aquaculture and livestock monitoring
Predict growth or health outcomes from environmental sensor readings. Needs a cooperating farm with instrumentation, which is the binding constraint.

How to choose between them

Rank your candidate topics on three things in this order: can I get the data, is the question specific, and can I explain the method. Data access is first because it is the one you cannot solve with effort later.

Then narrow. “Predicting student performance” is not a thesis. “Predicting first-year attrition in a BS Accountancy program from first-semester grades and attendance, across five cohorts” is a thesis. The specificity is what makes it defensible.

And be realistic about scope against your defense date. A tightly scoped project finished properly beats an ambitious one finished badly, every time.

Have a topic but not sure it is feasible?

The fastest way to find out is to talk it through with someone who has built these before. Kleio develops machine learning models for thesis and capstone projects across the Philippines, and will tell you early if the data will not support the question.

AI for thesis & capstone projects

Related questions

The follow-ups people ask after reading this.

Which course has the easiest AI thesis topics?

Ease depends far more on data access than on course. Business and education students often have the smoothest path because the records they need already exist administratively. Engineering topics map most cleanly onto a real decision but tend to have small datasets. The easiest project for you is the one where you already have access to relevant data.

Do I need to know how to code to do an AI thesis?

You need to understand what the model does, why it was chosen, and where it fails, because that is what you will be asked to defend. You do not necessarily need to implement it yourself, and many students in non-IT programs work with a technical collaborator. Check what your institution requires you to disclose, and disclose it.

Can I use a public dataset instead of collecting my own?

Yes, and it is often the sensible choice, particularly in agriculture and engineering. The trade-off is novelty: a public dataset many people have used needs a genuinely new question or a local angle to be interesting. Applying it to Philippine conditions and being explicit about transferability is frequently that angle.

How specific does my research question need to be?

Specific enough that the inputs, the outcome and the population are all named. Compare 'predicting student performance' with 'predicting first-year attrition in a BS Accountancy program from first-semester grades and attendance across five cohorts'. The second tells a panel exactly what you did and exactly what you are claiming.

Keep reading

How to incorporate AI into a thesis

The decision order that protects you: question first, data second, method third.

Choosing a machine learning model

Why the most sophisticated architecture is usually the wrong answer.

Steel Connections Classifier

A worked example: an engineering thesis model, built and deployed.

Still have a question this didn't answer?

Bring it to a free discovery call. No pitch, straight to the developer, and an honest answer even when it is that we're not the right fit.

Book a free discovery call