Starting with a tool
A plan to use a neural network, chatbot or a particular software package is not yet a research question. Begin with the problem and evidence, then decide whether the tool is justified.
UG and MSc research guide
A useful topic starts with a focused problem and an answerable question. This guide helps you check scope, evidence, method, evaluation and feasibility before you commit to a computing project.
Book a free 20-minute consultationLook for a computing or data problem that interests you: a task that is difficult to perform, a decision that could be better informed, a system that needs evaluation, or an area where published findings are incomplete. A module, placement, research paper or public dataset can prompt an idea, but it does not by itself establish a research gap.
Read recent academic work and note what has been studied, how evidence was gathered, and what limitations or open questions the authors identify. Use those sources to explain why your proposed question is worth investigating. Do not assume a topic is novel just because you have not seen it before.
Projects may involve machine learning, deep learning, data analysis, software evaluation or a systematic review. Your course may have different expectations for implementation and contribution, so check its handbook before selecting a project type.
Explore the Computer Science and Data Science subject guide for examples of dissertation designs.
A broad area needs boundaries before it can become a project. Specify the task or problem, the context or population, the evidence you will use, and the outcome you intend to examine. A clear scope helps you estimate implementation, analysis and writing effort.
For example, “deep learning for medical images” is too broad. A more bounded direction could compare a small number of image-classification approaches on a specified, permitted dataset, using a fixed validation plan and metrics chosen for the error types that matter. It is an illustration, not a ready-made topic: data access, ethics, course rules and supervisor advice still need checking.
A research question describes what you will investigate, not simply what you will build. It should identify the task and scope clearly enough to guide your reading, method and evaluation. Your aim states the overall purpose; objectives describe the steps that lead to an answer.
A computing project may contribute through a careful comparison of methods, an application in a justified context, an evaluation of a system, an analysis of a dataset or a synthesis of existing research. A new algorithm is not required for every undergraduate or master's dissertation. Confirm the expected contribution with your supervisor.
Check that each objective can be addressed with the evidence you can realistically collect or access. Avoid promising broad impact, causal conclusions or generalisation that your design cannot support.
Before finalising the topic, identify the likely data source and inspect a sample if permitted. Check who collected it, what each record represents, how labels or measurements were produced, what is missing, whether classes or groups are imbalanced, and what licence or terms apply. Consider whether the data is representative of the question you want to answer.
Look for risks such as duplicated records, target leakage, a mismatch between training and test populations, outdated measurements or labels that do not measure the concept you intend to study. A public dataset can be useful, but availability alone does not make it suitable. Follow institutional rules for personal, sensitive and restricted data, including approval and data-handling requirements.
Estimate the complete workload: literature review, data permissions, ethics review if needed, preparation, implementation, validation, analysis, writing and revision. Confirm access to software and compute. A modest prototype can expose technical or data-access problems before they threaten the schedule. Keep a feasible fallback that still answers the research question.
For more on selecting a technical approach, see the guides to machine learning and deep learning.
Decide what evidence is needed and how it will answer the question. A model-comparison study needs consistent preprocessing and validation. A software evaluation needs clear requirements or use cases and a suitable test or user-evaluation plan. A data-analysis study needs measures and an analysis strategy that fit the data. A systematic review needs a focused review question, documented search strategy and transparent selection criteria.
Explain why the design fits the question, how cases or data will be selected, what steps you will take, and what limitations follow. If you combine methods, check that each part has a purpose and that the extra workload is realistic. Avoid choosing a method simply because you already know a particular tool.
Related guidance: research methodology, systematic literature review, and quantitative analysis.
Evaluation should test the claims your research question makes. Define the comparison, metrics and validation design before running experiments. Use a simple, credible baseline and make comparisons under consistent data, preprocessing and tuning conditions.
For classification, accuracy may be misleading when classes are imbalanced; precision, recall, F1, confusion matrices or precision-recall analysis may be more informative depending on the errors that matter. For regression, consider whether MAE or RMSE aligns with the task. For a software artefact, identify test cases, benchmarks or user tasks that reflect its requirements.
Keep a final test set separate from model selection, avoid data leakage, and report variation or uncertainty where appropriate. Inspect failure cases, record settings and seeds, and explain what the results do and do not show. A strong evaluation includes limitations, not only the best score.
See the detailed machine-learning evaluation guide and deep-learning guide for more on validation and experiment records.
A plan to use a neural network, chatbot or a particular software package is not yet a research question. Begin with the problem and evidence, then decide whether the tool is justified.
Topics such as artificial intelligence in healthcare can contain many separate projects. Bound the task, data, population or setting, and intended outcome.
A large dataset may still have unsuitable labels, missing fields, leakage, sampling bias or unclear provenance. Inspect it and confirm the research question it can realistically support.
Comparisons need consistent data splits, preprocessing and tuning rules. Keep a final test set separate from model selection and include a credible baseline.
A single accuracy or headline metric can hide important errors. Select measures that fit the task, inspect failure cases and explain practical and methodological limits.
Data cleaning, permissions, compute setup, debugging, validation and writing all take time. Make a small feasibility prototype before committing to a complex design.
Use this list before you take a proposal to your supervisor. Any “not yet” answer is a prompt to clarify, narrow or verify part of the plan.
Prepare a short outline with the problem, research question, aim and objectives, relevant literature, evidence source, method, evaluation criteria, timeline and known risks. Check it against your handbook and discuss it with your supervisor before committing to implementation.
For an individual discussion, book a free consultation or contact Matt Scholar on WhatsApp.