Research Problems, Questions, and Proposals

Introduction and Context

Research begins not with a method or a dataset, but with a well-defined problem. In artificial intelligence and machine learning (AI/ML), researchers frequently encounter broad concerns such as unreliable medical prediction, biased facial-recognition systems, inefficient recommendation engines, or models that perform poorly when deployed outside laboratory conditions. These concerns are important, but they are not yet research problems. A research problem is a specific, evidence-based issue that can be investigated systematically within available time, resources, and ethical limits.

A strong research problem connects three elements: an unsatisfactory condition, a defensible reason why it matters, and a feasible way to generate knowledge about it. The problem is transformed into research questions, objectives, variables, hypotheses where appropriate, and a proposal that explains how the investigation will be conducted.

For AI/ML students, this process is especially important because technical novelty alone does not guarantee research value. A new model may be unnecessary if an existing method solves the problem adequately. Conversely, a modest improvement in interpretability, fairness, energy efficiency, or robustness may have substantial practical importance.

Learning Objectives

After completing this chapter, students should be able to:

1. Distinguish a research topic, problem, gap, purpose, question, and objective.

2. Identify and evaluate research gaps using scholarly and practical evidence.

3. Define scope, constructs, variables, assumptions, constraints, and delimitations.

4. Formulate suitable research questions and testable hypotheses.

5. Explain how research paradigms and designs influence research proposals.

6. Prepare a feasible proposal containing methods, resources, risks, timeline, and evaluation criteria.

7. Recognize ethical, legal, governance, and reproducibility requirements in AI/ML research.

From Topic to Researchable Problem

A topic is a broad area of interest, such as deep learning for agriculture. A problem identifies a deficiency within that area: for example, crop-disease classifiers trained on controlled laboratory images may fail on photographs taken by farmers under varying illumination. A purpose statement expresses what the study intends to accomplish, such as evaluating whether domain adaptation improves field performance. A research question states what the study will answer, while an objective states an action undertaken to answer it.

A useful progression is:

Topic → Context → Problem → Gap → Purpose → Questions → Objectives → Method → Contribution

For example:

• Topic: Machine learning for crop-disease detection.
• Context: Small farms use mobile phones to photograph leaves.
• Problem: Models trained on laboratory images may generalize poorly to field images.
• Gap: Existing studies report high benchmark accuracy but provide limited evidence about cross-environment generalization.
• Purpose: To compare adaptation methods for improving field-domain performance.
• Question: Which adaptation method most improves performance when training and deployment image conditions differ?
• Objective: Compare baseline fine-tuning, data augmentation, and domain-adaptation methods using a predefined field-image test set.

A problem statement should normally identify the context, affected stakeholders, evidence of the problem, consequences, and the knowledge needed to address it. Statements such as “AI is changing agriculture” are informative but too broad. A stronger statement is: “Although laboratory-trained classifiers report high accuracy, their reliability under field conditions remains insufficiently characterized, creating uncertainty for mobile decision-support systems.”

Identifying Research Gaps

A research gap is a meaningful limitation, inconsistency, absence, or unanswered question in existing knowledge or practice. A gap is not simply a topic that has not been studied by the student. It must be demonstrated through a critical examination of credible evidence.

Common types of gaps include:

• Knowledge gap: An important relationship or phenomenon is insufficiently understood.
• Methodological gap: Existing studies use weak, narrow, or unsuitable methods.
• Contextual gap: Findings have not been tested in a different population, environment, language, industry, or geographic setting.
• Population gap: An underrepresented group has been excluded.
• Evidence gap: Claims are based on small samples, short periods, or inadequate comparisons.
• Implementation gap: A technically successful method has not been evaluated in realistic use.
• Contradiction gap: Credible studies report conflicting results.
• Evaluation gap: Studies use incomplete metrics, such as accuracy without calibration, fairness, latency, or resource consumption.

Gap identification requires more than listing papers. Students should compare studies according to research questions, data, methods, baselines, evaluation measures, limitations, and assumptions. A literature matrix is useful:

| Study | Data/context | Method | Main finding | Limitation | Relevance to proposed study |

|---|---|---|---|---|---|

| A | Public images | CNN | High accuracy | Controlled conditions | Establishes baseline |

| B | Field images | Transfer learning | Moderate improvement | Small sample | Supports contextual gap |

| C | Field images | Domain adaptation | Mixed results | No external test set | Supports evaluation gap |

A credible gap statement should answer three questions: What is already known? What remains uncertain or inadequate? Why does resolving the uncertainty matter? The statement should avoid exaggerated claims such as “no one has studied this.” It is safer to write “Few studies have evaluated…” or “Existing evidence remains limited because…” unless a systematic search supports a stronger assertion.

Research Paradigms and Design Choices

A research paradigm comprises assumptions about reality, knowledge, evidence, and appropriate methods. At undergraduate level, three broad orientations are especially useful.

Positivist or post-positivist research emphasizes measurement, observable evidence, hypothesis testing, and control of alternative explanations. Experiments, benchmark comparisons, and statistical analyses commonly reflect this orientation. Absolute objectivity is difficult, so post-positivist work recognizes uncertainty, measurement error, and researcher decisions.

Interpretivist research examines how people understand and experience phenomena. Interviews, observations, and qualitative analysis may be appropriate when the question concerns trust in AI, user interpretations of explanations, or organizational adoption.

Pragmatic research selects methods according to the problem rather than committing exclusively to one philosophical tradition. A mixed-methods study might measure model performance and interview users about whether the model is usable and trusted.

Design should follow the question. Descriptive designs characterize a phenomenon; correlational designs examine associations; experimental designs investigate effects under controlled conditions; case studies provide contextual depth; survey designs collect standardized responses; and mixed-methods designs integrate quantitative and qualitative evidence. A proposal should explain why its design is suitable rather than selecting a fashionable method.

Formulating Research Questions

A good research question is clear, focused, researchable, significant, and ethically answerable. It identifies the population, system, intervention, comparison, phenomenon, or context when these are relevant.

Useful forms include:

• Descriptive: What types of errors occur when a language model summarizes legal documents?
• Comparative: How does method A compare with method B under the same evaluation protocol?
• Relational: What relationship exists between training-data diversity and model performance across demographic groups?
• Causal: Does calibrated uncertainty reporting improve users’ decisions compared with uncalibrated predictions?
• Exploratory: How do developers perceive the usefulness of model cards in deployment decisions?

Questions should not contain hidden conclusions. “Why does the superior model improve accuracy?” assumes superiority and improvement. “How do the models differ in accuracy and error distribution?” is more neutral.

A proposal normally has one central question and several subordinate questions. For example:

Central question: How effectively does data augmentation improve the robustness of a skin-lesion classifier across image-acquisition conditions?

1. How does augmentation affect performance on the original test distribution?

2. How does it affect performance on a shifted distribution?

3. Does its effect differ across relevant demographic or clinical subgroups?

4. What trade-offs occur in calibration, computation, and inference time?

The questions should be answerable by the proposed data and analysis. A student cannot responsibly ask whether an AI system “is fair” in general; fairness depends on a defined population, task, outcome, metric, and threshold.

Objectives, Scope, and Contributions

Objectives convert questions into observable research actions. They should use verbs such as identify, compare, estimate, evaluate, classify, examine, design, validate, or interpret. “Understand AI” is too vague. “Evaluate calibration error for three classifiers on an independently held-out dataset” is specific.

A useful objective structure is:

• General objective: State the overall purpose.
• Specific objectives: Divide the purpose into sequential, measurable tasks.

Specific objectives should collectively answer the research questions. They should not merely describe activities. “Collect data” is an activity; “construct a balanced evaluation set representing three acquisition conditions” is a research objective.

Scope defines what the project covers. It may specify the task, population, dataset, geographical setting, model family, time period, metrics, and deployment environment. Delimitations are deliberate boundaries chosen by the researcher, such as evaluating image classification rather than object detection. Constraints are limitations not fully controlled by the researcher, such as limited computing capacity, restricted data access, or a small sample.

A contribution is the value added by the study. Possible contributions include:

• empirical evidence about performance or generalization;
• a dataset, benchmark, annotation scheme, or reproducible pipeline;
• a methodological comparison;
• a theoretical explanation;
• a software artifact or prototype;
• design guidance for practitioners; or
• evidence relevant to policy and governance.

A contribution need not be a revolutionary algorithm. A carefully controlled replication, negative result, or analysis showing that a popular method fails under realistic conditions can be valuable.

Variables, Constructs, and Operational Definitions

A construct is an abstract concept, such as trust, robustness, fairness, or usability. A variable is a measurable representation of a construct or condition. Operationalization specifies exactly how the variable will be observed or calculated.

In a supervised ML experiment, the independent variable may be training strategy, the dependent variable may be test performance, and control variables may include model architecture, training budget, preprocessing, and random-seed policy. In a user study, the independent variable might be explanation type, while decision accuracy, time, and confidence are dependent variables.

Variables may be categorical, ordinal, discrete, or continuous. They may also be manipulated, observed, or controlled. A proposal must distinguish the target population from the sample and the training data from the evaluation data.

Operational definitions prevent ambiguity. “Robustness” could mean performance retention under image corruption, resistance to distribution shift, or sensitivity to adversarial perturbation. The proposal must specify the chosen meaning and evaluation procedure.

For classification, accuracy is commonly defined as

Accuracy = (TP + TN) / (TP + TN + FP + FN)

where TP and TN are true positives and true negatives, and FP and FN are false positives and false negatives. This measure is appropriate when class costs and class distributions make overall correctness meaningful. It can be misleading with severe class imbalance. Precision is TP/(TP+FP), and recall is TP/(TP+FN); each requires a nonzero denominator. Metrics should be selected according to the research question and consequences of errors, not convenience.

Hypotheses, Assumptions, and Propositions

A hypothesis is a logically derived, empirically testable statement about an expected relationship or difference. For example: “Models trained with condition-specific augmentation will have higher macro-averaged F1 scores on the shifted test set than models trained without augmentation.” The null hypothesis would state that no difference exists according to the prespecified comparison.

A hypothesis is not mandatory for exploratory or qualitative research. Forcing hypotheses into an exploratory study can create artificial certainty. Qualitative proposals may instead use sensitizing concepts and open research questions.

Assumptions are conditions accepted as reasonable for the study, such as the validity of labels, independence of test examples, or adequate representation of the target population. Assumptions should be stated because violations can threaten validity. For example, random train-test splitting assumes that near-duplicate images do not cross the split; if they do, performance may be inflated.

Constructing a Feasible Research Proposal

A proposal is a reasoned plan, not a promise of a particular result. A strong proposal commonly contains:

1. Title: Specific, concise, and informative.

2. Background and problem statement: Establishes context and significance.

3. Literature review: Synthesizes evidence and identifies the gap.

4. Purpose, questions, and objectives: Aligns the investigation.

5. Conceptual framework: Shows expected relationships among constructs.

6. Methodology: Describes design, data, sampling, procedures, tools, and analysis.

7. Ethics and governance: Addresses consent, privacy, bias, security, and ownership.

8. Evaluation plan: Defines success criteria and comparison conditions.

9. Resources and budget: Identifies personnel, software, hardware, data, and access.

10. Timeline and milestones: Demonstrates feasibility.

11. Risks and mitigation: Anticipates failure and ethical hazards.

12. Expected contributions and dissemination: Explains who may use the findings.

13. References and appendices: Supports scholarly verification.

Alignment is a central quality criterion. The problem should justify the questions; the questions should determine the objectives; the objectives should determine the data and methods; and the evaluation should determine whether the objectives were achieved. Misalignment occurs when a proposal asks about fairness but measures only accuracy, or claims causal effects from a purely observational dataset.

Evaluation Criteria

Evaluation criteria should be established before analyzing results. They may include predictive performance, statistical uncertainty, generalization, computational cost, interpretability, fairness, usability, reliability, and reproducibility. A proposal should specify baselines, validation strategy, test-set separation, and acceptable thresholds where possible.

For model comparisons, repeated runs or confidence intervals may be appropriate because random initialization and sampling can affect results. Statistical significance does not by itself establish practical importance. A small improvement may be statistically detectable but operationally irrelevant, while a practically important effect may require a larger study to estimate precisely.

Timeline and Resources

A realistic timeline includes literature review, protocol design, ethics approval, data acquisition, preprocessing, pilot testing, implementation, analysis, writing, and revision. Dependencies should be made explicit: data access may be required before preprocessing; ethics approval may be required before collecting participant data.

Resources include computing hardware, storage, software libraries, annotation tools, participant compensation, domain expertise, and supervisory support. Students should estimate whether the project can be completed with available resources rather than assuming unlimited cloud computing or proprietary data.

Risk Management

Typical risks include unavailable data, label noise, data leakage, model instability, insufficient compute, participant recruitment failure, privacy breaches, and schedule delays. Mitigation may involve a backup dataset, a smaller baseline experiment, version-controlled code, access controls, anonymization, preregistered analysis, and staged milestones. A proposal should identify decision points at which the scope will be reduced without invalidating the central question.

Worked Example: An Illustrative AI/ML Proposal

Background: A university wants to deploy a model that predicts whether students may need academic support. Historical records contain attendance, assessment, and course information. A highly accurate model could still be harmful if it systematically produces more false alarms for particular groups or uses information that should not influence support decisions.

Problem: Existing predictive studies often report aggregate accuracy but provide limited analysis of subgroup error rates, calibration, and the practical consequences of false positives and false negatives.

Purpose: To evaluate the predictive and governance characteristics of three baseline classifiers for early academic-support identification.

Questions:

1. Which classifier provides the best discrimination on a temporally held-out cohort?

2. Are error rates and calibration comparable across predefined student groups?

3. What information is necessary for prediction, and which features raise governance concerns?

Objectives: The study will define an approved feature set, train three baseline models using historical data, evaluate them on a later cohort, compare subgroup performance, and produce recommendations for human review rather than automated sanction.

Design: A retrospective comparative study will use a temporal split: earlier records for training and a later cohort for testing. This better reflects deployment than a random split when future prediction is intended. The outcome must be defined independently of the model, such as participation in an approved support program, while acknowledging that participation may reflect unequal access rather than academic need.

Evaluation: The study will report discrimination metrics, confusion matrices, calibration, subgroup comparisons, and resource requirements. No model will be deployed solely on the basis of aggregate performance. A domain expert will review feature legitimacy and possible proxy variables.

Ethics: Student records require authorization, data minimization, secure storage, controlled access, and, where applicable, institutional review. Students should not be denied opportunities because of a prediction. The proposal must explain human oversight, contestability, retention, and communication.

Risks and mitigation: If labels are biased, the study will report this limitation and avoid presenting predictions as objective truth. If subgroup sample sizes are too small for stable estimates, the study will report uncertainty and refrain from overinterpreting comparisons. If performance is inadequate, the valid outcome is a documented non-deployment recommendation.

Learning points: The example shows that a research proposal evaluates not only whether a model works, but also what “works” means, for whom, under what conditions, and with what consequences.

Ethical, Legal, and Governance Considerations

AI/ML research can affect privacy, autonomy, equality, safety, intellectual property, and access to services. Researchers must obtain appropriate approval before collecting human data, explain participation and withdrawal rights, minimize sensitive data, and secure stored information. Anonymization is not always irreversible; combinations of variables may re-identify individuals.

Researchers should examine consent and secondary data use, licensing restrictions, copyright, data provenance, biometric sensitivity, and applicable data-protection law. They should document dataset composition, exclusions, labeling procedures, known biases, and intended uses. Model cards, datasheets, audit logs, and reproducible experiment records support accountability.

Ethical analysis must continue after approval. A technically valid study can still be irresponsible if results are likely to be misused or if limitations are concealed. Responsible reporting includes negative findings, uncertainty, failed experiments, data exclusions, and conflicts of interest.

Contemporary Developments and Best Practices

Current AI research increasingly emphasizes foundation models, synthetic data, federated learning, privacy-preserving computation, responsible AI, energy-aware computing, and evaluation under distribution shift. These developments create new research opportunities but also demand careful scope. A student proposal involving a large foundation model should specify model version, prompts or fine-tuning procedure, access conditions, data governance, cost, and reproducibility limitations.

Best practices include:

• conduct a structured literature search and maintain a source matrix;
• define the primary outcome before inspecting results;
• separate training, validation, and final testing data;
• prevent data leakage and document preprocessing;
• use meaningful baselines rather than only comparing advanced models;
• report uncertainty and practical significance;
• analyze relevant subgroups without stereotyping or exposing identities;
• preregister hypotheses or analysis plans when appropriate;
• use version control and experiment tracking;
• preserve a reproducible environment where licensing permits;
• distinguish exploratory findings from confirmatory claims; and
• write limitations as part of the contribution, not as an afterthought.

Chapter Summary

A research project develops from a broad topic into a defensible problem, gap, purpose, question, objective, method, and contribution. Research gaps may be knowledge-based, methodological, contextual, population-based, evidential, implementation-related, contradictory, or evaluative. Research questions must be clear, focused, feasible, significant, and ethically answerable. Objectives translate questions into observable actions.

A proposal defines scope, constructs, variables, operational measures, assumptions, constraints, and expected contributions. Hypotheses are appropriate when relationships or differences can be tested; exploratory and qualitative studies may use open questions instead. Alignment among problem, questions, objectives, data, methods, and evaluation is essential. Feasibility requires realistic timelines, resources, risks, and fallback plans. In AI/ML, responsible research also requires attention to privacy, bias, data provenance, fairness, security, reproducibility, and the consequences of deployment.

Key Terms

Research problem: A specific issue or deficiency that can be investigated systematically.

Research gap: An important limitation, absence, inconsistency, or unanswered question in existing knowledge or practice.

Research question: A focused question that an investigation is designed to answer.

Research objective: A specific action undertaken to address a research question.

Purpose statement: A concise statement of what the study intends to accomplish.

Scope: The boundaries of what a study covers.

Delimitation: A boundary deliberately selected by the researcher.

Constraint: A limitation imposed by circumstances or available resources.

Construct: An abstract concept represented through measurable indicators.

Variable: A characteristic or condition that can take different values.

Operational definition: A precise description of how a construct or variable will be measured.

Hypothesis: A testable statement about an expected relationship or difference.

Assumption: A condition accepted as reasonable for conducting the study.

Research paradigm: Assumptions about reality, knowledge, evidence, and inquiry.

Research design: The overall plan connecting questions, data, procedures, and analysis.

Baseline: A reference method used for meaningful comparison.

Data leakage: Unintended use of information during training that would not be available at prediction time.

Generalizability: The extent to which findings apply beyond the studied sample or setting.

Contribution: The knowledge, method, evidence, artifact, or guidance added by a study.

Research proposal: A structured plan explaining what will be studied, why, and how.

Review Questions

Short Answer Questions

1. Distinguish a research topic from a research problem.

2. What is a research gap, and why must it be supported by evidence?

3. Define scope, delimitation, and constraint.

4. Why is operationalization important in AI/ML research?

5. State two ethical issues that may arise when using human-related data.

Long Answer Questions

1. Explain the progression from topic to research contribution, using an AI/ML example.

2. Compare positivist, interpretivist, and pragmatic research orientations.

3. Discuss the relationship among research questions, objectives, variables, and hypotheses.

4. Describe the essential components of a feasible research proposal.

5. Explain why evaluation should include more than predictive accuracy in an AI/ML study.

Analytical / Application Questions

1. A student proposes to “develop an accurate and fair hiring AI system.” Critique the proposal and rewrite its problem statement, research question, scope, and evaluation criteria.

2. Design a research proposal outline for testing whether data augmentation improves model performance under distribution shift. Identify the baseline, variables, risks, and success criteria.

3. A model performs well on a random test split but poorly on data collected six months later. Analyze possible causes and recommend changes to the research design and reporting.

Multiple Choice Questions with Answers

Question 1

Which statement best describes a research gap?

A. A broad subject that interests the researcher

B. A meaningful limitation or unanswered issue in existing knowledge

C. A list of all papers found during a literature search

D. A prediction about the final research result

Correct Answer: B. A meaningful limitation or unanswered issue in existing knowledge

Question 2

Which is the most appropriately focused research question?

A. How will AI change the world?

B. Is machine learning useful?

C. How does condition-specific augmentation affect classifier performance under image-distribution shift?

D. Why are neural networks superior?

Correct Answer: C. How does condition-specific augmentation affect classifier performance under image-distribution shift?

Question 3

In an ML experiment, the dependent variable is generally the variable that is

A. manipulated by the researcher

B. measured as an outcome

C. held constant in every experiment

D. excluded from the evaluation set

Correct Answer: B. measured as an outcome

Question 4

Which item is a deliberate delimitation?

A. Limited access to cloud computing

B. Unexpected participant withdrawal

C. Choosing to study image classification rather than object detection

D. Unavailability of historical records

Correct Answer: C. Choosing to study image classification rather than object detection

Question 5

What is the main purpose of a baseline model?

A. To guarantee the proposed method succeeds

B. To provide a meaningful reference for comparison

C. To replace the final evaluation dataset

D. To eliminate the need for documentation

Correct Answer: B. To provide a meaningful reference for comparison

Question 6

Which practice most directly helps prevent data leakage?

A. Using the test set repeatedly during model tuning

B. Separating information available at prediction time from future information

C. Increasing the number of model parameters

D. Removing all validation procedures

Correct Answer: B. Separating information available at prediction time from future information

Question 7

Accuracy is calculated as

A. TP/(TP+FP)

B. TP/(TP+FN)

C. (TP+TN)/(TP+TN+FP+FN)

D. FP/(TN+FP)

Correct Answer: C. (TP+TN)/(TP+TN+FP+FN)

Question 8

Why may accuracy be misleading for a severely imbalanced classification problem?

A. It cannot be computed for classifiers

B. It may appear high even when the minority class is poorly detected

C. It always measures calibration instead

D. It requires qualitative interviews

Correct Answer: B. It may appear high even when the minority class is poorly detected

Question 9

Which proposal feature best demonstrates feasibility?

A. A claim that unlimited resources will be available

B. A realistic timeline, resource plan, and risk-mitigation strategy

C. A promise of a revolutionary result

D. A decision to omit limitations

Correct Answer: B. A realistic timeline, resource plan, and risk-mitigation strategy

Question 10

Which statement about hypotheses is correct?

A. Every qualitative study must include a statistical hypothesis

B. A hypothesis is a testable statement about an expected relationship or difference

C. A hypothesis is the same as a research topic

D. A hypothesis guarantees the predicted result

Correct Answer: B. A hypothesis is a testable statement about an expected relationship or difference