CSEG3060 · Research Methodology in Computer Science
From the research problem to the finished document: objectives, hypotheses, significance, scope, feasibility — and the ten components
A proposal is read three times: as a promise, as a plan, and later as a contract you will be measured against.
Unit IV · Lecture 2 of 3 · 60 minutes · CO4 · B.Tech (CSE) Semester V
Dr. Mohsin Furkh Dar · School of Computer Science, UPES
→ / Space next · ← back · O overview · N speaker notes · F full screen
Learning outcomes · CO4
Distinguish a research problem, a research question and a research hypothesis
Identify and refine a research problem in computing using structured techniques, and state it to SMART criteria
Formulate measurable objectives and testable hypotheses — and recognise when hypotheses do not apply
Articulate significance, scope and feasibility with rigour, not rhetoric
Construct all ten components of a proposal document, including work plan, timeline and budget
Evaluate a sample proposal against standard academic criteria
What a proposal is
It convinces reviewers, funders or supervisors that the work is worthwhile, feasible and methodologically sound. Nothing else in the document matters if this fails.
It forces you to think systematically about problem, approach, resources and schedule — before committing eighteen months to them.
Once approved it becomes the benchmark against which your progress and deliverables are evaluated. What you promised is what you will be asked for.
The first half builds the intellectual core — problem, objectives, hypotheses, literature, significance, scope, feasibility. The second half builds the document that carries it. The order is not arbitrary: a beautifully formatted proposal with an ill-defined problem is rejected, while a rough draft of a sharp idea can be rewritten.
Orientation
| Research problem | Defines the gap the study addresses |
| Research questions | Operationalise it into answerable inquiries |
| Objectives | State the concrete aims |
| Hypotheses | Testable predictions, where applicable |
| Literature review | Context, gaps, no duplication |
| Significance | Why it matters |
| Scope | What is and is not covered |
| Feasibility | That it can realistically be executed |
| 1 Title page | Identifies project and investigators |
| 2 Abstract | Single-paragraph summary of the whole |
| 3 Introduction | Context, background, motivation |
| 4 Objectives | Specific goals |
| 5 Methodology | Methods, tools, datasets, analysis |
| 6 Work plan | Tasks, phases, responsibilities |
| 7 Timeline | Gantt chart or milestone table |
| 8 Budget | Financial resources required |
| 9 Expected outcomes | Deliverables, publications, impact |
| 10 References | Consistent citation style |
Intellectual core · 1 of 8
This connects directly to Unit III: a gap is a claim about the literature, and must be evidenced by a documented search rather than asserted because you have not read enough.
Intellectual core · 2 of 8
Generate broad themes, then cluster them into narrower sub-issues.
Apply What, Why, Who, Where, When, How to dissect a broad topic into its components.
Map causes, effects and the core issue — separating symptoms from the thing worth solving.
Population — the data or the system users
Intervention — your proposed technique or model
Comparison — the baseline or state of the art
Outcome — the measured metric: accuracy, latency, energy
Timeframe — the evaluation period
PICO(T) is the fastest route from a vague interest to a comparable, measurable claim — it forces a baseline and a metric into existence.
Identify the limitations that authors of seminal papers state themselves. The “Limitations” and “Future work” sections of strong papers are the cheapest source of legitimate research problems available to you.
Intellectual core · 3 of 8
A defined sub-domain, not a field
Precision, recall, F1, throughput
Within your technical and resource capacity
A recognised gap or stakeholder need
Completable in the project duration
“Machine learning is not very accurate.”
“Existing convolutional neural network architectures for skin lesion classification achieve a maximum F1-score of 0.81 on the ISIC 2019 dataset, which is insufficient for clinical deployment where false negatives carry high patient risk. This study seeks to improve classification accuracy while maintaining interpretability.”
Intellectual core · 4 of 8
Primary (general) objective — the overarching goal that captures the spirit of the research.
Secondary (specific) objectives — discrete, sequential sub-goals that together deliver it.
Objectives are the benchmark against which the project’s success is later judged. They define what success looks like.
design develop evaluate compare analyse implement validate
Too broad — “To study machine learning”.
Objectives confused with methods — “To use Python and TensorFlow”.
Activities listed as outcomes — “To conduct a literature review” is a task, not an objective.
Avoid study, understand and explore unless followed by a measurable complement.
Worked example
To analyse the limitations of existing fixed-time and actuated traffic control systems through a systematic literature review
To develop a multi-agent reinforcement learning model that adjusts signal phases from real-time traffic density
To implement the framework in a simulated urban grid using SUMO and benchmark it against baseline controllers
To evaluate performance on average vehicle delay, throughput, queue length and emissions
To assess the transferability of the trained policy across varying traffic demand profiles
Read the sequence: survey → build → implement and benchmark → measure → generalise. Each objective is a verb plus a measurable complement, and each depends on the one before it — that ordering is what makes the work plan on slide 18 possible.
Intellectual core · 5 of 8
| Type | Description | Example |
|---|---|---|
| Null (H₀) | No significant effect or difference | “There is no significant difference in classification accuracy between the proposed CNN and baseline ResNet-50 on CIFAR-10.” |
| Alternative (H₁ / Hₐ) | Asserts a meaningful effect | “The proposed CNN achieves significantly higher classification accuracy than ResNet-50 on CIFAR-10.” |
| Directional | Specifies the direction of the effect | “Increasing transformer layers beyond 12 will degrade performance on low-resource translation tasks due to overfitting.” |
| Non-directional | An effect, without direction | “There is a relationship between dataset size and model robustness.” |
Testable by statistical or empirical means · falsifiable in principle · specific about variables and their relationship · grounded in theory or prior literature · concise, ideally one declarative sentence.
The problem is the deficiency. The question makes it answerable. The hypothesis is a prediction about the answer. Examiners test this three-way distinction routinely.
Intellectual core · 6 of 8
where μ is the population mean of a performance metric — accuracy, latency, throughput. You collect sample data and run a statistical test (paired t-test, Wilcoxon signed-rank) to decide whether to reject H₀ at significance level α, conventionally 0.05.
In exploratory, descriptive and design-science research. A study proposing a new software architecture pattern does not require hypothesis testing; it is evaluated by design principles, heuristics and qualitative criteria such as utility, elegance and reusability.
Forcing a hypothesis onto design-science work is a common and visible error — reviewers read it as a misunderstanding of your own paradigm.
Are you comparing measurable outcomes between conditions? If yes, state H₀ and H₁ and name the test. If you are building an artefact and arguing it is better by construction, state evaluation criteria instead.
Intellectual core · 7 of 8
| Narrative | Broad synthesis — doctoral proposals |
| Systematic | Protocol-driven, quality-assessed |
| Mapping study | Categorises the landscape — SE, HCI |
| Meta-analysis | Statistical aggregation of effect sizes |
| Scoping | Breadth and depth — emerging topics |
Background → theoretical foundations → state of the art → critique → gap identification. The last move must be explicit: state what is missing and how your work addresses it.
Annotation without synthesis — summaries, not integration
Citation bias — Western, English-language, industry-heavy
Outdated evidence — ignoring recent preprints and benchmarks
No critical voice — claims accepted at face value
Mini case — FL for health records: Konečný (2016) → Wei (2020) → Li (2023) surfaces recurring limits that justify a new secure aggregation protocol.
Intellectual core · 8 of 8
A new complexity bound or model
Improved procedures, tools, evaluation frameworks
New datasets, benchmarks, experimental insight
Solves an industry problem; informs policy
Ethical, environmental, equity considerations
“This study is significant because it [addresses / resolves / extends] [specific gap] in [domain]. The findings will [advance theory / inform practice / support stakeholders] by [specific mechanism].”
Filled in: “…addresses the lack of lightweight intrusion detection models for edge devices in IoT networks. The framework will advance theory in resource-constrained deep learning and offer practitioners a deployable artefact that identifies threats in real time without cloud infrastructure.”
Topic (algorithms, systems, phenomena) · theoretical (frameworks adopted or excluded) · methodological (methods, tools, datasets) · geographic or demographic (social computing, HCI, policy) · temporal (data collection or operation period).
Sample: “Limited to the evaluation of PPO, SAC and DQN for adaptive traffic signal control in simulated urban grids of up to 20 intersections. Real-world deployment, metropolitan-scale networks and pedestrian-priority optimisation fall outside this scope.”
Can it actually be done?
| Technical | Are the tools, libraries and hardware accessible? |
| Data | Are datasets available, accessible and legally usable? |
| Operational | Are institutional approvals (IRB, ethics) in place? |
| Schedule | Is the timeline realistic given task dependencies? |
| Economic | Is the budget adequate for compute, travel, dissemination? |
| Human resources | Does the team have the expertise? |
SWOT analysis · Gantt chart for scheduling and milestones · risk register (risk, likelihood, impact, mitigation) · proof of concept — a small preliminary implementation that validates your assumptions.
Data — insufficient or biased datasets, licensing
Compute — hardware failure, GPU shortage, cost overrun
Algorithmic — non-convergence, instability, poor generalisation
Operational — staff attrition, third-party API dependency
Ethical — personal data, dual-use in security research
“The study is technically feasible because all required libraries (PyTorch 2.1, Ray RLlib) are open-source, and computation can be performed on the institutional GPU cluster (4× NVIDIA A100). The PEMS-BAY traffic dataset is publicly available through the California Department of Transportation. A preliminary PoC using SUMO on a 4-intersection grid has been completed, confirming that a single training episode requires approximately 12 minutes — well within the project’s computational budget.”
Part 2 begins
Everything so far is the argument. The remaining ten components are the vehicle that carries it to a reviewer who has forty proposals to read and will spend four minutes on yours before deciding how carefully to read the rest.
Title page
and Abstract
Introduction
and Objectives
Methodology
the heart
Work plan
timeline, budget
Outcomes
and References
The single test the whole document must pass: every objective is addressed by the methodology, produces a deliverable in the outcomes, and occupies time in the work plan. Reviewers check this alignment first, because misalignment is the cheapest signal that a proposal was assembled rather than designed.
Components · 1 and 2
Title — 10–15 words, descriptive, containing the key technical terms that will index it in databases · PI and Co-PIs with designations, affiliations, contacts · institution and department · funding agency and call identifier · submission date and version · declarations and signatures where required.
Weak — “A Study on Machine Learning”
Strong — “A Hybrid Deep Learning Framework for Real-Time Anomaly Detection in IoT-Enabled Smart Grid Networks”
No abbreviations unless universally understood (AI, ML, IoT) · no terminal full stop · consistent title or sentence case.
1 Context (1–2 sentences) — background and domain
2 Problem / gap (1–2) — what is missing
3 Objective (1) — what this research aims to do
4 Methodology (2–3) — how it will be done
5 Expected outcomes (1–2) — deliverables and impact
6 Keywords (4–6) — for indexing
Four mistakes: citations or equations in the abstract · an abstract that does not match the proposal · exceeding the word limit — grounds for desk rejection · undefined jargon.
Worked abstract (condensed): “The rapid proliferation of IoT devices in healthcare has introduced unprecedented cybersecurity vulnerabilities in patient monitoring. [context] Existing intrusion detection struggles to identify zero-day attacks while maintaining low false-positive rates. [gap] This research proposes a hybrid framework combining CNNs for spatial feature extraction and LSTMs for temporal pattern recognition. [objective + method] The model will be trained on CIC-IoT-2023 and tested in a simulated smart hospital. The expected outcome is real-time detection exceeding 98 percent accuracy with response latency under 50 ms. [outcomes]”
Component · 3
1 Background — the foundational knowledge needed to understand the problem
2 Problem statement — what is wrong with current approaches, and why does it matter?
3 Motivation — why now: prevalence with statistical evidence, economic or regulatory implications, technological enablers
4 Literature gap — what is missing after the survey
5 Contribution summary — the enumerated claims
Four sentences take the reader from a field to a specific unresolved issue. Each step discards everything the next step does not need.
“However, none of the existing approaches have addressed…”
“To the best of the authors’ knowledge, no prior study has investigated…”
“A critical limitation of current methods is…”
“The major contributions are: (i) a novel hybrid CNN-LSTM architecture for IoT anomaly detection, (ii) a curated benchmark dataset for smart healthcare scenarios, (iii) extensive empirical evaluation demonstrating superior accuracy and latency, and (iv) an open-source reference implementation for reproducibility.”
Component · 5
CS proposals carry definitions, equations and algorithm sketches. Define every symbol:
L(θ) = −(1/N) Σᵢ [ yᵢ·log(ŷᵢ) + (1−yᵢ)·log(1−ŷᵢ) ]binary cross-entropy loss; θ the parameters, N the sample count, yᵢ the true label, ŷᵢ the predicted probability
Algorithm 1 · Hybrid CNN-LSTM detection In: traffic sequence X = {x₁ … x_T} Out: score s ∈ [0,1] 1: F = CNN(X) spatial features 2: H = LSTM(F) temporal dependencies 3: s = σ(W·H + b) anomaly scoreOpen-source repository (GitHub, GitLab) · public datasets or clear licensing for proprietary data · hardware specified (NVIDIA RTX 3090, 64 GB RAM) · random seeds and hyperparameter settings.
Languages Python, R, Julia, C++, Java · Frameworks TensorFlow, PyTorch, scikit-learn, Keras · Data pandas, NumPy, Matplotlib, Seaborn · Experiment management MLflow, Weights & Biases, DVC · Cloud AWS SageMaker, Google Vertex AI, Azure ML
Components · 6 and 7
| Work package | Title | Months | Lead | Deliverable |
|---|---|---|---|---|
| WP1 | Literature review and problem refinement | 1–3 | PI | Review report |
| WP2 | Dataset collection and preprocessing | 3–5 | Co-PI | Curated dataset |
| WP3 | Model design and implementation | 5–9 | Research scholar | Working prototype |
| WP4 | Experimental evaluation | 9–13 | Research scholar | Results and analysis |
| WP5 | Deployment and real-time testing | 13–16 | Industry partner | Deployed system |
| WP6 | Documentation, publication, reporting | 16–18 | All | Papers, final report |
M1 (month 3) literature review report complete · M2 (month 5) dataset finalised and preprocessed · M3 (month 9) working prototype at baseline performance · M4 (month 13) full experimental evaluation complete · M5 (month 18) final report and project closure.
Every WP has a clear, verifiable deliverable — a report, dataset, code module or publication. Dependencies between WPs are stated explicitly. Every major WP carries a risk mitigation line: “if GPU procurement is delayed, cloud instances will be used as fallback”.
The timeline renders this chronologically as a Gantt chart — rows are tasks, columns are months, shaded cells are activity. Overlapping bars are not a flaw; they are how you show parallel work and where the critical path actually runs.
Component · 8
| Manpower | JRF/SRF salaries, project staff |
| Equipment | Hardware and software purchases |
| Consumables | Cloud credits, datasets |
| Travel | Conference registration, airfare, per diem |
| Contingency | Typically 5–10 percent |
| Institutional overhead | Often 10–20 percent of total |
| Publications | APCs, open-access fees |
| Item | Qty | Total (INR) |
|---|---|---|
| GPU workstation (RTX 4090, 64 GB) | 1 | 250,000 |
| JRF stipend (24 months) | 1 | 744,000 |
| Cloud compute (AWS) | 12 mo | 120,000 |
| Conference travel | 2 | 150,000 |
| Publications (2 journals) | 2 | 200,000 |
| Contingency | — | 100,000 |
| Total | 1,564,000 |
Components · 9 and 10
2–3 journal papers in SCI/Scopus venues · 1–2 conference papers · one Ph.D. thesis where applicable
Open-source toolkit · curated and annotated benchmark dataset · prototype or proof-of-concept system
Trained research scholars and postgraduate students · workshops and training programmes conducted
Reduced cyberattack incidents in the target deployment · technology transfer to industry partners · policy recommendations
“An article will be submitted to IEEE Trans. Information Forensics and Security by month 18.” · “The framework will achieve at least 98 percent detection accuracy on the benchmark.” Both checkable; “we expect good results” is not.
| IEEE | Most engineering and CS journals and conferences |
| ACM | Computing conferences and journals |
| APA 7 / Springer | Social-science-adjacent CS; LNCS proceedings |
Tools: Zotero · Mendeley · EndNote · BibTeX / BibLaTeX
The reviewer's first check
“Privacy-Preserving Federated Learning for Cross-Organisation Medical Imaging Analysis”
Introduction — medical AI matters; HIPAA and India’s DISHA block data sharing; FL is the emerging answer.
Objectives — (1) an FL framework enabling multi-hospital collaboration without sharing raw data; (2) integrated differential privacy with formal guarantees; (3) validation on chest X-ray, brain MRI and histopathology benchmarks.
Methodology — federated averaging, privacy budget ε, simulated hospital clients.
Work plan — six WPs over 24 months. Timeline — Gantt with quarterly milestones. Budget — ₹22 lakh: JRF, GPU server, cloud, travel.
Outcomes — two SCI publications, an open-source framework, a radiologist workshop.
References — 45 IEEE-format entries including McMahan et al. (2017) and Dwork & Roth (2014).
Consolidation
A full research proposal in the prescribed format (10 marks, CO4). Use the ten components as your section headings and the alignment test as your self-check before submission. Due at lecture V.3.
Kitchenham SLR guidelines · Konečný et al. (2016) · Wei et al. (2020) · McMahan et al. (2017) · Dwork & Roth (2014) · your target agency’s own proposal template — DST-SERB, MEITY or AICTE
Next — Lecture 3: Presenting and Defending Your Proposal: presentation architecture, audio-visual aids, communication skills, handling committee queries, and the rubric you will actually be scored against.