What research is · what you promise to produce · and what you will produce it from.
Definition, objectives, and the five types: exploratory, descriptive, analytical, theoretical, empirical.
Turning an aim into SMART objectives an examiner can tick off.
Primary vs secondary, qualitative vs quantitative — and the three collection tools.
Your own type, three SMART objectives, and a named data source with an access route.
Dr. Mohsin Furkh Dar · School of Computer Science · 5th Semester
B.Tech CSE · 55 minutes
Navigate: → / Space next · ← back · O slide list ·
D dark mode · F full screen
1 “AI coding assistants help students write code faster.”
An opinion. Nobody counted anything. It cannot be wrong, because nothing was claimed precisely enough to check.
2 “In a survey of 400 UPES students, 78% report using an AI assistant at least weekly.”
A fact about a sample. True, useful, checkable — and it explains nothing. This is descriptive work, and it is a legitimate first step.
3 “Students using an assistant finished the debugging tasks 22% faster (n = 120, 3 tasks, counterbalanced) but left 1.6× more silent defects; the speed gain did not survive when the defect-repair time was added back.”
A research claim: population, comparison, measure, magnitude, and a result that could have come out the other way.
Three sentences about one topic. The difference is not the subject and not the technology — it is how the sentence was produced and what could refute it.
Notice the word every definition contains: systematic. Not clever, not novel, not difficult — systematic.
Research is a systematic, documented investigation that produces a claim about the world which someone else could check — and which could turn out to be wrong.
This is the same falsifiability test from Lecture I.1. Sentence 1 on the previous slide fails it; sentence 3 passes it.
1 · Systematic — a stated procedure, not a wander.
2 · Empirical or logical — settled by data or by proof, never by authority.
3 · Bias-controlled — the design tries to prove you wrong.
4 · Replicable — a stranger with your write-up can repeat it.
5 · Cumulative — it stands on named prior work.
6 · Communicable — unpublished research is not yet research.
Two different questions hide under the word objective. Keep them apart.
Kothari's four classical groups — a study usually sits mainly in one:
In four verbs, research exists to describe, explain, predict and design — computing does a great deal of the fourth.
The three to five sentences in your proposal that say what this study will produce. They are:
Part B of this hour is entirely about writing these so they can be ticked off.
The link between the two: the kind of objective you have (explore / describe / diagnose / test) decides the type of research you must run — which is the next slide.
“What is going on here?”
No hypothesis yet. Small, flexible, mostly qualitative. Output = sharper questions, categories, a hypothesis worth testing.
CS: 12 interviews with clinic staff before you design a triage tool.
“What is the case?”
Measures and characterises without manipulating anything. Surveys, measurement studies, repository mining.
CS: mining 50,000 GitHub repositories to report how often CI is left red, and for how long.
“What does it mean?”
Takes facts or data already available and analyses them critically to find relationships and evaluate claims.
CS: re-analysing those same repositories to test whether long red-CI spells predict defect density, controlling for team size.
“What must be true?”
Abstract and formal. Proofs, bounds, models, frameworks. No data collection at all; the warrant is the argument.
CS: proving no deterministic protocol achieves consensus under asynchrony with one crash fault.
“What actually happens?”
Data-based and experience-based. Ideally you manipulate something, control the rest, and repeat.
CS: two schedulers, same workload, 3 seeds, reported as mean ± SD with an effect size.
Exploratory work finds the question. Descriptive work measures. Analytical work relates. Theoretical work proves. Empirical work tests.
| Type | Question | Evidence | Typical output | Where it goes wrong |
|---|---|---|---|---|
| Exploratory | What is going on, and what should I even ask? | Interviews, field notes, pilot probes, open-ended responses | Categories, themes, a testable hypothesis | Reported as if it proved something; n = 8 becomes “users prefer…” |
| Descriptive | What is the case, and how much of it? | Surveys, censuses, measurement studies, mined repositories | Distributions, rates, characterisations, a public dataset | Slides into causal language: “because”, “leads to”, “improves” |
| Analytical | What relationships hold in data I already have? | Existing data + statistical or critical analysis | Associations, models, a critical evaluation of prior claims | Correlation read as causation; confounders never named |
| Theoretical | What must hold for every input, always? | Definitions, axioms, derivation, proof | Theorems, bounds, impossibility results, formal models | A valid proof of an assumption that no real system satisfies |
| Empirical | What happens when I change this and hold the rest fixed? | Controlled experiments, benchmarks, A/B tests, user studies | Comparative results with variance and effect size | Weak baseline, one seed, tuned on the test set |
Two more pairs you will meet in the literature and should be able to place: basic vs applied (is the goal knowledge or a use?) and qualitative vs quantitative (is the evidence meaning or magnitude?). Neither is a sixth type — they are different axes cutting across the five above.
So “is this descriptive or empirical?” is often the wrong question. A measurement study is both. Ask instead: what would make this claim wrong?
“Does an AI assistant help first-year students debug?”
The type follows the question, and the question changes as the project moves.
Item 1 of 6 · Score: 0
—
The direction of travel, in plain language. Not measurable, and that is allowed.
“To make automated crop-disease diagnosis usable on real farms in Uttarakhand.”
The research question restated as an intention. Bounded, but still one sentence.
“To measure and explain the accuracy degradation of a PlantVillage-trained classifier on field-captured images for three Uttarakhand crops.”
The steps that, if all completed, discharge the general objective. Each one is SMART.
assemble · reproduce · quantify · attribute · measure
The test that catches almost every weak objective set: if I completed all your specific objectives and nothing else, would your general objective be answered?
If the answer is “no, you would also need to…”, that missing thing is an objective you forgot. If the answer is “yes, and more besides”, you have an objective that belongs to a different project.
Names the population, task, baseline and setting. No umbrella nouns.
Test: two strangers reading it design the same study.
Names the metric and the threshold — not just “accuracy” but “top-1, ≥ baseline + 5 pp”.
Test: you can state, today, the number that would count as failure.
Fits your data, compute, skill, access and calendar — which multiply; any zero is fatal.
Test: a crude end-to-end version exists within one week.
Serves this research question, and someone would act differently on the answer.
Test: the “so what” chain ends at a named person or group.
Carries a date or week number from your work plan, not “by the end”.
Test: it appears as a bar on your Gantt chart with a finish week.
One worked objective, all five letters visible: “Quantify the top-1 accuracy drop (M) of the ResNet-50 PlantVillage baseline (S) on 900 field-captured images of tomato, potato and maize collected in Dehradun (S), reporting mean ± SD over 3 seeds against the published 99.2% figure (M), using the annotation set already collected (A), to establish whether the deployment gap reported by extension officers is real (R), by week 7 (T).”
| As written | What fails | Repaired |
|---|---|---|
| “To study deep learning techniques for medical image analysis.” | S, M, T — a syllabus, not an objective; no observable at all. | “To compare U-Net, nnU-Net and SAM-Med2D on the BUSI breast-ultrasound set, reporting Dice and HD95 over 5 folds, by week 6.” |
| “To improve the accuracy of our classifier.” | M — improve from what, to what? No baseline and no threshold, so no result can fail. | “To raise macro-F1 on the held-out Hinglish test split from the 0.71 BERT baseline to ≥ 0.78, over 3 seeds, by week 9.” |
| “To develop a fully autonomous campus delivery robot.” | A — hardware, permissions, safety approval and six person-months you do not have. | “To evaluate three path-planning policies in a Gazebo model of the campus loop on 50 scripted pedestrian scenarios, reporting collisions per km, by week 8.” |
| “To count the weekly commits in our project repository.” | R — perfectly specific, measurable, achievable and dated, and it answers nothing. | “To test whether weekly commit frequency predicts post-release defect count across 200 repositories, controlling for team size, by week 10.” |
The fourth row is the one to remember: SMART is a filter, not a generator. A worthless objective can pass all five letters. Relevance is the letter students score themselves highest on and reviewers score them lowest on.
| Use | Avoid |
|---|---|
| measure · quantify · compare · characterise · evaluate · attribute · implement · validate · replicate · derive | understand · study · explore · look into · investigate · analyse (when unqualified) · gain insight into · work on |
The right-hand column names states of mind. You cannot show an examiner that you have understood something; you can show them a table.
The artefact rule. Each specific objective maps to one experiment, one table or figure, and one subsection of your report. An objective that produces no artefact is a wish.
1 · Measurable but meaningless. The metric drifts from the construct you actually care about — you promise “95% accuracy” on a set where the majority class is 94%. Passes M; means nothing. (The validity gap, Lecture I.4.)
2 · Achievable but trivial. You shrink until the objective is certain to succeed. If no possible result would surprise anyone, you have written a task, not a research objective. (Lecture I.2's task/project/research distinction.)
Calibration: three to five specific objectives for a semester. More objectives than working weeks is the single most common failure in Assignment 2 — and the one supervisors spot in ten seconds.
Item 1 of 6 · Score: 0
—
Buys you: exact fit to your objective, known
provenance, control over quality, and a contribution in its own right.
Costs you: time, money, ethics approval, recruitment, and the risk that the whole
semester depends on people replying to you.
Buys you: scale, speed, low cost, history you
could never collect yourself, and comparability with prior work.
Costs you: a fit gap, unknown collection bias, licence limits, stale definitions, and no
way to add the variable you wish someone had recorded.
Always search secondary first. The cheapest primary study is the one you did not have to run — and the reviewer will ask why you collected data that already exists. Then ask the two questions that decide whether you may use it: what was it collected for, and who is missing from it? Plus the licence, always.
Words, behaviour, artefacts. Answers why and how. Small purposive samples, flexible instruments, analysis by coding and themes, agreement reported with a statistic such as Cohen's κ. Generalises to theory, not to a population.
Numbers on a defined scale. Answers how much, how often, is the difference real. Larger samples, fixed instruments, analysis by descriptive then inferential statistics, reported with variance and effect size.
| Quantitative | Qualitative | |
|---|---|---|
| Primary | Your own latency traces; a 400-response Likert survey; timings from a controlled task study | Your interviews with clinic staff; think-aloud transcripts; structured observation field notes |
| Secondary | ImageNet; NASA/PROMISE defect data; monthly UPI transaction volumes; a Stack Overflow dump | Existing issue-tracker discussions; published incident post-mortems; policy documents; archived interview transcripts |
Mixed methods is the normal case, not a compromise: qualitative work tells you what to measure; quantitative work tells you whether it holds.
Use when you need self-reported facts, opinions or experiences from more people than you could ever sit with.
Use when what people do differs from what they say they do — which is almost always.
ts | participant | task | code | verbatim | observer note
Example
00:04:12 | P07 | book-lab | BACKTRACK | “wait, which slot did I pick?” | returned to step
2, third time
Six columns, one line per event. A log you cannot analyse mechanically is a diary — and a diary is not data.
Computing's own instrument: the apparatus that produces your numbers. In this field it replaces the psychologist's questionnaire as the default tool.
Privacy: a log becomes personal data the moment it holds a user id, an IP or content. Minimise fields, pseudonymise, set a retention period — and put all three in your ethics section.
Item 1 of 6 · Score: 0
—
★ = non-negotiable. Six starred items.
Tick the items on the left that are already true of your own project — not the ones you intend to do.
The two that end projects. Access — confirm in writing, in week 1, that you can actually obtain the data; a smaller slice of inaccessible data is still inaccessible. Leakage — hold out the test set before any modelling, split by group (patient, user, repository) not by row, and touch it once.
If you remember one sentence: an objective nobody can tick off, and data you cannot actually obtain, are the two ways a good research question still dies.
source | primary/secondary | qual/quant | tool |
access route | licence/consent | confirmed by. The last column is a name and a
date, not “should be fine”.Stamp the entry \version{5}{...}{after Lecture
I.7}.
QUIZ 1 — 25 marks, CO1, the full slot. 25 scenario questions in 30 minutes, covering all of Unit I. 25 August (CCVT) · 26 August (Full Stack AI).
Not definitions — every question hands you a situation to diagnose, exactly like the three drills in this deck. Assignment 1 is collected in the same slot: problem statement, scope and objectives.
You now have a question, a bounded scope, a type, objectives an examiner can tick off, and a data source with a name on it. That is CO1, complete.
Reading: Kothari, Research Methodology: Methods and Techniques, ch. 1 · Melville & Goddard, Research Methodology: An Introduction for Science & Engineering Students, ch. 1–3 · Ganesan, Research Methodology for Engineers · OECD, Frascati Manual (2015), §2 · Doran (1981) on SMART goals.