← Research Methodology · Unit I
Unit I · Lecture 4

Data Collection, Analysis, Interpretation & Instrumentation

A research problem is only as trustworthy as the evidence behind it. This lecture is about that evidence — how you gather it, make sense of it, and the tools you need to do so honestly.

⏱️ ~20 min read 🎯 CO1 · CO2 📘 Data · Analysis · Instrumentation

🎯 By the end of this lecture you will be able to…

  • Distinguish data collection, analysis, and interpretation — and why conflating them is dangerous.
  • Choose appropriate data-collection methods and sources for a computing problem.
  • Identify the instrumentation (tools, benchmarks, metrics) a study requires.
  • Avoid the classic evidence errors: bad sampling, over-claiming, and correlation-as-causation.

In Lecture 3 you chose an approach. That approach now has to produce evidence. Three distinct activities turn a plan into a finding: collect the data, analyse it, then interpret what it means. They are not the same thing — and the gap between the last two is where most research goes wrong.

① COLLECT gather raw data logs · timings · surveys ② ANALYSE find patterns mean · p95 · tests ③ INTERPRET explain meaning ⚠ most errors hide here
Three distinct steps. Analysis tells you what the numbers say; interpretation tells you what they mean — don't skip straight to conclusions.

1Data collection

Data collection is the systematic gathering of the observations you'll reason about. In computing, "data" ranges from benchmark timings to survey responses to system logs.

Primary vs. secondary data

Primary dataSecondary data
DefinitionYou collect it yourself, first-handCollected by others, reused by you
Computing exampleLatency you measured; a survey you ranPublic datasets (ImageNet, MNIST), GitHub archives, Kaggle
ProsFits your exact question; you control qualityFast, cheap, large, comparable to others
ConsTime-consuming; needs instrumentationMay not fit; unknown biases; licensing/ethics
DATA SOURCES PRIMARY first-hand — you gather it benchmarks you ran a survey you conducted SECONDARY reused — collected by others ImageNet, MNIST Kaggle, GitHub archives
Two origins of evidence. Primary fits your exact question but costs time; secondary is fast and comparable but carries hidden biases.

Common collection methods in computing research

📌 Sampling: the quiet make-or-break decision

You rarely measure everything; you measure a sample and generalise. If the sample is biased, so is every conclusion. Testing your recommender only on power-users, or benchmarking only on tiny inputs, produces confident but wrong claims. Ask: does my sample represent the population I want to generalise to?

⚠️ Pitfall: the single-run measurement

Running a benchmark once and reporting the number ignores variance from caching, background processes, and thermal throttling. Fix: repeat runs (e.g., 30×), discard warm-up, and report a central value with its spread (mean ± std, or median + IQR).

2Instrumentation — the tools of evidence

Instrumentation means everything you use to obtain and record data reliably: the measuring apparatus of your study. In a lab it's oscilloscopes; in computing it's a stack of software and defined metrics.

Instrumentation typeExamplesUsed to capture…
Profilers & timersperf, gprof, time, JMHexecution time, CPU, memory
Benchmark suitesSPEC, TPC, MLPerf, GLUEstandardised, comparable workloads
Monitoring / loggingPrometheus, Grafana, ELKsystem behaviour over time
Survey platformsGoogle Forms, Qualtricshuman responses
Statistical / analysis toolsPython (pandas, SciPy), Ranalysis & visualisation
Defined metricsaccuracy, F1, latency p95, throughputthe quantities you actually report
✅ Tip: choose metrics before you run anything

Deciding what counts as success after seeing results invites cherry-picking ("well, it wins on this metric"). Pre-register your primary metric. For classification, know why you're using F1 vs. accuracy; for latency, report a tail percentile (p95/p99), not just the mean — averages hide the bad experiences.

🤔 Pause & think

You benchmark a web server and report "average response time = 120 ms." A reviewer says that number could hide a serious problem. What are they worried about, and which single metric would reveal it?

3Data analysis

Analysis transforms raw data into patterns: summarising, testing, and visualising. Two broad families:

💡 Analysis example

Model A averages 91.2% accuracy, Model B 91.8%. Is B really better? Descriptive stats say B is higher. Inferential analysis (e.g., a paired test across cross-validation folds) tells you whether a 0.6-point gap is statistically significant or just noise. Without the second step, you might crown a winner that isn't one.

📌 Visualise — but honestly

A good chart reveals structure a table hides (distributions, outliers, trends). A dishonest chart — truncated y-axis, missing error bars — misleads. Always show the spread of your data, not just the headline number.

4Interpretation — what does it mean?

Interpretation is where you explain the significance of your results: connecting numbers back to the research question, comparing with prior work, and stating what we now know. Analysis says "B is 0.6% higher, p = 0.03." Interpretation says "B's small but significant gain likely comes from its data augmentation, suggesting augmentation matters more than architecture for this task."

⚠️ Pitfall: correlation is not causation

Your logs show users who use feature X churn less. Does X cause retention? Maybe — or maybe already-loyal users happen to explore more features. Observational data shows association; causal claims need controlled experiments (e.g., an A/B test). Over-claiming causation is one of the most common — and most penalised — errors in student research.

CORRELATION association only Loyal-user type? (hidden) Feature X Retention a hidden cause may drive both CAUSATION proven, not assumed Feature X Retention only a controlled A/B test can establish this arrow
Correlation ≠ causation. A hidden variable can make X and Y move together; only a controlled experiment proves that X actually causes Y.
⚠️ Pitfall: over-generalising

"Our method beats the baseline on MNIST, therefore it's the best image classifier." One dataset, one domain, one setting cannot support a universal claim. Interpret within your scope (Lecture 3). Honest limits build credibility; overreach destroys it.

🤔 Your turn

Label each statement as analysis or interpretation: (a) "The median query time was 44 ms." (b) "The 3× slowdown on joins suggests the index isn't being used." (c) "Accuracy differences were not significant (p = 0.4)."

✅ Check your understanding

1. Reusing the public ImageNet dataset is primary or secondary data? Give one risk.
Secondary data (collected by others, reused by you). Risks include hidden biases in how it was labelled/collected, licensing/ethics constraints, and it may not perfectly fit your specific question.
2. Why report p95 latency instead of just the mean?
The mean can look fine while a minority of requests are catastrophically slow. p95/p99 (tail latency) exposes the worst experiences real users hit — often the numbers that matter most for a system's reputation.
3. Your A/B-free log analysis shows feature X correlates with retention. What can and cannot you claim?
You can claim an association between X and retention. You cannot claim X causes retention — that needs a controlled experiment (A/B test) to rule out confounders like user type.
4. From "Your turn": which are analysis and which is interpretation?
(a) Analysis (a descriptive statistic). (b) Interpretation (explaining why/what it means). (c) Analysis (an inferential test result). Interpretation is the one that reaches beyond the numbers to their meaning.

🧠 Key takeaways

  1. Collection → analysis → interpretation are three distinct steps; the biggest errors hide in the last one.
  2. Know your data source: primary (you gathered it) vs. secondary (reused) — and watch for sampling bias.
  3. Instrumentation = the tools and metrics that record evidence; choose metrics before running experiments.
  4. Analysis is descriptive and inferential — a higher number isn't "better" until significance says so; always show spread.
  5. Interpret within scope: never confuse correlation with causation, and never over-generalise beyond your evidence.
🎓 Unit I complete — where you're headed

You can now find a problem, judge its quality, scope it, set objectives, pick an approach, and gather trustworthy evidence. Unit II goes deeper into research methods specific to computing science — models of argument, proof methods, deduction vs. induction, and performance evaluation.

Research Methodology in CS (CSEG3060) · Unit I · Lecture 4  ·  Dr. Mohsin Furkh Dar · UPES Dehradun