COMP322 · Human computer interaction

Evaluation

How do we know it is any good?

Usability testing, the methods around it, and a couple of laws that let you judge a design before anyone touches it.

The problem with your own design

You cannot see your own interface

Once you know where the button is, you can never again experience not knowing. This is the curse of knowledge, and it is why self review fails.

The designer

Knows the menu structure, the labels, the hidden gesture. Everything is obvious. Of course it is.

The new user

Sees a wall of choices, no memory of your intentions, and a goal that has nothing to do with your architecture.

The gap

Evaluation exists to measure the distance between those two people. You cannot guess it. You have to look.

The working assumption

You are not the user. You will never be the user again. Test with someone who is.

A ballot layout that changed an election

Left column Candidate A Candidate B Candidate C Right column Candidate D Candidate E 1 2 3 4 5 Correct: B is the 2nd name, but its hole is the 3rd. Instinct: punch the 2nd hole, and you vote for D.
  • Names ran down two facing columns with a single row of punch holes between them, so the second name on the left aligned with the third hole.
  • Voters reaching for the second name punched the second hole and recorded a vote they never intended.
  • Over 19,000 ballots in that county were spoiled by double punching.
Why it belongs in this lecture

The certified statewide margin was 537 votes. This was not a software bug, it was a mapping failure in a paper layout that no one tested with ordinary voters. Recall Norman: when action and result do not line up, people make errors that are not their fault.

The three hundred million dollar button

$300m/ year
extra revenue, reportedly, from removing one barrier that testing exposed
  • A large retailer forced shoppers to register before buying. The checkout showed a Register button and a login form.
  • In testing, first time buyers were furious. They did not want a relationship, they wanted shoes. Returning customers could not remember if they had an account.
  • The team replaced Register with a single Continue button and a quiet line: you do not need an account to buy.
  • Purchases jumped immediately. One button, found by watching people, not by arguing in a meeting.
Why it matters

Nobody on the team had predicted this. They were too close to it. A handful of real sessions found money that years of internal debate had missed.

The same root cause, decades apart

Two more failures, one shared cause

Both systems were shipped without watching real people under real conditions. Both broke down in public, at the worst possible moment.

Three Mile Island · 1979

A control panel showed that a relief valve had been commanded shut, not that it was physically stuck open. Operators trusted the light while coolant drained away. Meanwhile hundreds of alarms lit at once with no priority. A feedback and mapping failure in a nuclear control room.

Healthcare.gov · 2013

The United States health insurance exchange launched with almost no end to end usability or load testing. On the first day a mere handful of people managed to enrol before the site collapsed under demand. Evaluation skipped at the scale where it mattered most.

The pattern

Different decades, different domains, one lesson. A system never tested with real people under real load will still find its problems. It just finds them in front of the public instead of in a quiet room with five users.

Before you test anything

Plan it: the DECIDE framework

Evaluation is not "grab a friend and poke the app". It is planned. Six questions, in order.

D · Determine

What are the goals of this evaluation? What do you actually need to learn?

E · Explore

What specific questions will answer those goals?

C · Choose

Which approach and methods fit? Lab, field, or inspection?

I · Identify

Practical issues: who, where, what kit, how long, how many?

D · Decide

How will you handle ethics? Consent, privacy, the right to stop.

E · Evaluate

Collect, analyse, interpret and present the data honestly.

Use it as a checklist

Most weak evaluations skip D and E at the front. They run sessions without deciding what a useful answer would even look like, then drown in notes.

The map of methods

Three ways to evaluate, one question each

With users · controlled

Usability testing

Set tasks, watch closely, in a quiet room. You control the conditions.

Precise, slightly artificial.
With users · natural

Field studies

Watch real use in the real setting, on a bus, in a shop, at 2am.

Realistic, hard to control.
Without users

Inspection and models

Experts judge against principles. Or you predict performance with a formula.

Fast and cheap, blind to surprises.
Where this lecture goes

We spend most time on usability testing, because it is the workhorse. Then we end with the without users methods, because being able to judge a design with no people in the room is a quiet superpower.

Where these tasks come from, and where they lead

One evaluation is a chain, not a single act

Each step exists because the one before it produced something the next step needs. Nothing here is busywork.

1

Goal

DECIDE: what must we learn?

2

Tasks

Turn the goal into real user goals

3

Session

Observe, think aloud, stay quiet

4

Measures

Completion, time, errors, SUS

5

Findings

Themes and severity ratings

6

Revise

Fix, then test the fix

Step 6 feeds step 2 again. Evaluation is a loop, not a line, and that loop is exactly what your Phase 3 must show.
Where each method attaches across the project
Discover
Interviews, personas
Design
Heuristic inspection, Fitts models
no users
Prototype
Usability testing
a few users
Live
Analytics, A/B
many users

The workhorse method

A usability test is just structured watching

Put a representative person in front of the real thing, give them a real goal, and watch without helping.

What you do
  • Recruit people who match your users
  • Give them realistic tasks, not a tour
  • Ask them to think aloud
  • Observe, time, count, stay quiet
What you do not do
  • Demo the feature you are proud of
  • Explain where the button is
  • Ask "wouldn't it be nice if..."
  • Argue with the participant
The mindset

You are not selling the design and you are not defending it. You are collecting evidence about it, and the most useful evidence is usually the part that stings.

Mathematical model + replications

Why five users is usually enough

  • Nielsen and Landauer modelled how problems are found. Each user reveals roughly the same fraction of issues, so discoveries pile up fast then flatten.
  • With about five users you have already seen around 85 percent of the problems. The sixth mostly repeats what you saw.
  • So run small, fix, and test again. Three rounds of five beats one round of fifteen.
~85%of issues
found by the fifth participant, in the classic model
The honest caveat

The 85 per cent assumes each user finds about a third of the problems, which holds for a focused interface but not for a sprawling, varied one. To measure a number precisely you also need far more people. The argument over this single figure has run for thirty years.

The skill nobody practises

Give a goal, not a recipe

The wording of a task decides whether you learn anything. Tell people what to achieve, never how.

Leading · tells them how

"Click the Registration menu at the top, choose Add Course, then type COMP322 in the search box."

You just tested whether they can follow instructions. Useless.
Goal · gives them a reason

"It is registration week. Add COMP322 to your schedule for a section that does not clash with your morning lab."

Now you see where the real journey breaks.
Why the framing matters

A good task has a context and a clear finish line, so both you and the participant know when they are done. It should be something a real person would actually want to do.

Hearing the reasoning, not just the clicks

Think aloud: narrate the struggle

Ask participants to say what they are thinking as they work. Clicks tell you what happened. Words tell you why.

What it sounds like

"I am looking for register... maybe under this menu... no, that is settings... wait, why is there a second login box... I will just try this one."

How to prompt
  • "Keep talking, what are you thinking now?"
  • Stay neutral, never "is it confusing?"
  • When they go silent, gently nudge
It is genuinely unnatural

Nobody narrates their own thoughts in daily life, so people forget. Your job is to keep the stream running without putting words in their mouth.

The hardest job in the room

The facilitator's only real skill: silence

Every word you add contaminates the data. The instinct to rescue a struggling user is the instinct you must defeat.

Do not lead

"What would you do here?" beats "would you click that button?". The second hands them the answer.

Do not rescue

If they are stuck, that is a finding. Wait. Let the silence stretch. The struggle is the result you came for.

Do not react

No wincing when they miss it, no smiling when they get it. Your face is feedback too.

The bias to name

People want to please you. They will praise a design to your face and abandon it the moment you leave. Neutral questions and a neutral face are how you get the truth instead of the compliment.

Watch one session

A test replay, narrated

ritaj.birzeit.edu / registration
Welcome back
Results
COMP322 Human computer interaction
COMP333 Software engineering
done
COMP322 added to your schedule
Task: add COMP322 to your schedule
0:00
time
0
wrong turns
running
status
Press play to watch the participant think aloud.

Numbers and stories together

Measure what happened, and why

Quantitative data tells you that something is wrong. Qualitative data tells you why. You need both, or you are guessing.

Quantitative · the what
  • Task completion rate, the headline number
  • Time on task, how long the journey took
  • Errors, wrong turns and recoveries
  • SUS, a single satisfaction score
Qualitative · the why
  • Think aloud quotes at the moment of failure
  • Where they hesitated, sighed, backtracked
  • The mental model they expected
  • What they said they would do next time
The trap to avoid

A completion rate with no quotes is a thermometer with no diagnosis. You know there is a fever. You have no idea what to treat.

One number for satisfaction

The System Usability Scale, and its scoring trick

Ten statements, each rated 1 to 5. The wording alternates on purpose, odd items positive and even items negative, so agreeing means something different each time. Before you can add the answers, every one has to land on the same 0 to 4 scale where higher always means better.

Odd items · positive wording

"I would like to use this often." Agreeing is good, so the top answer 5 must become the best score. Subtract 1, and raw 1 to 5 becomes 0 to 4.

score = raw − 1
Even items · negative wording

"I found it unnecessarily complex." Agreeing is bad, so the top answer 5 must become the worst score. Subtract from 5, and the scale flips, 1 to 5 becomes 4 to 0.

score = 5 − raw
Raw answerOdd: raw − 1Even: 5 − raw
104
213
322
431
540
Both columns cover 0 to 4, they simply run in opposite directions, because a 5 means the opposite thing on a positive versus a negative statement.
The mnemonic

Odd, minus one. Even, from five. Subtracting from 5 reverses the scale, so disagreeing with a negative statement correctly counts as a good thing rather than a bad one.

From ten answers to one number

SUS scored: a worked example

One participant's ten answers, converted item by item with the two rules, then summed and scaled. Synthetic example, participant P2.

ItemRawRuleScore
1Would use it often44 − 13
2Too complex25 − 23
3Easy to use44 − 13
4Needs tech help15 − 14
5Well integrated44 − 13
ItemRawRuleScore
6Too inconsistent25 − 23
7Quick to learn33 − 12
8Cumbersome25 − 23
9Felt confident44 − 13
10Lots to learn first35 − 32
Sum of the ten converted scores 29
Then 29 × 2.5 = 72.5
--
press score
Field A/B results

A/B testing, and its famous limit

Show version A to half your users and version B to the other half, then measure which performs better. At the scale of millions you can detect tiny effects, but you only learn which option wins, never why.

Done well · Obama campaign, 2008

The team tested splash page variants. A "Learn more" button paired with a family photo beat the original by about 40 per cent more sign ups, an estimated 2.8 million extra subscribers and tens of millions in donations.

Taken too far · Google, 2009

A debate over a link colour was settled by testing 41 shades of blue. A senior designer resigned, arguing that data had replaced design judgement entirely.

The limit of pure numbers

A/B testing picks the better of two options you already have. It cannot tell you that a third, better idea exists, and it cannot explain behaviour. It climbs the hill you are standing on, never asking whether you are on the right hill.

Established law · Paul Fitts, 1954

Fitts's law, made interactive

Drag the sliders and the predicted time and the cursor obey the law. Pointing time depends on just how far the target is and how wide it is.

D · distance W · width
MT = a + b · log2( 2D / W )
Difficulty 2.00 bitsPredicted 350 ms
Double the distance DDifficulty rises by one bit, so the move takes clearly longer.
Halve the width WAlso exactly one bit harder, the same slowing as doubling the distance.
Double both at onceThe two effects cancel, so difficulty and time stay the same.

Fitts's law, applied

Edges and corners are infinite targets

  • You cannot overshoot a screen edge. Slam the cursor up and it stops at the wall, so the target is effectively infinitely tall.
  • That is why the Mac menu bar sits glued to the top edge. You flick upward without aiming and you are there.
  • A corner is infinite in two directions, the easiest place on the whole screen to hit. Hot corners and the Start button live there for a reason.
  • A right click menu appears at the cursor, distance near zero, the fastest target of all.
Feel it yourself

Click the two targets in turn. Watch which one your hand finds faster.

last move: --
Design payoff

Put your most important, most frequent controls at edges and corners, or right under the cursor. The law was published in 1954 and your operating system is quietly obeying it right now.

Non negotiable

You are studying people, so behave like it

Every method in this lecture involves a real human giving you their time and sometimes their frustration. Respect comes first, data second.

Before
  • Informed consent, signed, before you start
  • Explain the purpose and that they can stop anytime
  • Ask permission before recording
During and after
  • Test the system, never the person
  • Anonymise: P1, P2, no names or ID numbers
  • Store data safely, delete it when done
The phrase that protects everyone

Open every session with it: "We are testing the design, not you. If something is confusing, that is the design's fault, not yours." It relaxes the participant and gives you honest behaviour.

No method is enough alone

Triangulate, because each method is half blind

Inspection

Heuristic evaluation, fast and cheap, no users needed.

Misses the surprises only real people produce.
Testing

Real users, real tasks, the unexpected made visible.

Slow, small samples, says little about scale.
Analytics and models

Millions of data points, or a formula with none.

Tells you what, almost never why.
The professional move

Inspect early to catch the obvious for free. Test with five users to find the real surprises. Use analytics to confirm at scale. Each method covers another's blind spot. That is triangulation, and it is what separates a real evaluation from a vibe check.

The lower bar you should clear today

Watching one user beats watching none

1user
is infinitely more than zero. The first person you watch teaches you the most.
  • You do not need a lab, a one way mirror, or special software. A laptop, a real person, and a quiet corner is a usability test.
  • Steve Krug calls it "rocket surgery made easy". The barrier is mostly in your head.
  • Test a little, early, often. A rough prototype tested this week beats a perfect one tested never.
The habit to build

The teams that improve fastest are not the ones with the best first design. They are the ones that put their work in front of a real person soonest and were willing to be wrong.

Where this lands for you

Your Phase 3, mapped to today

Everything in this lecture is the toolkit for the Phase 3 deliverable, due 24 June. Match each rubric line to a method you now have.

The test itself
  • Three participants, distinct from your Phase 1 interviewees
  • Goal based tasks from your Phase 1 scenarios
  • Think aloud, signed consent, anonymised
The evidence
  • Completion, time on task, error counts
  • SUS score averaged across participants
  • Themed quotes, plus a before and after heuristic comparison
The point of the phase

Phase 3 is not about proving your design was perfect. It is about showing you can close the loop: test honestly, find the problems, and say clearly what you would change next. Iteration is the grade.

The thread through today

Evaluation is humility,

made into a method.

You cannot see your own design. So you plan a question, watch a real person, measure what happened and why, and let the evidence change your mind.

Next time

Data gathering in depth: interviews, observation, and questionnaires done properly, the raw material every method here depends on.

N notes  ·  F fullscreen  ·  arrows to move