Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
On AB Testing
Search
Hector Zarate
September 16, 2016
Technology
2
420
On AB Testing
My slides on my presentation on AB Testing.
As presented in NSSpain 2016.
Logroño, Spain.
Hector Zarate
September 16, 2016
Tweet
Share
More Decks by Hector Zarate
See All by Hector Zarate
On Debugging
ioscowboy
0
540
Inheritance, Interfaces and Composition
ioscowboy
0
460
iOS @ Spotify
ioscowboy
0
130
Consistent, Dumb and Thin
ioscowboy
3
16k
Other Decks in Technology
See All in Technology
Fargateを使った研修の話
takesection
0
140
顧客が本当に必要だったもの - パフォーマンス改善編 / Make what is needed
soudai
24
6.8k
IaC運用を楽にするためにCDK Pipelinesを導入したけど、思い通りにいかなかった話
smt7174
1
120
Gradle: The Build System That Loves To Hate You
aurimas
2
150
【若手エンジニア応援LT会】AWS Security Hubの活用に苦労した話
kazushi_ohata
0
170
来年もre:Invent2024 に行きたいあなたへ - “集中”と“つながり”で楽しむ -
ny7760
0
480
小規模に始めるデータメッシュとデータガバナンスの実践
kimujun
3
600
Autify Company Deck
autifyhq
1
39k
使えそうで使われないCloudHSM
maikamibayashi
0
170
CyberAgent 生成AI Deep Dive with Amazon Web Services / genai-aws
cyberagentdevelopers
PRO
1
480
新卒1年目が向き合う生成AI事業の開発を加速させる技術選定 / ai-web-launcher
cyberagentdevelopers
PRO
7
1.5k
「視座」の上げ方が成人発達理論にわかりやすくまとまってた / think_ perspective_hidden_dimensions
shuzon
2
6.6k
Featured
See All Featured
Intergalactic Javascript Robots from Outer Space
tanoku
268
27k
Music & Morning Musume
bryan
46
6.1k
Teambox: Starting and Learning
jrom
132
8.7k
Save Time (by Creating Custom Rails Generators)
garrettdimon
PRO
27
800
[Rails World 2023 - Day 1 Closing Keynote] - The Magic of Rails
eileencodes
32
1.8k
XXLCSS - How to scale CSS and keep your sanity
sugarenia
246
1.3M
Designing for Performance
lara
604
68k
Principles of Awesome APIs and How to Build Them.
keavy
126
17k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
26
1.4k
個人開発の失敗を避けるイケてる考え方 / tips for indie hackers
panda_program
92
16k
Easily Structure & Communicate Ideas using Wireframe
afnizarnur
191
16k
Why Our Code Smells
bkeepers
PRO
334
57k
Transcript
AB Testing Hector Zarate @ChocoChipset
AB Testing Hector Zarate @ChocoChipset
Hector Zarate (@ChocoChipset) Software Engineer at Spotify Boxun Zhang Data
Scientist at Spotify
1. What?
None
None
YOU ARE PART OF AN EXPERIMENT
AB Testing
None
Samson, 3
Max, 30 Samson, 3 +5% matches
Share of Matches 20% 23.75% 27.5% 31.25% 35% Week 0
Week 1 Week 2 Week 3 Week 4 With Sunglasses Sunglasses
Share of Matches 20% 23.75% 27.5% 31.25% 35% Week 0
Week 1 Week 2 Week 3 Week 4 With Sunglasses Control Sunglasses
Share of Premium Conversion 20% 23.75% 27.5% 31.25% 35% Week
0 Week 1 Week 2 Week 3 Week 4 Variation A Control Sunglasses
User Retention 20% 23.75% 27.5% 31.25% 35% Week 0 Week
1 Week 2 Week 3 Week 4 Variation A Control Sunglasses
None
spotify:user:chocochipset 92b2976bb15d26c9008
1 2 3 4 5 6 7 8 9 10
p ( x ) x 1 / n
1 2 3 4 5 6 7 8 9 10
p ( x ) x 1 / n
1 2 3 4 5 6 7 8 9 10
p ( x ) x 1 / n 0% 60%
Markets
Demographics
User Attributes
None
Login resolveABFlags() response(ABFlags) ABBA Cache AB Values Load Cached or
Default AB Flags
Example button.color = [UIColor spotifyGreen]; BOOL isButtonPink = ([abFlags[@“pink-buttons"] isEqual:@"1"]);
if (isButtonPink) { // alternate path here: button.color = [UIColor spotifyPink]; } Key Value charts Enabled pink-buttons 0 buffer-quality low gallery-artist Control
Key Value charts Enabled pink-buttons 0 buffer-quality low gallery-artist Control
Example button.color = [UIColor spotifyGreen]; BOOL isButtonPink = ([abFlags[@“pink-buttons"] isEqual:@"1"]); if (isButtonPink) { // alternate path here: button.color = [UIColor spotifyPink]; } Key Value charts Enabled pink-buttons 1 buffer-quality low gallery-artist Control pink-button 1 Variation A Control
Control Variation
Analytics impressions and interactions
* just a personal preference
2. How?
1. Hypothesis 2. Design 3. Run 4. Analysis
1. Formulate a Hypothesis
2. Design the Test
# shuffle plays 2.1 Target Metrics
# matches Samson, 3 # replies # walks in the
park
2.2 Test Group more test bandwith, less confidence, smaller effect
in business less test bandwith, more confidence, bigger effect in business Small Large
2.3 Duration
3. Run the test
1 2 3 4 5 6 7 8 9 10
p ( x ) x 1 / n Test A Test A Test B Test B Test A / Test B Watch out for conflicting tests!
Don’t cut them short! Bruno Cesar 48’ Ronaldo 89’ Morata
90’ + 4’ 2 -1
0 25 50 75 100 D1 D2 D3 D4 D5
D6 D7 D8 Don’t cut them short!
4. Analyze the results
1. Formulate a Hypothesis new
3. Case Studies
3.1. Losing Calories
None
None
None
None
None
None
Test: Tab Bar Navigation Hypothesis: By switching to TBN, we
expect an increase in the share of users who click at least one menu item. 1
Test: Tab Bar Navigation Hypothesis: By switching to TBN, we
expect an increase in the share of users who click at least one menu item. 2.1
10% New Users 1% Existing Users 2.2
2.3 8 day period
4 Clicks on Menu Items 10% 20% 30% 40% 50%
60% 70% 80% 90% 100% 110% 120% 130% Control Tab Bar 130% 100%
4 Clicks Overall 10% 20% 30% 40% 50% 60% 70%
80% 90% 100% 110% 120% 130% Control Tab Bar 109% 100%
4 1. Increased clicks in Tab Items 2. Decreased clicks
in non-tab items.
WHY AB TEST? QUANTIFY THE IMPACT OF A SPECIFIC CHANGE
3.2. Home in Mexico
My Rock Your Rock
My Home Your Home
My Home Your Home
Test: Home Hypothesis: We will measure an increased second week
retention by using the new ‘Home’ as start page. 1
Test: Home Hypothesis: We will measure an increased second week
retention by using the new ‘Home’ as start page. 2.1
US UK Germany Austria Mexico 4
1. Is the localization good? 2. Are recommendations relevant for
the market? 3. Technical restrictions we are not aware of? 4
12.34 Mbps US 13.70 Mbps UK 13.42 Mbps Germany 15.48
Mbps Austria 7.4 Mbps Mexico * 3.7 Mb 4
1. Hypothesis 2. Design 3. Run 4. Analysis
Test: Home Hypothesis: By compressing the data sent for Home,
more users will play on day one, two and be active during their second week. 1
4. Pitfalls
NO DATA
Max, 30 TOO MANY CHANGES
Max, WOLF AT WALL STREET TOO MANY CHANGES
DONT LISTEN TO YOUR HEART (LISTEN TO THE DATA)
5. Wrap Up
None
None
None
None
None
KNOWLEDGE IS YOUR ROI
A test is an investment Design to maximize learning
NEGATIVE RESULTS ARE STILL POSITIVE
TRY AB TESTING
None
Gracias! Hector Zarate @ChocoChipset spotify.com/jobs
Preguntas! Hector Zarate @ChocoChipset spotify.com/jobs