Describing Patterns and Disruptions in Large Scale Mobile App Usage Data
4 likes1,777 views
The document analyzes patterns and disruptions in mobile app usage, highlighting the impact of major events such as Euro 2016, Brexit, and New Year's Day on user engagement. It discusses demographic correlations and distinct user types, finding that app usage follows predictable daily and weekly patterns but can experience significant disruptions during notable events. The authors propose further research to detect events based on disruptions and to explore user profiling in relation to these changes.
Describing Patterns and Disruptions in Large Scale Mobile App Usage Data
- 1. Describing Patterns and Disruptions in Large Scale Mobile App Usage Data Steven van Canneyt*, Marc Bron, Mounia Lalmas and Andy Haines *Work carried out as part of Steven Canneyt internship at Yahoo London
- 2. This work § Why is it important to understand patterns of app usage? § What is known about patterns of app usage? § What we did ● Disruption of app usage behaviour through major sport (Euro 2016), political (Brexit) and social (new year day)
- 3. Why is it important to understand patterns of app usage? ● Increasing usage of mobile devices and mobile applications (apps) ● Emergence of online marketplaces and APIs à developers, market intermediaries & consumers develop, disseminate, and use apps ● Advertising industry wants to improve targeting and experience with apps ● Marketplace operators want to identify popular or problematic apps à provide effective app recommender systems ● Developers want to understand why their apps are liked or disliked à improve app design ● Insights for Yahoo London Ad Sales
- 4. What is known about patterns of app usage? ● Relationship between demographics and app usage ● Identify distinct types of users based on their app usage (e.g. evening learners, screen checkers, game addicts) ● Simple features app category, time of day, workday versus weekend (Malmi & Weber, 2016; Zhao etal, 2016)
- 5. Teenager app usage (UK) 13-17: peak in morning 13-17: drop during day school! 13-17: increase in evening Weekday Percentage of app sessions per 15-minute window by age
- 6. Teenager app usage (UK) order ~ age young people wake up later? Weekend Percentage of app sessions per 15-minute window by age
- 7. General patterns of app usage • Average number of visited app per week and per user is 7 with 5 app categories within a day social network > search > e-commerce • Users interact between 10 to 200 times a day on average with session length between 10-250 seconds • Mostly short sessions with 80% of apps used ≤ 2 minutes • Overwhelmingly only one app per session • App usage often focused at specific time (news app in morning) (Yang etal, 2015; Falaki etal, 2010)
- 8. Diurnal patterns § Different diurnal patterns for different categories of apps ● News apps in early morning ● Sports apps in evening ● Games apps peak after standard work hours § App usage varies during the day ● grow from 6am to first peak around 11am, then declining slightly between 11am to 12pm ● 32% of app usage performed during 7pm to 11pm, reaching maximum around 9pm, then decline reaching minimum around 5am à consistent with human habit (Xu etal, 2013; Li etal, 2015)
- 9. What is known about patterns of mobile app usage? ● App usage follows regular patterns, in terms of which app, and when during the day or the weekday they are mostly used ● So what about cases when these patterns are disrupted?
- 11. Flurry Library that mobile developers to integrate in their apps to measure app usage and allow in-app advertising default app events triggered by user actions (app start event) custom-based events Popularity based engagement metric number of sessions a user has with an app based on the app start event Sample data May 2016 230K mobile apps 600M daily unique users US and UK https://developer.yahoo.com/analytics/
- 12. App categories 27 categories ranging from work related (productivity) to leisure (games) and other popular categories (news)
- 13. Engagement patterns in app usage
- 14. General daily engagement patterns (US) weekday: peak during morning Users active later during weekend than during week weekend: stable during day Similar patterns reported in other studies → no or little bias from Flurry inventory
- 15. Daily engagement patterns by category (US) Similar patterns reported in other studies → no or little bias from Flurry inventory
- 16. App engagement patterns per day of the week and category (US) Week: productivity Weekend: sports, entertainment Fridays and Saturdays: shopping Similar patterns reported in other studies → no or little bias from Flurry inventory
- 17. Disruptions in engagement patterns in app usage in major UPCOMING KNOWN events SPORT: EURO 2016 POLITICAL: BREXIT SOCIAL: NEW YEAR DAY
- 18. Data processing and measurement • Target day and reference days If target event occurs on Saturday then take a number of Saturdays before event • Remove outliers from reference days (outage, new app release, other major event) number of start session events either ≤ 1st quartile – 1.5 or ≥ 3rd quartile + 1.5 • Day divided into time segments (e.g., 15 minutes) and normalize • avgt: expected number of sessions per time segment t estimated by averaging normalized number of sessions for reference days • stdt: standard deviation for reference days • Behaviour “significantly” disrupted: normalized number of sessions during target period ≥ avgt + 2·stdt or ≤ avgt − 2·stdt
- 19. Case study 1: EURO 2016
- 20. Euro 2016: The Data (UK) 16M viewers (25% of UK population) watched Portugal beat France in the final on BBC 1 Typical mobile engagement for match played on Saturday as average engagement of all Saturdays between November 2015 & June 2016 Same process used to model typical app engagement on reference days counterpart to each of the match days Each event day has 30 reference days
- 21. Number of sessions during games (UK) England 2 – 1 Wales 14:00 Slovakia 0 – 0 England 20:00 Portugal 1 – 0 France 20:00 ● Green bars: average same weekday based on 30 weeks before EURO 2016 ● Blue lines: 2 x std ● Green dots: similar as average ● Yellow dots: < avg - std or > avg + std ● Red dots: < avg - 2x std or > avg + 2 x std England 1 – 1 Russia 20:00 Sports apps: x3.7! half time half time half time half time
- 22. England 1 –2 Iceland 20:00 before game half time App engagement during Euro 2016 games • app engagement during games not lower than during same time on an average day for any of the matches • during half-time app engagement is significantly higher than during same time on an average day for England – Wales and England – Slovakia games BUT: after game
- 23. Case study 2: BREXIT
- 24. BREXIT (UK) European Union membership referendum – Brexit – took place on Thursday June 23, 2016 in the UK to gauge support for the country’s continued membership of the European Union § Study whether outcome coincides with disruptions in app usage § Reference days are all weekdays in June before June 24 § Top 10 app categories with largest percentage change in session volume compared to average usage Result was announced in early morning of June 24, 2016: overall vote to leave the EU by 51.9% on a national turnout of 72% unstability in financial markets & turmoil in UK political landscape
- 25. Day of the referendum result 24 June, 2016 (UK) The UK in shock The pound crashing app engagement increase by 114% for finance 43% for news
- 26. First week after the referendum result (UK) The UK slowly calming down but still concerned?
- 27. Second week after the referendum result (UK) The UK calming down further
- 28. Third week after the referendum result (UK) Wimbledon final (Andy Murray won)
- 29. Fourth week after the referendum (UK) Took 4 weeks for UK to be back to normal?
- 30. Percentage of UK sessions of finance apps during June- August 2016 Lower usage on weekends EU referendum result (24 June) Back to normal after 4 weeks
- 31. Case study 3: New year’s day
- 32. New year day (US) • New Year Day, first day of the new year, observed in most Western countries on January 1. • Common traditions include attending parties, making resolutions for the new year, watching fireworks displays and calling friends and family Examine whether New Year Day coincides with disruptions in app usage patterns Week days between December 15, 2015 & January 15, 2016, without January 1, used as reference days
- 33. Percentage of sessions for the 10 categories with the largest percentage change in app engagement Users take photos on New Year day Increased use of social media
- 34. Conclusions and what next
- 35. Some final thoughts § We are able to detect disruption of app engagement patterns § A tool to judge people habit, mood, interest, concern, etc What next § We want to look at country difference § Automatically detect events based on disruptions § Profile users based on disruptions § Study of “mobile addiction”