How to read a user’s mind?
Designing algorithms for contextual recommendations
Bharath Mohan
CEO, Sensara.tv
Google Now Airport Card
You are taking a flight
What is known? How is it known?
The flight starts at 6:45 PM
Airport is 45 minutes away
Your email
Flight database
GPS + Nav + Traffic
Google Now Currency Card
You are in an airport
What is known? How is it known?
A foreign country
A country that speaks a different
language
GPS
Different from your home
You’ve never spoken that
language
Foursquare Notification
You are near a restaurant
What is known? How is it known?
You like this cuisine
It is lunch time
GPS
Based on past outings
Clock
Sensy Recommendations
You are near a TV
What is known? How is it known?
You like English Action
You have someone with you
Same WiFi as TV
Based on past views
Her phone in same WiFi
Watch with Vidya
Channels
Your friend likes English Action Based on past views
The Context Engine
What if this super-smart engine guesses exactly what you need, in the context you
are in, and gives you something useful?
Understanding context
Reason about a person by modelling his
past and future
Analyse context around every dimension
- location, meetings, actions, time, etc
Some thumb rules for
recommendations
If context is novel, recommend the popular
If context is routine, recommend the novel
If context is novel, recommend the popular
At 1PM, a near-by and popular vegetarian restaurant in Mumbai
normally eats
actionable
out of place - home
in Bangalore
best recommendation
generally visits vegetarian
If context is routine, recommend the novel
At home at 8 PM, an action movie that’s airing for the first time on TV
normally watches TV
generally watches action
most novel
Life as a Context
Engine
Context engines have to be there with you, when
you are doing things.
#1
The curse of any recommender system is that a user never asks for one.
Context engines have to neatly fold into the experience of something you are doing already.
Context engines have to be interesting and precise.
#2
Do you recall Clippy? That annoying personal assistant on MS Office, that’d popup and tell you the obvious.
Context engines cannot make errors. Even a right thing, if told at a wrong time is annoying.
Context engines are personal, and should grow closer.
#3
I may love Mediterranean food, but only during lunch time. You may love Greek, but only on weekends.
A personal assistant that gathers your trust, must grow on it – on continued usage.
It should offer explanations, ask for feedback and constantly learn and react.
Personal assistants are
long term companions.
#4
Its like marriage.
Personal assistants have to find that sweet spot where users will continue to have
them even after the honeymoon phase.

How to read a user's mind

  • 1.
    How to reada user’s mind? Designing algorithms for contextual recommendations Bharath Mohan CEO, Sensara.tv
  • 2.
    Google Now AirportCard You are taking a flight What is known? How is it known? The flight starts at 6:45 PM Airport is 45 minutes away Your email Flight database GPS + Nav + Traffic
  • 3.
    Google Now CurrencyCard You are in an airport What is known? How is it known? A foreign country A country that speaks a different language GPS Different from your home You’ve never spoken that language
  • 4.
    Foursquare Notification You arenear a restaurant What is known? How is it known? You like this cuisine It is lunch time GPS Based on past outings Clock
  • 5.
    Sensy Recommendations You arenear a TV What is known? How is it known? You like English Action You have someone with you Same WiFi as TV Based on past views Her phone in same WiFi Watch with Vidya Channels Your friend likes English Action Based on past views
  • 6.
    The Context Engine Whatif this super-smart engine guesses exactly what you need, in the context you are in, and gives you something useful?
  • 7.
    Understanding context Reason abouta person by modelling his past and future Analyse context around every dimension - location, meetings, actions, time, etc
  • 8.
    Some thumb rulesfor recommendations If context is novel, recommend the popular If context is routine, recommend the novel
  • 9.
    If context isnovel, recommend the popular At 1PM, a near-by and popular vegetarian restaurant in Mumbai normally eats actionable out of place - home in Bangalore best recommendation generally visits vegetarian
  • 10.
    If context isroutine, recommend the novel At home at 8 PM, an action movie that’s airing for the first time on TV normally watches TV generally watches action most novel
  • 11.
    Life as aContext Engine
  • 12.
    Context engines haveto be there with you, when you are doing things. #1 The curse of any recommender system is that a user never asks for one. Context engines have to neatly fold into the experience of something you are doing already.
  • 13.
    Context engines haveto be interesting and precise. #2 Do you recall Clippy? That annoying personal assistant on MS Office, that’d popup and tell you the obvious. Context engines cannot make errors. Even a right thing, if told at a wrong time is annoying.
  • 14.
    Context engines arepersonal, and should grow closer. #3 I may love Mediterranean food, but only during lunch time. You may love Greek, but only on weekends. A personal assistant that gathers your trust, must grow on it – on continued usage. It should offer explanations, ask for feedback and constantly learn and react.
  • 15.
    Personal assistants are longterm companions. #4 Its like marriage. Personal assistants have to find that sweet spot where users will continue to have them even after the honeymoon phase.