The New Era Of Autonomous Search: Inferring Intent On Mobile Devices

[Google CEO Eric] Schmidt said that while the Google Instant predictive search technology helps shave an average of 2 seconds off users’ queries, the next step is “autonomous search.” This means Google will conduct searches for users without them having to manually conduct searches. As an example, Schmidt said he could be walking down the streets of San Francisco and receive information about the places around him on his mobile phone without having to click any buttons. “Think of it as a serendipity engine,” Schmidt said. “Think of it as a new way of thinking about traditional text search where you don’t even have to type.”  – eWeek

When users type phrases into Google, they are searching, but also expressing intent. To create the “serendipity engine” that Eric Schmidt envisions would require a system that infers users’ intentions.

Here are some of the input signals a mobile device could use to infer intent.


Location: It is helpful to break location down into layers, from the most concrete to the most abstract:

1) Latitude and Longitude – raw GPS coordinates

2) Venue – mapping of lat / long coordinates to a venue.

3) Venue Relationship to User – is the user at home, at a friend’s house, at work, in her home city etc.

4) User Movement – locations the user has visited recently.

5) Inferred User Activity – if the user is at work during a weekday, she more likely in the midst of work. If she is walking around a shopping district on a Sunday away from her home city, she is more likely to want to buy something. If she is outside, close to home, and going to multiple locations, she more likely to be running erands.

Weather: during inclement weather user is less likely to want to move far and more likely to prefer indoor activities.

Time of day and date: around mealtimes the user is more likely to be considering what to eat. On weekends the user is more likely to be doing non-work activities. Outside at night, the user is more likely to be looking for bar/club/movie etc. Time of days also let’s you know what venues are open & closed.

News events near the user: they are at the pro sporting event, an accident happened nearby, etc.

Things around the user: knowing not just venues, but activities (soccer game), inventories (Madden 2011 is in stock at BestBuy across the street), events (concert you might like is nearby), etc.

These are just a few of the contextual signals that could be included as input signals.


The more you know about users’ tastes the more likely you can infer their intent. It is silly to suggest a great Sushi restaurant to someone who dislikes Sushi. At Hunch we model taste with a giant matrix. One axis is every known user (the system is agnostic about which ID system – it could be Facebook, Twitter, a mobile device, etc), the other axis is things, defined very broadly: product, person, place, activity, tag etc.  In the cells of the matrix are either the known or predicted affinity between the person and thing.  (Hunch’s matrix currently has about 500M people, 700M items, and 50B known affinity points).

Past expressed intent

  • App actions:  e.g. user just opened Yelp, so is probably looking for a place to go.
  • Past search actions: user’s recent (desktop & mobile) web searches could be indications of later intent.
  • Past “saved for later” actions:  user explicitly saved something for later e.g. using Foursquare’s “to do” functionality.

behaviour of other people

  • Friends:  The fact that a users friends are all gathered nearby might make her want to join them.
  • Tastemates: That someone with similar tastes just performed some actions suggests the user is more likely to want to perform the same actions.
  • Crowds: The user might prefer to go toward or avoid crowds, depending on mood and taste.

How should an algorithm weight all these signals? It is difficult to imagine this being done effectively anyway except empirically through a feedback loop. So the system suggests some intent, the user gives feedback, and then the system learns by adjusting signal weightings and gets smarter.  

With a machine learning system like this it is usually impossible to get to 100% accuracy, so the system would need a “fault tolerant” UI. For example, pushing suggestions through modal dialogs could get very annoying without 100% accuracy, whereas making suggestions when the user opens an application or through subtle push alerts could be non-annoying and useful.