About our Science
“We have been thinking deeply about these problems for a very long time.” – David Rosenberg, Chief Scientist
Sense Networks has a long history of understanding people and places from location data. Our founders started the company with an approach grounded in scientific inquiry and academic rigor that continues today within our science team.
The Science of Place
One of Sense Networks’ original messages was that you can’t really understand a person from where they go until you attach some “semantic” meaning to the places. In other words, raw latitudes and longitudes aren’t very useful, home neighborhood is somewhat useful, but knowing that East Village and West Village are rather similar, and each quite different from Upper West Side, can be quite useful for understanding similarities and differences between people.
Our earlier focus on understanding the “DNA” of places uncovered a wealth of learning. For example, one can understand a good deal from simply looking at taxi rides between places.
Did You Know?
The Science Team recently presented research on Collaborative Location Prediction at the New York Academy of Sciences Seventh Annual Machine Learning Symposium.
Scientific Challenges “Beyond Place”
These days, the science team is focused on understanding people, in addition to places. This introduces several new challenges, which we discuss below. For some of these challenges, we have solutions, others are in-progress, and all have been researched extensively:
+ Doing the POI lookups at scale. We’ve gone through several rounds of development in building our lookup TreeServer™, which uses kd-Tree and Priority R-Tree algorithms. We do billions of lookups per day.
+ Handling the uncertainty in the location fixes: First, some location readings are flawed, and we need to filter them out so we don’t corrupt our features – we’ve developed algorithms for detecting these anomalies. Next, we need to handle the more traditional uncertainty in location. Seamlessly combining observations with different levels of accuracy is one of many challenges we’ve addressed. Our platform incorporates the accuracy of the underlying location-measurement source. For example, GPS location is weighted more heavily than network-based location.
+ Deriving features at different time resolutions: For identifying car shoppers, we want to estimate the frequency of visits to car dealerships in the past day or week. On the other hand, for identifying international travelers, recent history won’t suffice: we look at the number of trips people have taken abroad since the beginning of our records. If we generically want to know how often an individual does something “these days,” we need to balance between going far back in time (which would give us a lot of data for estimating the rate), with focusing on the most recent observations (which ensures the rate is up-to-date).
+ Trading off between testing and scaling: Since we frequently get new ads and new users we need to trade off a “testing phase”, in which we try an ad with different types of people, with a “scaling phase”, in which we show the ad to those people who responded the best in the test phase. This is a classic “exploration/exploitation” problem, and it fits nicely into the Contextual Bandit framework. Sense Networks’ unique challenge is to leverage a user’s historical location features to assist with the ad targeting.
Science : Results
Sense Networks has confronted and solved many interesting location data challenges over the last few years. This includes fundamental new research efforts to advance location data science, as well as commercializing the Sense Networks platform to drive value for our customers and partners.