Scalable Distributed Inference of Dynamic User Interests for Behavioral Targeting
Abstract
Historical user activity is key for building user pro les to
predict the user behavior and anities in many web ap-
plications such as targeting of online advertising, content
personalization and social recommendations. User pro les
are temporal, and changes in a user's activity patterns are
particularly useful for improved prediction and recommen-
dation. For instance, an increased interest in car-related web
pages may well suggest that the user might be shopping for
a new vehicle.In this paper we present a comprehensive sta-
tistical framework for user pro ling based on topic models
which is able to capture such e ects in a fully unsupervised
fashion. Our method models topical interests of a user dy-
namically where both the user association with the topics
and the topics themselves are allowed to vary over time,
thus ensuring that the pro les remain current.
We describe a streaming, distributed inference algorithm
which is able to handle tens of millions of users. Our re-
sults show that our model contributes towards improved
behavioral targeting of display advertising relative to base-
line models that do not incorporate topical and/or tempo-
ral dependencies. As a side-e ect our model yields human-
understandable results which can be used in an intuitive
fashion by advertisers.
predict the user behavior and anities in many web ap-
plications such as targeting of online advertising, content
personalization and social recommendations. User pro les
are temporal, and changes in a user's activity patterns are
particularly useful for improved prediction and recommen-
dation. For instance, an increased interest in car-related web
pages may well suggest that the user might be shopping for
a new vehicle.In this paper we present a comprehensive sta-
tistical framework for user pro ling based on topic models
which is able to capture such e ects in a fully unsupervised
fashion. Our method models topical interests of a user dy-
namically where both the user association with the topics
and the topics themselves are allowed to vary over time,
thus ensuring that the pro les remain current.
We describe a streaming, distributed inference algorithm
which is able to handle tens of millions of users. Our re-
sults show that our model contributes towards improved
behavioral targeting of display advertising relative to base-
line models that do not incorporate topical and/or tempo-
ral dependencies. As a side-e ect our model yields human-
understandable results which can be used in an intuitive
fashion by advertisers.
