A Miscellaneous Joint-Learning Model for Audience Segmentation in Online Advertising

Abstract

Audience segmentation in online advertising is most often generated from browsing information found on server logs and identifiers like cookies. Beside the primary use for advertisement targeting, those segments optimize the distribution on many other tasks downstream. One bottleneck for performance improvement is the integration of trivial and heterogeneous features in large scale. Manual feature engineering is generally required to achieve higher accuracy. Traditional classification models further have deficits in learning heterogeneous features such as browsing history together with text. In this work, we append other features typically present in digital advertising setups, such as user search queries, to the device browser history. Then we propose a model that integrates heterogeneous features by joint-training on three low-coupled networks of radial basis function (RBF)-, deep- and recurrent network. Each network can separately learn on different features. The weights of each network can be updated simultaneously through a unified loss function. We discover that the joint-training model is also effective for sparse features. Experiments on two online advertising datasets validate our model performs better than the broadly adapted ensemble tree model.