Learning Campaign Representations Without Explicit Labels

In Microsoft Advertising platform, an ad campaign is a collection of `similar’ ads that are intended to achieve a specific goal as per advertisers. Large number of campaigns are created every day by our advertisers to create brand awareness about their products and drive sales up by  providing discounts. Advertisers come to our platform and create ads under a campaign, associate a set of keywords, assign budget to them, choose target location and audience. Since these campaigns are such meaningful entities, it is important that we understand them well and have a way to represent them using different signals that we have (advertiser provided data, user provided clicks). There has not been much work in this field due to lack of consensus on how to define a campaign, how to evaluate them, and how to acquire labelled data to train a supervised machine learning model. Long running campaigns for which we have enough information can be represented using historical data while this is not the case for new campaigns and thus would lead to cold-start issues. In this work we tackle this very problem of learning campaign representations, first for long running campaigns (using ad-query co-occurrence data) then extend this learning to novel campaigns scenario, without need of any explicit labels. We propose various use cases where such embeddings can be used (evaluated), and perform initial experiments that prove that our embeddings are of good quality, encode useful information, and can help improve the performance of important tasks such as Click Prediction and Bid pricing.