AdKDD 2019  |   Questions - organizers@adkdd.org  |  Contact Us

In-app Purchase Prediction Using Bayesian Personalized DwellDay Ranking

Shonosuke Harada, Kazuki Taniguchi, Makoto Yamada and Hisashi Kashima

With the increasing popularity of mobile devices like smartphones,the number of display advertisements is also increasing; especially,advertisement for mobile game applications is one of the largestbusiness among them. In recent years, advertisers claim that LifeTime Value (LTV) is even more important metric for advertisingin the acquirement of new users than Cost Per Install (CPI). There-fore, we need to predict users who will purchase XXXX in theapplications in future and show the ads to them. To this end, weconsider the in-app purchase prediction problem to promote in-apppurchases to user in a personalized manner. In contrast with installprediction, purchase prediction is difficult in the sense that thenumber of in-app purchases is much smaller than that of instal-lations. To handle this issue, following the idea of the Bayesianpersonalized ranking (BPR) framework, we exploit installation ofan application as an intermediate feedback between purchase andunobserved feedback and propose the install-enhanced BPR. Morespecifically, in the proposed method, we enhance the quality ofthe intermediate information by combining the dwell time on theapplication. Through experiments using the real world dataset, weshow that the proposed method outperforms baselines as well asseveral insights on user activity.