Predicting response in mobile advertising with hierarchical importance-aware factorization machine

Abstract

Mobile advertising has recently seen dramatic growth, fueled by the global proliferation of mobile phones and devices. The task of predicting ad response is thus crucial for maximizing business revenue. However, ad response data change dynamically over time, and are subject to cold-start situations in which limited history hinders reliable prediction. There is also a need for a robust regression estimation for high prediction accuracy, and good ranking to distinguish the impacts of different ads. To this end, we develop a Hierarchical Importance-aware Factorization Machine (HIFM), which provides an effective generic latent factor framework that incorporates importance weights and hierarchical learning. Comprehensive empirical studies on a real-world mobile advertising dataset show that HIFM outperforms the contemporary temporal latent factor models. The results also demonstrate the efficacy of the HIFM's importance-aware and hierarchical learning in improving the overall prediction and prediction in cold-start scenarios, respectively.

User Response Learning for Directly Optimizing Campaign Performance in Display Advertising

“…(iii) As for CTR, we find that EU achieves the highest CTR and RR also performs better than CE. Here FM has higher CTR than the CE model because it could learn feature interactions via the latent vector inner product [21]. However, FM obtains relatively less profit gain and ROI than CE, which shows that FM does not care enough about those auctions with high return value.…”

“…The response prediction is a probability estimation task [19] which models the interest of users towards the content of publishers or the ads, and is used to derive the budget allocation of the advertisers [23]. Typically, the response prediction problem is formulated as a regression problem with prediction likelihood as the training objective [23,9,1,21]. From the methodology view, linear models such as logistic regression [14] and non-linear models such as tree-based model [10] and factorization machines [19,21] are commonly used.…”

“…Typically, the response prediction problem is formulated as a regression problem with prediction likelihood as the training objective [23,9,1,21]. From the methodology view, linear models such as logistic regression [14] and non-linear models such as tree-based model [10] and factorization machines [19,21] are commonly used. Other variants include Bayesian probit regression [9], FTRFL [24] in factorization machine, and convolutional neural network learning framework [17].…”

“…There are 4 deployed models: EU, RR, CE and FM, where the first three have been discussed in Section 4.4 and FM is a factorization machine model [21] with non-hierarchical feature engineering. To show the comparable performance of user response prediction, we set the same linear bidding function for all prediction models including baselines.…”

“…First, the CTR/CVR 1 estimation is formulated as a binary regression problem, which can be solved by such as logistic regression with SGD [23] or FTRL learning [18], Bayesian probit regression [9], gradient boosting regression trees [10], and Factorization Machines [21]. The common objective in this stage is to make the estimation as accurate as possible by, for instance, minimizing the error between the predictions and ground truth user responses.…”