PROJECT TITLE :

App Popularity Prediction by Incorporating Time-Varying Hierarchical Interactions

ABSTRACT:

The task of predicting an app's future popularity based on its current behaviors is an important part of the process of developing mobile services. This task is known as app popularity prediction. It offers benefits ranging from the development of apps to investment in specific areas. Popularity can be influenced by two different types of factors: those that are internal, such as reviews, and those that are external, such as interactions between apps. On the other hand, the vast majority of related studies only investigate internal factors and ignore external ones. Because it is the promoting and/or inhibiting influence that is the result of app interaction, external factors play an important role in popularity prediction modeling. In fact, they play a significant role. The interaction with the app has two primary characteristics, namely, interactivity and dynamicity, both of which present difficulties in predicting the popularity of the app for two reasons: 1) Interactivity — It is difficult to evaluate the existence of interactions and the degree to which they influence; 2) Dynamism — The nature of an interaction's influence, such as whether it is promoting or inhibiting popularity, and the intensity of that influence can change over time. In this article, we present DeePOP, a model for predicting popularity that makes novel use of time-varying hierarchical interactions. To begin, in order to organically characterize the relationship between apps and the influence they have on one another, we propose using the Hierarchical Interaction Graph, which is first studied in this work. Second, when it comes time to construct the prediction model, DeePOP takes into account both time-varying hierarchical interactions and internal factors as inputs. It does this by first developing multi-level modules that are based on Recurrent Neural Networks with attention mechanisms, and then by fusing the outputs of these modules to produce multi-step time series predictions. Experiments performed on a dataset derived from the real world demonstrate that DeePOP outperforms methods considered to be state-of-the-art in terms of prediction accuracy. As a result, the Root Mean Square Error (RMSE) was effectively reduced to 0.088.