Modeling feature interactions for context-aware QoS prediction of IoT services

Accurate Quality of Service (QoS) prediction of service is a key measure to accomplish successful applications such as QoS-aware service recommendation and composition in Internet of Things (IoT) environments. The key of this task is to consider contextual information like geographic location, network address and type of service, since they have subtle but powerful effects on QoS of IoT services. Recently proposed context-aware QoS prediction for IoT services follow two general paradigms: clustering contextual information for calculating similarity between users and services or integrating contextual information by extending latent factor models. However, the simple clustering contextual information or learning the latent feature of contextual information do not go much further to discover complex and intricate user–service interaction patterns. To this end, we propose a context-aware feature interaction modeling (CFM) approach for IoT services to perform QoS prediction, considering context as an additional feature similar to users and services and modeling their interactions. The proposed method can capture both low-order and high-order feature interactions using contextual information and user's invoked records, which consists of three phases: (1) learn low-order feature interactions by decomposing the sparse user–service QoS matrix with factorization machine; (2) learn high-order feature interactions explicitly and implicitly with a multilayer perceptron and deep cross network; (3) aggregate the output of low-order and high-order feature interactions with a parametric-matrix-based fusion. Experimental results on a large-scale QoS dataset demonstrate that the proposed method consistently outperforms state-of-the-art baselines in terms of various evaluation metrics.