PKE: A Model for Recommender Systems in Online Service Platform

Graph embedding is a technique that has grown attention in recent years. Apart from mining implicit information in a graph representation data, graph embedding can be used in recommender system. Recommender system is a tool that can help e-sellers collect users' information more easily. This is beneficial for platform providers to mine users' interests with more data. However, it is easy to be distracted by other unrelated information when too many data are collected. In this paper, we proposed an idea called information matrix to combine information from different data. This idea considers data as graphics; hence, connects data with common nodes and extends to vectors through keyword embedding. In addition, we constructed a model that illustrates the efficiency and ability of the information matrix. This model was tested on data provided by e-sellers. Our aim was to combine browsing data and order data, and transfer these data into information matrix with high accuracy. Although the accuracy rate drops after transferring information into embedded type, it provides an idea for potential solution of cold start, a common problem in recommender systems.