PROJECT TITLE :

Heterogeneous Network Representation Learning A Unified Framework with Survey and Benchmark

ABSTRACT:

Traditional homogeneous networks have been replaced with the more powerful, realistic, and generic heterogeneous networks in recent years. This is due to the fact that real-world objects and the interactions between them frequently involve more than one mode and more than one type (graphs). In the meantime, representation learning, also known as embedding, has recently been the subject of extensive research and has been demonstrated to be effective for a variety of different network mining and analytical tasks. This work aims to provide a unified framework for deeply summarizing and evaluating previous research on heterogeneous network embedding (HNE). While a standard survey is a part of this, our goal is to go beyond the scope of a survey and provide a more comprehensive analysis. We provide a generic paradigm for the systematic categorization and analysis over the merits of various existing HNE algorithms as the first contribution of this article because there has already been a large body of HNE algorithms. This is our first contribution because there has already been a large body of HNE algorithms. In addition, even though they are generally claimed to be generic, existing HNE algorithms are frequently evaluated using a variety of datasets. It is understandable due to the application favor of HNE; however, such indirect comparisons significantly hinder the proper attribution of improved task performance towards effective data preprocessing and novel technical design. This is especially true when taking into consideration the numerous ways it is possible to construct a heterogeneous network from data derived from real-world applications. As a result, as our second contribution, we create four benchmark datasets from a variety of sources, each with their own unique scale, structure, attribute/label availability, and other characteristics, with the goal of making evaluations of HNE algorithms more convenient and objective. As part of our third contribution, we carefully refactored and modified the implementations of 13 well-known HNE algorithms, as well as created user-friendly interfaces for these algorithms. Additionally, we compared these algorithms comprehensively across a variety of tasks and experimental settings. We aim to provide a universal reference and guideline for the understanding and development of HNE algorithms by placing all existing HNE algorithms inside of a unified framework. This will allow us to accomplish this goal. In the meantime, by making all of the data and code publicly available, we plan to provide the community with a fully operational benchmark platform that can be used to evaluate and evaluate and compare the performance of existing and future HNE algorithms.