Position Recommendation Based On Graph Convolutional Network

GitHub has become more than just a code-hosting platform.The open source communities it spawned are not only benefiting developers,but increasingly affecting our daily lives.Many developers use GitHub as their second resume,because there is a clear track record of their actions in GitHub.However,for recruiters,these concepts may be vague,so the main work of this paper is to match developers and Job advertisement,so as to automatically recommend jobs and solve the semantic gap between developers and recruiters.Recommendation algorithms require user-item rating dataset.But the original Github log data and job advertisement datase are isolated from each other,so the first work of this paper is to establish a rating mechanism.The workflow of the algorithm is as follows:Firstly,use named Entity Recognition(Named Entity Recognition,NER)and named Entity disambiguation(Named Entity Disambiguation,NED)to extract IT entities from GitHub log and job advertisements.Then use these entities to build one-hot vectors for developers and recruitment information.Then use word embedding to optimize one-hot vectors and use cosine similarity to calculate the similarity,the similarity is used as the scoring basis between the developer and the job advertisement,and the scoring dataset is constructed based on this.Inspired by graph convolution network(Graph Convolutional Network,GCN),a collaborative filtering model based on GCN is developed after obtaining the rating data set of developer-job advertisements.The core idea of the algorithm is that the developerjob advertisements dataset is a graph data,and the existing work usually obtains the user's embedding only from the user's characteristics(such as ID and attribute),which ignores the interaction information between the user and the job advertisements in the graph data.In this paper,we capture the interaction by embedding propagation.The model is compared with MF,Neu MF,CMN,and GC-MC.The advantages of the model are verified by comparing Recall,MRR,HIT,and Precision.In addition,since feature transformation and nonlinear activation in graph convolution network are added to the development of GCN recommendation algorithm,a large number of ablation experiments are conducted to prove that discarding feature transformation and nonlinear activation can make the model achieve better performance.At the same time,the penalty term is introduced in the final embedded combination to solve the phenomenon that the graph propagation may be too smooth,and the validity of the penalty term is verified by experiments.