Deep Semantic Learning Based Software Developers Intelligent Recommendation

Intelligent collaborative software developer recommendation has become a hot research direction in recent years.Software development as an intensive intellectual activity,the developers’ collaboration efficiency is an important factor which may affect the efficiency and quality.Therefore,how to improve the developers’ collaboration efficiency is an important research topic.Traditional developer recommendation approaches usually suffer from data sparse and cold start problems,which may lead to low recommendation accuracy,poor diversity and interpretability.How to effectively mine the characteristics of developers and tasks,analyze their explicit and implicit relationships,and improve the overall collaboration efficiency is an urgent problem to be solved.To solve this problem,this thesis conducts an in-depth study on different developer recommendation scenarios(e.g.,technical Q&A community,open-source software community,crowdsourcing development platform).In particular,we propose a series of novel developer recommendation methods.The main contributions of this thesis can be summarized as the following aspects:(1)In the Q&A community scenario,existing developer recommendation methods(i.e.,Q&A expert recommendation)usually ignore the deep explicit and implicit relations between questions and developers,developer capabilities,and the interpretability of recommendation results.To solve those problems,this thesis proposes a developer recommendation approach based on the high-order embedding propagation of knowledge graphs.First,we propose to build a domain knowledge graph for Q&A community,and generate feature vectors of questions and developers through knowledge graph embedding learning.Secondly,we propose to refine the feature vector with the help of knowledge graph high-order propagation and attention mechanisms.By this means,the problems of sparse interaction between questions and developers and the inefficient embedding propagation are effectively solved,and the accuracy of recommendation is effectively improved.Finally,we explicitly model the developer’s capabilities and generate a high-order embedding propagation and capability awarded recommendation framework.We conduct extensive experiments based on the Stack Overflow,and the results demonstrate its superiority in recommendation accuracy and interpretability.(2)In the open-source software community scenario,existing developer recommendation methods(i.e.,code reviewer recommendation)usually ignore the deep correlation between developers and the textual semantics of code snippets.This may lead to poor recommendation accuracy,especially under the conditions of data sparse and noisy problem.To solve this problem,this thesis proposes an attention mechanism-based neighbor embedding enhanced code text semantic feature learning and code reviewer recommendation approaches.First,we propose to leverage the Transformer model and multi-head attention mechanism to learn the semantic features of code text and then generate the code and reviewers’ representations.Secondly,to enhance the representation of code and reviewers,we leverage the embedding propagation based on the attention mechanism and graph neural network.We conduct extensive experiments based on the Git Hub real-world datasets,and the results demonstrate its superiority.(3)In the crowdsourcing development platform scenario,existing developer recommendation methods(i.e.,crowdsourcing developer recommendation)are usually based on the developer’s feature vector of ID representation learning.Those methods may have problems such as low recommendation accuracy and poor diversity in the case of sparse data and noise problem.To solve this problem,this thesis proposes a text enhanced graph embedding learning and diversity fused developer recommendation framework.First,we learn the text feature vectors of developers and tasks,and aggregate those text feature vectors with their ID feature vectors.Secondly,to enhance the ability of developer and task’s representation learning,we propose to leverage the high-order graph embedding learning to propagate and stack their neighbors’ embeddings.Then we calculate the matching relations between tasks and developers based on the final enhanced embeddings.Finally,we model the diversity of team members and propose a boundary-restricted greedy algorithm to approximate the matching relation between developers and tasks and the diversity of team members.We conduct extensive experiments based on the Topoder real-world datasets,and the results demonstrate its superiority.