Understanding Career Mobility Behavior Based On Attributed Graph Mining

Career mobility refers to the talent exchange phenomena among organizations,regions,or industries,which is caused by individuals' job change behavior.It is rooted from the individual's occupational pursuit,skill level as well as the educational background,and usually affected by the economy and labor market.The dynamics of career mobility have considerable impacts on family life,organizational structure,and even national talent strategy.In recent years,the economic globalization has led to a more active global labor market which brings new opportunities and challenges to the research community of career mobility.Traditionally,career mobility has been studied by human resource management and organizational behavior community.However,due to the limitation of data sources,the research work had been restricted on either small-scale or coarse-grained qualitative analysis.Nowadays,billions of resumes have been digitized on the Internet,which makes it possible to gain more insights into large-scale and fine-grained career mobility.Meanwhile,the growing computational capability and data mining technology have opened up new paradigms for data analysis and modeling.To this end,we propose to study career mobility by mining attributed graphs.Specifically,the framework is built on the basis of a graph representation of career mobilities.The conversion among different granularities is implemented by the reconstruction of graphs.In this way,the research issues of career mobility are converted into mathematical modeling problems,and we study two key issues in detail,namely human resource locating and talent flow prediction.The main contributions of this dissertation are summarized as follows:First,we analyze the common characteristics of career mobility among different granularities and introduce the attributed graph representation.In specific,all career mobilities include a source,a destination,a relation between the source and the destination,as well as additional information.These characteristics make the behavior suitable to be represented by the attributed graph.Therefore,we introduce the formalization of the attributed graph and propose a method to extract job trajectory graph from individuals' work experience.In addition,we show several statistics of the dataset used in the dissertation,and analyze some key properties of the dataset including the data sparsity,distribution of tenures and correlation with external variables.Then,we propose two reconstruction methods of the attributed graph,which is used to convert career mobility among different granularities.In the context of attributed graph,the conversion can be implemented by reconstructing graphs,i.e.,constructing both intra-organizational and inter-organizational graph from job trajectory graph.The inter-organization graph is reconstructed by a heuristic algorithm.On the contrary,reconstructing intra-organizational graph is non-trivial and has several challenges,including the unobservable job positions ranks,the variability of individual expertise,and the conflicting of job trajectories.In response to these challenges,we propose the concept of the difficulty of promotion to quantify the rank of job positions.A Gaussian Bayesian graphical model is designed to model the joint distribution of promotion difficulty and ranks.The analytic form of posterior distribution of the ranks is further derived.Experiments on a large set of real data are conducted to evaluate the performance of the model.The result shows that the proposed model is more precise than benchmark models.Furthermore,a clustering model on the static inter-organizational graph is proposed to solve the talent sourcing problem.Specifically,talent sourcing is the process of locating the potential organizations which contain candidates during a recruitment.In general,talent sourcing is an efficient way to reduce the cost of recruiting but is a difficult task,facing the challenges of massive search space,high diversity of talents and complicated recruiting demands.The task is converted into a clustering problem and an efficient model is proposed.The proposed model first defines the similarity among nodes according to the distribution of career mobility.Then the imaginary edge among nodes are generated according to the clustering status,and an objective function is designed to measure the difference between the weight of imaginary edges and actual edges.Next,a simulated annealing algorithm is implemented to search for an optimal solution of the objective function.Experiments on a real dataset show that the proposed model outperforms several benchmarks.Several case studies are also conducted to show the usability of the results.Finally,we develop a model to forecast talent flows on the basis of dynamic attributed graphs.Talent flow forecasting refers to the quantitative prediction of career mobility among organizations,cities or industries,but it is difficult with two obstacles.First,the factors that may affect talent flow are complicated,including the economy,commercial performance,reputation and so on.Second,the data are distributed unevenly among organizations such that there is a data sparsity problem.We formalize the forecasting task as an edge weight estimation problem in a dynamic graph.The influential factors are bound to the graph as node attributes,while external sequences are as dynamic attributes for alleviating the data sparsity problem.We propose a model to fuse all the attributes together and make a multi-step prediction of future edge weights.The model leverages several recurrent neural networks for coding these attributes and uses the attention mechanism to make the prediction within an encoder-decoder structure.Experiments are conducted on the prediction of talent flows among a number of public companies.The results show that the proposed model has lower prediction error rate than baselines and the performance is improved with the help of external data.