Ontology Matching Tuning Based On Machine Learning

Ontology,as a formal definition and description of common recognized knowledge in the domain,brings the problem of ontology heterogeneity while sharing domain knowledge.Ontology Matching is an effective way to solve the problem of ontology heterogeneity.In an ontology matching system,there are many adjustable matching parameters and matching strategies.If we can choose the best matching parameters and strategies for different matching tasks,we can get better matching results.This process is Ontology Matcher Tuning.The existing matching tuning work is mainly accomplished by experienced experts,and it is difficult for ordinary users to start.In this paper,the automatic matching tuning technology is studied from two parts:parameter and strategy.The main contents are as follows:(1)A parameter tuning method based on machine learning is proposed.The problem of parameter tuning is regarded as a multi-output regression problem.Firstly,feature engineering is designed according to matching task,and the training set training model is constructed by collecting the approximate optimal parameter group of particle swarm optimization tuning method under historical matching task.This method has better robustness to unknown matching tasks.It can realize automatic tuning of parameters without reference matching and improve the usability of matching system.Compared with the default parameter configuration built in the system,the optimal parameters calculated by the system can effectively improve the matching accuracy.In some experimental ontology matching systems,their F1 values increased by an average of 4.7% on OAEI benchmark datasets.(2)proposed two strategy tuning methods based on machine learning.From the point of view of supervised learning,the strategy tuning problem is regarded as a classical multiclassification problem,and the candidate sets of matching strategies in each historical matching task are searched in depth first,and the optimal matching strategies of each task are searched to construct the training set.Finally,the trained model can automatically calculate the optimal matching strategy of the task according to the characteristics of the new task.?From the point of view of reinforcement learning,the strategy tuning process is decomposed into a statebehavior graph applicable to reinforcement learning,and then the Q-Learning reinforcement learning algorithm is used to solve the optimal matching strategy for a given task.Compared with the default matching strategy built in the system,the two strategy tuning methods can effectively improve the matching accuracy and ease of use of the matching system.(3)proposed a method to accelerate the construction of training set.The training set relied on in the tuning process takes too long to construct.By using graph sampling technology,sampling operation is carried out on large ontology.The sub-ontology task after sampling is searched for the optimal solution,and the search results are directly taken as the solution of the original task,which greatly reduces the search time.Experiments show that this method can effectively shorten the construction time of training set while retaining a certain tuning effect.(4)developed an ontology matching system Lily-TM which can automatically tune matching parameters and strategies.The system not only integrates the current mainstream ontology matching algorithm,but also has built-in parameter and strategy tuning algorithm based on machine learning.It can effectively tune matching parameters and strategies while realizing general ontology matching function.