Data Mining Based On Multi-source Data Fusion

With the advent of the information age,the sources of information have become more diverse.Multi-source data can describe an object from multiple angles,and there may be correlations between data sources.To discover more hidden information of multi-source data sets,multi-source data fusion methods are widely used.The fusion of multi-source data can expand the functions based on the existing methods and provide interpretability for complex models;it can also provide more accurate predictions by mining comprehensive information from various angles.Current research ignores deep connection among data sources,which leads to unsatisfying results.This thesis explores integration of different data sets in two fields.Firstly,in urban computing,to address the city functional region identification problem,we find that text with geographic coordinates can bring interpretability to the system.We integrate online review text data,propose an explainable functional region identification model EFRI,and verify EFRI on a real-world dataset.EFRI not only can effectively cluster regions based on city functions,but also has intuitive interpretability.Secondly,in bioinformatics,to predict Drug-Target Interaction(DTI),we assume the similarity between drug features and target features are positively correlated with their corresponding DTI values.With this assumption,we propose a multi-task learning deep model called Multi-DTI to predict DTI.This model uses co-attention mechanism and cosine similarity to fuse features.Experiments show that Multi-DTI is outperformed than existing models.