A Learning-based Approach For Visualization Of Big Data

Interactive data exploration(IDE)is a classical visualization research topic.Most existing techniques achieve rapid query response by designing a data structure to store aggregate values of massive queriable attribute range combinations.These traditional models show some limitations in flexibility.First,they have a high storage overhead.Second,users can not change the aggregation function during the exploration.Third,they only return aggregate values,without the support of constructing views showing original records.This paper presents a new IDE technique named Learned Visualization Index Structure(called LVI below for simplicity).Unlike existing IDE techniques outputting aggregate values,LVI returns original records satisfying an IDE query by training regression models to predict the location of records within user specified attribute ranges,and uses a search strategy to avoid bias.The features bring better use flexibility.First,it has low storage overhead and enables high-resolution query and display on more finely-binned attributes.Second,it enables interactive change of visualized aggregate measures during the exploration.Third,it enables the construction of views showing both aggregate patterns and original records.We design three optimization strategies for LVI,which can be used simultaneously to further optimize the performance of LVI in construction,storage,execution and update.One update strategy and one parallel computation strategy are also proposed.The update strategy improves the update efficiency of LVI,and the parallel computation strategy compresses the online computation time overhead of LVI.Several experiments are conducted on seven open datasets.Experimental results demonstrate LVI’s better use flexibility,significantly smaller size,comparable query speed to existing techniques,and the sizes of LVI are only 1.3%-5.3% of Nanocubes.Moreover,LVI enables high update efficiency in real-world scenarios and enables interactive exploration over larger datasets and more dimensions with more hardware resources.