Decision-Understanding Oriented Interactive Visualization Chart Recommendation

Data visualization is an important part of data analysis task.In order to reduce the burden of analysts and assist them to obtain meaningful visualizations,visualization recommendation technology has emerged as the right moment with the rapid growth of data volume and complexity.Inspired by the successful application of machine learning models in various automated tasks,solving visualization recommendation problems with machine learning methods become a research hotspot.However,the existing research has the following limitations: First,there is a lack of efficient end-to-end visualization recommendation scheme combining rules;At the same time,the recommendation model does not support human-in-the-loop,unable to integrate the advantages of human and machine.To solve the above problems,this paper proposes Vis Pal,an interactive chart recommendation framework for decision understanding.The specific research results are as follows:(1)To solve the problem of lack of efficient recommendation scheme with and rulebased end-to-end approach,a chart recommendation scheme based on deep reinforcement learning is proposed.Firstly,this paper proposes a diagram syntax template constructed according to the chart generation logic.The chart recommendation problem is abstracted to the action token sequence generation problem based on the data table according to the grammar template,and a special feature transformation network is defined to capture the feature information of the action token better when embedding.Then,according to the idea of reinforcement learning,the corresponding state space,action space and reward value are defined,and the Chart Recommendation Deep Q-Network(CRec DQN)with copy mechanism is proposed to complete the sequential decision process.Finally,the chart recommendation problem is solved by heuristic beam search based on fitting function.The effectiveness of Vis Pal recommendation scheme is demonstrated by comparing with several popular visualization recommendation systems on open datasets.(2)To solve the problem of recommendation model that doesn’t support human-inthe-loop,propose a human-machine collaboration visualization chart recommendation scheme based on Hooks mechanism.Firstly,to realize the user’s decision understanding and forward and backward interaction with the model,Hooks mechanism is introduced in this paper,and adaptively designed for the chart recommendation task.The Hooks mechanism presents all the effective actions in the action space of each state in interface,which increases the interpretability of the decision process by displaying the value of the action,and at the same time completes the semantic interaction between users and interfaces through state rollback.To make the model aware of user modification and not affect the stability of the recommendation model,we design a network,independent of the recommendation model,Hooks Net,and add the output of the network and the results of the recommendation model by weight,to achieve the purpose of influencing the subsequent decisions according to the user’s actions.Finally,the case study proves that the collaborative chart recommendation scheme proposed by us has high user satisfaction.