Diverse Interaction Recommendation for Public Users Exploring Multi-view Visualization using Deep Learning

Original link: http://vis.pku.edu.cn/blog/%E5%88%A9%E7%94%A8%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4% B9%A0%E4%B8%BA%E5%A4%A7%E4%BC%97%E6%8E%A2%E7%B4%A2%E5%A4%9A%E8%A7%86%E5%9B% BE%E5%8F%AF%E8%A7%86%E5%8C%96%E8%BF%9B%E8%A1%8C%E5%A4%9A%E6%A0%B7%E5%8C%96/

The figure is a visualization system with multiple views, which is used to help explore the travel trajectories of the four poets in the Song Dynasty at various stages of their lives, as well as literary creations, analyzing the poets’ lives from multiple perspectives. The visualization system can be placed in scenes such as museums as a visualization case for the general public. Utilize its highly interactive and rich content to help the public understand relevant domain knowledge in a unique format.

The Song Dynasty Poet Trajectory and Literary Creation Visualization System

However, such complex interactive visualizations in public popular science scenarios such as museums are often not fully utilized by users. This is because users usually lack background knowledge about domain knowledge and data analysis, and visualization systems also lack corresponding help guides. Typically, mass users only interact randomly, focus on a single view, a single interaction, and quickly lose interest in visualization systems. In this process, it is usually impossible to obtain relevant exploration discoveries. In order to solve this problem, researchers from Fudan University and Tongji University proposed that real-time diversified interactive recommendations can be embedded in the visualization system to help users make more discoveries when exploring the multi-view visualization system and improve related scenarios. Utilization of the visualization system.

The researchers believe that the user’s interaction behavior is affected by the previous interaction operation and the information display of the visualization system, and try to use the long short-term memory network (LSTM) model for interaction prediction, and use the prediction result as an interaction recommendation. To achieve this goal, it is first necessary to encode the user’s interaction sequence and the information presentation of the visualization system. According to previous research [1], interaction can be divided into three high abstract types, and each type of interaction has corresponding parameters to describe the actual interaction behavior. Taking the visualization system introduced at the beginning of the article as an example, all the interactive operations and the corresponding required parameter list are shown in the following figure.

Three high-abstract interaction types (left), and the specific interaction types involved in the visualization of Song Dynasty poets (right)

In this way, a single interaction can be turned into an interaction log in JSON format and further encoded into a vector for use in machine learning models. In addition, the encoding of the data displayed by the visualization can be understood as the encoding of the information on each view. The information involved in the view is disassembled, reorganized into keyword vectors, and further dimensionality reduction operations are performed, which can also be converted into vectors. For example, a map view with punctuation can be transformed into a vector of the information combination of each point on the map, including the position, color, size and other information of each point. Then, dimensionality reduction is performed on the long information vector, and the tSNE algorithm is used to obtain a shorter vector for use in the machine learning model.

When collecting and arranging training data, the author also set up two filters to improve the diversity of recommended interactions and ensure that user interactions can have certain discoveries. The author calculates the current interaction type and the entropy of the interaction object, and establishes a diversity threshold to filter out the training data below the threshold; in addition, according to the preset valuable exploration area, evaluate the recommended interaction and filter the low-value exploration. recommend.

Training module and real-time update module for system model

Then, taking a single interaction operation and the corresponding visualization state as a group, as the state of the LSTM model, and controlling the sliding window to continuously slide, the model can be used to continuously predict the next interaction operation. The interaction behavior predicted by the model is presented to the user as a recommendation, highlighted on the visualization. When the user accepts the corresponding recommendation, the recommendation is considered valid; when the user does not accept the recommendation, the model will update the model in real time according to the user’s choice as a supplement to the training data.

The authors conducted user experiments to verify the effectiveness of the method. Subjects were divided into three groups and explored for about 8 minutes on the visualization without the recommender system, the visualization with random recommendations, and the visualization system recommended by the model. The subjects were then asked to rate the 8 provided statements to verify how much information was learned during the exploration; the latter two groups of subjects were also asked to rate the provided recommended interactions.

The scoring results of 8 statements, of which Q1-Q6 are correct statements and Q7-Q8 are false statements

From the scoring results, it can be seen that users who use the model to provide recommended interactions give higher agreement on correct statements (Q1-Q6) and lower agreement on false statements (Q7-Q8). This shows that users who use the model recommendation interactive system can obtain more information in the exploration. In addition, experiments also show that even random interactive recommendation can greatly improve the diversity of users’ actual interactive operations. At the same time, subjects using model recommendations rated the recommended interactions higher than those who received random recommendations.

This work explores the use of real-time interactive recommendation methods to help users who lack the background of data analysis knowledge to explore complex multi-view visualization systems, and achieves good results. This work can greatly improve the utilization rate of visualization systems displayed in museums and other scenarios, allowing more users to use visualization systems for data exploration.

[1] JS Yi, Y. a. Kang, J. Stasko, and JA Jacko. Toward a deeper understanding of the role of interaction in information visualization. IEEE Transactions on Visualization and Computer Graphics, 13(6):1224–1231, 2007. doi: 10.1109/TVCG.2007.70515

[2] Li, Y., Qi, Y., Shi, Y., Chen, Q., Cao, N., & Chen, S. Diverse Interaction Recommendation for Public Users Exploring Multi-view Visualization using Deep Learning. IEEE VIS 2022.

This article is reproduced from: http://vis.pku.edu.cn/blog/%E5%88%A9%E7%94%A8%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4% B9%A0%E4%B8%BA%E5%A4%A7%E4%BC%97%E6%8E%A2%E7%B4%A2%E5%A4%9A%E8%A7%86%E5%9B% BE%E5%8F%AF%E8%A7%86%E5%8C%96%E8%BF%9B%E8%A1%8C%E5%A4%9A%E6%A0%B7%E5%8C%96/
This site is for inclusion only, and the copyright belongs to the original author.