针对气味数据的交互式聚类可视分析框架  

A Visual Analytics Framework for Interactively Clustering Scent Data

作　　者：黄立 张静怡 吴昊 路奇[1,2,4] Huang Li;Zhang Jingyi;Wu Hao;Lu Qi(Academy of Arts&Design,Tsinghua University,Beijing 100084;The Future Laboratory,Tsinghua University,Beijing 100084;The Australian National University,Canberra 2600;Tsinghua University-Alibaba Joint Research Laboratory for Natural Interaction Experience,Beijing 100084)

机构地区：[1]清华大学美术学院,北京100084 [2]清华大学未来实验室,北京100084 [3]The Australian National University,Canberra 2600 [4]清华大学-阿里巴巴自然交互体验联合实验室,北京100084

出　　处：《计算机辅助设计与图形学学报》2020年第7期1026-1041,共16页Journal of Computer-Aided Design & Computer Graphics

基　　金：清华大学自主科研计划(20197010003).

摘　　要：自然界中的气味通常是多种成分的混合物,相比图像和声音具有特殊的数据复杂性,且由于缺少有效的表示方法导致气味研究相对滞后.聚类分析是研究气味的一个重要方法,有助于增进对气味的理解、识别、合成和设计,然而大部分高效的聚类方法都是被动的,用户往往期望参与到计算过程中,利用领域知识和经验进行交互式聚类.气味数据存在多种可能的聚类方式,用户需要对多组聚类结果进行比较,以确定最优的聚类.提出一个对气味数据进行交互式聚类的可视分析框架.首先使用人工神经网络结合多分子气味的高维化学信息和感官描述信息,计算一种低维度的气味嵌套表示;然后基于高斯混合模型的交互式聚类算法让用户对气味嵌套进行符合预期的聚类和重新聚类;最后用户可通过聚类可视化视图比较多组聚类结果的异同,增进对气味数据的理解,从而找到最佳的聚类方式.通过实验和2位领域专家的测试与评估,验证了该可视分析框架的有效性.Compared with images and sounds,data collected from scents can be more complex since most scents in nature are multi-molecular;and due to a lack of efficient representation methods,research progress for scents is slower.An important way of studying scents is clustering analysis,which can help understanding,recognition,synthesis,and design of scents.However,most of the efficient clustering methods are passive,while users desire involvement in computation by applying their domain knowledge and experiences to an interactive clustering process.In addition,since scent data can be clustered in multiple ways,users also need to compare the results to find the most appropriate clustering.Therefore,we present a visual analytics framework for interactively clustering scent data,which systematically comprises of the three steps for analyzing scents:an artificial neural network model is utilized to learn low-dimensional scent representations by combining scents’high dimensional chemical and perceptual information;a Gaussian mixture model-based interactive clustering algorithm enables users to restructure scent clusters according to their expectations;and multiple visualizations support comparison and analysis of clustering results to gain a deeper understanding of scent data and determine the most suitable result.The effectiveness of the proposed framework is proven by experimental analysis of clustering algorithms and two case studies where two domain experts provided positive feedback on user experience and functionalities.

关 键 词：气味化学感知嵌套 高斯混合模型 交互式聚类 聚类可视化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]