A Visual Analytics Framework for Interactively Clustering Scent Data

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.