A visual analytics approach to model learning

Abstract

The process of learning models from raw data typically requires a substantial amount of user input during the model initialization phase. We present an assistive visualization system which greatly reduces the load on the users and makes the process of model initialization and refinement more efficient, problem-driven, and engaging. Utilizing a sequence segmentation task with a Hidden Markov Model as an example, we assign each token in the sequence a feature vector based on its various properties within the sequence. These vectors are then clustered according to similarity, generating a layout of the individual tokens in form of a node link diagram where the length of the links is determined by the feature vector similarity. Users may then tune the weights of the feature vector components to improve the segmentation, which is visualized as a better separation of the clusters. Also, as individual clusters represent different classes, the user can now work at the cluster level to define token classes, instead of labelling one entry at time. Inconsistent entries visually identify themselves by locating at the periphery of clusters, and the user then helps refine the model by resolving these inconsistencies. Our system therefore makes efficient use of the knowledge of its users, only requesting user assistance for non-trivial data items. It so allows users to visually analyze data at a higher, more abstract level, improving scalability. INTRODUCTIONWith the tremendous growth in physical and online data collection technology, we are now experiencing an explosion of digital information. Since a large amount of these data are unstructured, various machine learning techniques have been developed to assign structure to these data to make them machine readable. This process can allow the machine to reason with and draw insight from data almost automatically. However, all such tasks depend heavily on large amounts of user-tagged data as the starting point, and use various semi-supervised learning methods [19]. Due to the high user input required, such tagged data is difficult to construct. Further, data is dynamic, and as a dataset grows and changes, we might need to supplement the tagged data from time to time. We propose to make this task simpler and interactive by designing a system where the user can obtain a visual overview of the dataset, and in that visual interface only tags those data elements that the system cannot easily resolve itself.One crucial idea behind our system is that given good feature vectors to represent each data point, points that are similar will be close-by in the feature vector space. Here, we mean data-points which though rich in semantics, do not have an explicit highdimensional feature vector automatically attached to them. In such cases we need to design feature vectors to represent the semantics and structure of the data-points. We aim to achieve this in our system by designing feature vectors which encompass a data point's structure, context, and location in the dataset. If some s...

An Affordance-Based Framework for Human Computation and Human-Computer Collaboration

“…In presenting a preliminary framework for describing and comparing systems involving human and machine collaborators, we aspire to lay the foundation for a more rigorous analysis of the tools and approaches presented by our field, thereby paving the way for the construction of an increasingly robust understanding of analytical reasoning and how to best support insight generation. [41] CrowdSearch [80] ParallelTopics [23] Dissimilarity [50] VH+ML [28] Implicit tagging [62] reCAPTCHA [77] VizWiz [10] Phetch [74] ESP Game [73] KissKissBan [33] LabelMe [59] Ka-captcha [21] PeekABoom [76] MRI [12] iView [83] iVisClassifier [18] Saliency [37] RP Explorer [3] DimStiller [36] WireVis [46] Action trails [65] NetClinic [47] Trajectories [4] Risk assessment [51] Automatic transfer functions [57] MDX [66] Automated+viz [68] CzSaw [39] Fold.it [20] HRI scripts [17] Animated agents for VR [56] VA Model-learning [29] EyeSpy [6] MonoTrans2 [35] CastingWords [16] Click2Annotate [15] Wrangler [40] Soylent [8] Crowdsourced solutions [67] Crowdsourced design [81] Stress OutSourced …”

“…2a), computational methods for manipulating large datasets have been used to help users navigate and make sense of massive text corpora [23]. It has also been utilized to refine classification models and performing dimension reduction [18,29,51], interactively cluster data [4], and automatically extract transfer functions from user-selected data [57]. It has been used to suggest informative data views [83], and even to help users externalize and understand their own insight generation process [15,39,41,46,65].…”

“…The study of general human-computer collaboration offers a plethora of examples of successful human/machine teams [15,23,37,39,41,46,50,65,66,68,83]. Developments in supervised machine learning in the visualization community present several vetted techniques for human intervention into computationally complex tasks [3,4,12,18,28,29,36,47,51,57]. The emerging field of human computation inverts the traditional paradigm of machines providing computational support for problems that humans find challenging, and demonstrates success using aggregated human processing power facilitated by machines to perform difficult computational tasks such as image labeling [21,33,73,74], annotating audio clips [44,49], and even folding proteins [20].…”

Leveraging Spatial Abstraction in Traffic Analysis and Forecasting with Visual Analytics

“…Much attention has been given to classification models [25][30] [69], including decision trees [24]. A classifier may be built for assigning new objects to previously obtained clusters [3] [26]. Visual analytics techniques also support derivation of linear trend models from multivariate data [32], regression models [51], and time series models [13] [34].…”

A visual analytics framework for spatio-temporal analysis and modelling

“…Garg et al (2008) suggest a framework where a classifier is built by means of machine learning methods on the basis of positive and negative examples (patterns) provided by the user through an interactive visual interface; the user finds the patterns using visualisations. Garg et al (2010) describe a procedure in which clusters of documents are built by combining computational and interactive techniques, then a classifier for assigning documents to the clusters is automatically generated, and then the user refines and debugs the model. This is similar to what is suggested by Andrienko et al (2009) for analysis of a very large collection of trajectories: first, clusters of trajectories following similar routes are defined on the basis of a subset of trajectories, second, a classification model is built and interactively refined, and, third, the model is used to assign new trajectories to the clusters.…”