The traditional method for clustering or grouping media is to use a large training set of labeled data. However, the Making System multi-touchless approach moves away from the use of large hand labelled training datasets. Instead, allowing the user to find natural groups of similar content based upon a handful of ”seed” examples using two efficient data mining tools originally developed for text analysis: min-Hash and APriori. This means that the user guides the correct grouping of the diverse media by identifying only a small subset of contradictory items  e.g. web, video, image and emails. The pair-wise similarity between the media is then collapsed into a 2D presentation for displaying on the touch table. More recent work has looked at automatically attaching linguistic tags to images harvested from the internet .
In addition to showing groups of similar media content, a summary of the media can be produced. This can efficiently identify the frequently reoccurring elements within the data. When all the events are used, this can provide a overall gist or flavour of the activities. Furthermore, discriminative mining can be used where a subset of (positive) events are mined against another subset of (negative) examples. Here we can highlight events or content that are salient in the positive set, allowing trends to be identified that, for example, describe activity that is particular to a specific time of day or day of the week.
Automatically identify shared features that satisfy user query
Re-project and visualise the content based on learnt rules
I use a compact representation of the content to enable real time operation. I propose an Image Signature, this is similar to a Bag of Words histogram, where the frequency count of the features firing in the piece of content builds a histogram for each piece of media.
Image Signature Similarity It uses the data mining tool min-hash, this was originally designed for duplicate text analysis, and estimates the set overlap of the image signatures. To do this min-hash compares multiple min hashes based on many random permutations of overall vocabulary of features. and the pair wise similarity is computed between all image signatures, as shown by the left distance matrix below Where the white points indicate the closest match, ideally all the matches would be within the green boxes indicating the class classification.
Pushing and Pulling Image Signatures However, there are a number of false positives, this is to be expected as there is no training done at all, therefore to remove these false positives, the user will select a number of examples from a class, and mis-classified examples from the same class. The underlying program will identify feature similarities between the positive media with respect to the negative, using APriori data mining to find correlations in vast number of images and videos. The positive rules are then emphasised across the content. This will update the distance matrix, resulting in new visual clusters, with the positive examples closer together.
Example from the Youtube Dataset Results The image belowshows the grouping for a single class from the YouTube dataset, based on the connecting blue lines
The user wishes to bring the two large circles of media together, so the user would select a number of pieces from each circle the approach brings the two circle into a single group.