ConceptCloud is a exible interactive tool for exploring, visualising, and analysing semi-structured data sets. It uses a combination of an intuitive tag cloud visualisation with an underlying concept lattice to provide a formal structure for navigation through a data set. ConceptCloud 2.0 extends the tool with an integrated map view to exploit the geolocation aspect of data. The tool's implementation of exploratory search does not require prior knowledge of the structure of the data or compromise on scalability, and provides seamless navigation through the tag cloud and the map viewer.
Semi-structured data such as product reviews or event logs contains embedded meta information that describe semantic elements, but does not conform to the formal structure of a data model. Many such data sets incorporate some geolocation aspect, for example the location of the vineyard for wine reviews, or the accident location for tra c data. This paper describes a map extension to ConceptCloud, a visualisation and exploration tool for semi-structured data sets.
ConceptCloud uses a formal concept lattice generated from the input data as
the underlying navigation structure [
The software allows the user to iteratively select an attribute or object tag in the tag cloud, and the tool adjusts the tag cloud to display all other tags attached to objects possessing the selected attribute tag(s). This is achieved by maintaining a focus concept from which a tag cloud is created.
Copyright c 2019 for this paper by its authors. Copying permitted for private and academic purposes.
Formally, the focus concept c := hO; Ai is the concept whose extent is the set of objects that share the set of currently selected attributes, F A, within the tag cloud. The focus concept can be further re ned by iteratively adding elements to F . When an additional attribute is added to F, the focus concept is updated by computing the meet of the current focus concept c and the concept introduced by the additional attribute. This strategy obviates the need to compute an entire concept lattice or implication base, which is not feasible for large data sets.
The explorative search process corresponds to the process of stepping through
a concept lattice, wherein the selection of an attribute moves the user to the point
in the lattice where all linked objects contain that attribute. As more attributes
are selected, the user moves further down the lattice. The reverse of this process
is also possible. ConceptCloud has been used in many domains as an e ective
method for knowledge discovery using tag cloud visualisation and navigation [
With the rapid increase in always-on, embedded GPS devices, geolocation
rich data is becoming more prevalent. This abundance presents a new
opportunity for knowledge discovery by exploring the geolocation aspect of the data,
but it also demands a di erent visualisation approach [
Historically, maps have most commonly been used to visualise geolocation information. However, while maps are immediately useful for visualising the geolocation aspect of data, fully integrating maps into an interactive tag cloud explorer requires that the exploration also be driven from the map. Thus an important function of this tool is to make the exploration of the data work bidirectionally, i.e., to update the map when changes are made in the tag cloud and vice versa. Moreover, the rate at which such data is being generated is increasing, resulting in ever larger sets. This requires that the visualisation and exploration tool be highly scalable in order to process large data sets. With maps providing a time-tested method of exploring the geolocation aspect of data, and tag clouds providing an e ective method for facilitating knowledge discovery and data visualisation for semi-structured data, the overall goal of the ConceptCloud 2.0 extension described here is therefore to merge these two proven methods into an integrated and scalable system. 2
ConceptCloud 2.0 generates map pins based on the objects' geolocation attribute, where available, and uses the Google Maps JavaScript API1 to render them on a fully interactive map. However, to prevent the generated pins from occluding each other, ConceptCloud 2.0 clusters them automatically within a shifting geolocational tolerance, based on the current zoom level of the map, and displays a count of the pins in each cluster.
Fig. 1 (Left) shows this view for an example crime data set which contains, for each crime, its category (e.g., Theft ) or sub-category (e.g., Stock theft ) and the name and location of the police station where the crime was reported.
The pins then function the same way as the tags in the tag cloud. The user
can thus explore the data by selecting an attribute in the tag cloud, which will
update the cloud as before [
The pin clusters mentioned above also have a formal correspondence in the
concept lattice in the form of biclusters. Ordinarily, a bicluster is de ned as a
pair (A,B) of inclusion-maximal sets of objects and attributes such that almost
all objects in A have almost all attributes in B. This technique has been
implemented and applied successfully to mine numeric data sets using triadic formal
concept analysis [
Fig. 2 shows the new tag cloud generated from biclusters from the central Cape Town area. The user can generate this bicluster tag cloud by right-clicking either a single or multiple pin clusters as well as individual pins. This bicluster tag cloud corresponds to a smaller lattice created from the original lattice, with its own focus concept, which can be navigated through independently of the main lattice. ConceptCloud 2.0 adopts a client-server architecture. Its server is written in Java and uses the Play! framework2 along with a PostgreSQL database to generate the concept lattice. The amount and nature of pre-processing that is required in order for ConceptCloud to consume data depends on the source. Typically some standard data cleanup and NLP techniques are applied, such as stemming and the removal of stop words. ConceptCloud accepts a JSON le of objects consisting of attributes only. Along with this, the user needs to specify the attribute types present in the object, in particular, the geolocation attribute which must be of format \lat, long".
The geolocation attribute of the objects are used to display the marker objects in the map viewer, although they also appear in the tag cloud as latitude and longitude pairs. When considering the underlying context table, each
lat-long pair is an attribute, with each object having at most one geolocation attribute.
The map and marker objects in the browser-based client are populated by a single specialised server call. The size of the visible map in the Google Map viewer is dependent on the size of the viewer window and the zoom level. Using the dimensions of this viewer, we then calculate the visible radius, and use this radius together with the zoom level and centre coordinates of the map viewer to only retrieve the map pins that are visible to the user. We keep track of movement in the map viewer and make additional server calls whenever the map is moved to a new location.
ConceptCloud allows the user to pre-con gure the data attributes they wish to appear in each map pin object, allowing for a far smaller, and therefore more responsive server call to create each marker. The server returns a set of objects, containing at least an identi er and the geolocation, from which a Google Maps marker object is created and populated with the pre-con gured attributes of that object. If desired and available, the user can populate a tool-tip text window of the marker with additional meta-data and links to external resources involving the object.
In order to maintain a single focus information retrieval navigation
algorithm [
Computing the disjunction of two or more objects involves assigning a new
meta-tag to the selected objects, and in doing so, generating a new temporary
lattice on the y. This new lattice consisting of the merged objects becomes the
subject of the new tag cloud window, and the user is free to explore the desired
objects without introducing concept broadening [
ConceptCloud 2.0 functions e ectively as a data exploration and visualisation
tool. A user study [
In our testing it became apparent that the Google Maps API could be a limiting factor, since it adds marker objects to the map individually before being clustered. This process is highly memory intensive, and the web browser becomes unresponsive when exceeding 250 000 markers. With the previously discussed front-end and back-end optimisation implemented, these limitations were overcome. The ConceptCloud software itself proved to be highly scalable, processing a full 2.5 million object data set without any error or decrease in performance. With regard to future work, research is being done around abstracting the map viewer, essentially a 2D Cartesian plane, to support other metric space visualisations.