Intro blurb Lorem ipsum dolor sit amet, consectetuer adipiscing elit. Pellentesque enim tortor, blandit vitae, dignissim sed, lobortis vitae, quam. Donec in enim. Curabitur vitae orci. Pellentesque lobortis arcu sit amet turpis. Sed sit amet massa sit amet mi condimentum porta. Cras urna risus, nonummy sed, ultricies ut, mollis vitae, lorem. Fusce ultrices metus eu tellus. Vestibulum ac risus a augue tincidunt tempor. In hac habitasse platea dictumst. Morbi varius. Donec eget nunc. Nam vel felis sit amet elit pulvinar molestie. Proin purus justo, vulputate sit amet, pretium vitae, dignissim at, ipsum. Mauris mattis mauris quis tellus.

 

Overview statistics

 

 

 

Count of Iconclass code occurrences, by average weight

The scatter plot above shows how many other codes typically exists alongside a given code, and the frequency of that code. Each code observation was assigned a weight equal to 1 / (total number of codes given to the poster); each poster has a total weight of 1. We can see that the most frequent code "birth control" (ie: condom) will typically be one of two codes assigned to the poster. We can also see that codes in class 3, Humans, typically have lower weight compared to codes in class 4, Culture. A more detailed analysis would be necessary to determine if this observation is statistically significant. One possible explanation may be that Cultural codes encapsulate multiple features in a single code, but Human codes are more discrete.

 

Temporal graphs

Stacked area graphs

 

Line graph

The graph below demonstrates how the total weight of each code is distributed over each year. The more common codes are more stable, allowing the dynamics of less common codes to be examined. It serves as a less rigorous burst analysis.

 

Temporal burst bar graphs

Sci2 was used for burst analysis of the Iconclass observations, and to generate multiple temporal bar graphs. The Iconclass codes were truncated to the specified depth for each graph.

 

 

Geospatial maps

Two maps were created using QGIS and some secondary data for country boundaries. Although they clearly show the physical distribution of the posters, no Iconclass information is given. Since the map uses both axes, any additional data must be encoded in the symbol. Such a design is possible, using bar or pie overlays for example, but this would quickly become messy and disorderly. If the data is aggregated by world region, an orderly map can be created, but it may be oversimplified and only useful as an introduction, summary, or footnote; not a primary visualization.

 

 

Networks

 

Heirarchies

 

v0: An hierarchical network is simulated by first generating rows consisting of each Iconclass code and its immediate parent (ie, truncate a single character from the right), and then using Sci2 to extract a coöccurrence network. Notice how there are no isolates, only pairs of nodes. However, some clusters did emerge where codes exist at multiple depths of the Iconclass hierarchy.

v1: This is a more structured attempt to construct a tree network using the Iconclass.org RDF service, and a Python script. Sci2 was used to generate a graphML file, and Gephi used for visual encoding and layout.

v2: Returning to v0 with the following changes. This time, new rows are created to fill gaps in the hierarchy. Then the count of each code is used for node size and nodes with a count of zero (ie, gap-fillers) are coloured grey.

v3: Same as v2, but the size scale matches the code count distribution (0 - 422). Gap-filler nodes disappear, and single occurrence codes look like barren, dead branches.

 

 

Semantic network

The goal of this network was inspired by the science of science map. What you see here is the top 10000 nodes of the full Iconclass hierarchy. The intent was to create the full hierarchy to be used as a base map for the AIDS poster data. Unfortunately, the computational power required was prohibitively high.

 

Bipartite network

Somewhat interesting. One can see which codes are shared among regions, but no noticeable trends or patterns.

 

Coöccurence networks

The following coöccurrence networks proved to be fruitful and interesting. The data for each was constructed from an SQL query, the network extracted using Sci2, and final rending with Gephi. Node data had to be joined using Gephi since Sci2 disregards it when extracting the network. The basic strategy for creating the input data is as follows

• Choose two attributes from the source data
• Aggregate by one
• Concatenate the other with a delimiter

Sample query:

 select p.id as id, 
     group_concat(ic_stem,"|") as ic_list
 from posters as p 
 join ic_obsv as o, 
     ic_codes as c 
 on p.id = o.id and c.ic_en_ = o.ic_en_;

The first network aggregates Iconclass codes, and concatenates poster id. This is an interesting result that could be developed further with colour partitions of the major categories. The problem with this direction of analysis is that it conflates multiple posters into a single node, making it difficult to encode year and origin data.

 

One of the primary benefits of this network is having each poster as a single node. This is a more concrete representation of the data. Also, clusters begin to emerge as a result of similarities in the codes used for each poster. The potential relevance of these clusters is that they represent emergent categories. Compare this to networks in the previous section, which merely mimic the Iconclass hierarchy. These clusters describe groups found in the data rather than groups imposed on the data.

 

Now that the data has been represented in a way that is both informative and satisfactory, additional attributes are encoded. The three following networks encode `region` as the node colour, and `year_from` as the node size. The first two show 1983 as the smallest and 2012 as the largest. The third reverses the size to emphasize early posters. Ultimately, there are benefits to both, and a possible future development of the visualization may include the ability to dynamically control this variable.

 

 

The network was then exported using the Sigma.js plugin of Gephi. Further customizations to this network are described in greater detail in the next section. The second image below shows the approximate boundaries of some emergent categories. The following observations can be made:

• North American posters frequently use faces. The use of close-up faces is a common marketing technique for emotional appeals.
• African posters are generally more instructive and informative, or community focused. One explanation may be the possibility that most of the posters in this region are funded or produced by organizations who emphasize direct dissemination of information, moral neutrality, and political correctness.
• European posters tend to be more provocative and sexually explicit.

 

 

---

Next >> Design