Putting Recommendations on the Map -- Visualizing Clusters and Relations

Putting Recommendations on the Map – Visualizing Clusters and Relations Emden Gansner, Yifan Hu, Stephen Kobourov, Chris Volinsky AT&T Labs – Research, 180 Park Ave, Florham Park, NJ 07932 {erg, yifanhu,skobourov,volinsky}@research.att.com ABSTRACT For users, recommendations can sometimes seem odd or counterintuitive. Visualizing recommendations can remove some of this mystery, showing how a recommendation is grouped with other choices. A drawing can also lead a user's eye to other options. Traditional 2D-embeddings of points can be used to create a basic layout, but these methods, by themselves, do not illustrate clusters and neighborhoods very well. In this paper, we propose the use of geographic maps to enhance the denition of clusters and neighbor- hoods, and consider the eectiveness of this approach in vi- sualizing similarities and recommendations arising from TV shows and music selections. All the maps referenced in this paper can be found in www.research.att.com/~volinsky/ maps. 1. INTRODUCTION Information visualization techniques are often essential in helping to make sense out of large data sets. High- dimensional data can be visualized as a collection of points in 2-dimensional space using principal component analysis [12], multidimensional scaling [15], force directed algorithms [6], or non-linear dimensionality reduction like LLE/Isomap [20, 22]. These embedding algorithms tend to put similar items next to each other. Visual examination often suces to identify the presence of clusters. Sometimes, however, the clusters are not as easy to see and additional visual clues are needed to highlight them. One possibility is to use cluster analysis algorithms, such as k-means or hierarchical clus- tering algorithms [11, 16] to explicitly dene clusters. The points and labels can then be colored based on the cluster- ing. While in small examples it is possible to convey the cluster information just with the use of colors and proxim- ity, this becomes dicult to do with large data. Common problems include dense clusters where labels overlap each other and clusters that lack clearly dened boundaries. In this paper we propose the use of maps as a way to achieve this explicit visual denition of clusters. There are several reasons that such a representation can be more use- ful. First, by explicitly dening the boundary of the clusters and coloring the regions, we make the clustering information clear. Second, as most dimensionality reduction techniques lead to a 2-dimensional positioning of the data points, a map is a natural generalization. Finally, while graphs, charts, and tables often require considerable eort to comprehend, a map representation is more intuitive, as most people are very familiar with maps and even enjoy carefully examin- ing maps. Applying this approach to a data set containing show-show similarities between 1000 TV programs results in the map in Figure 1, which conveys clustering information much better.1 We have considered several dierent approaches for gen- erating maps, depending on (1) how we obtain the 2D posi- tions for the data points, (2) how we cluster the data points, and (3) how we represent the resulting layout and clustering as a map. Depending on the application, some choices are more suitable than others. Consider the case when the data points are TV programs and the goal is to visualize TV- Land. In this case, it is highly desirable to obtain a map in which similar programs are placed close to each other and, even better, grouped in countries such as Sportsitania, Newsistan, ToddlerSprawl, etc. Once the data is represented as countries on a map, recommendations can be visualized statically and interactively. The map metaphor becomes more powerful as a user be- comes familiar with the canonical map layout. Humans are comfortable with map-related concepts: items within a country are similar to each other; areas separated by a mountain range are dicult to connect; islands might have atypical qualities, etc. We hope through this work to ex- tend the familiar map metaphor to the world of recommen- dations. Thus, by putting items like TV shows on a map, we can borrow map-related cognitive concepts. In a user- driven mode, a personalized heat map is generated, where regions of low interest are colored with cool colors and re- gions containing highly recommended shows are colored with a hot color. These maps are generated dynamically based on user preference and the available TV shows at this mo- ment in time. The user can interactively explore the map to nd related shows in the same region, or in neighbor- ing regions. Building roads between regions, visiting is- lands and traveling cross-country are all metaphors that can have meaning in the recommendation space. In a user- passive mode, recommendations are highlighted on the map. These maps help the user (as well the designers of the rec- ommender system) understand why certain shows are rec- ommended. By

…(Full text truncated)…

This content is AI-processed based on ArXiv data.