Description and characterization of place properties using topic modeling on georeferenced tags

Figures & data

Figure 1. Study area within inner London.

Table 1. Remaining numbers of Flickr images and users after applying each task of the data filtering.

Function	Images	Contributors
Original dataset	3,105,544	49,130
Accuracy filtering	1,047,003	31,092
Bulk-upload filtering	839,822	31,080
Camera generated contents (either titles or tags)	571,241	30,377
Inactive users	503,536	8143
Prolific users	404,329	8060
Null tags	371,752	7753

Figure 2. Number of users taking photographs in each grid cell.

Table 2. Median corpus distance and number of cells per topic as a function of the number of topics for different grid resolutions.

Resolution (m)	No. of topics	Median no. of cells	Median corpus distance
50	20	678.0	2.32
	40	349.0	3.07
	60	228.5	3.36
	80	166.5	3.59
	100	131.0	3.81
250	20	64.0	2.09
	40	32.5	2.80
	60	18.5	3.13
	80	14.0	3.30
	100	11.0	3.44
500	20	16.0	1.90
	40	10.0	2.41
	60	5.0	2.69
	80	4.0	2.96
	100	3.0	3.27
1000	20	5.5	1.52
	40	2.0	2.21
	60	2.0	2.46
	80	1.0	2.75
	100	1.0	2.93

Figure 3. Comparison between clusters of 40 topics with respect to the grid resolution. (a) 1 km vs. 500 m; (b) 500 m vs. 250 m; (c) 250 m vs. 50 m.

Figure 4. Change in median corpus distance for different number of topics with respect to grid resolution.

Figure 5. Corpus distance for topics associated with different numbers of cells at a resolution of 500m.

Figure 6. Number of topics associated with the 500m cells for each implementation of the model.

Figure 7. Topic coherence value for labeled and unlabeled topics.

Figure 8. Labeled topics (numbered cells) and example topics, labels, and tags (size as a function of probability).

Figure 9. Map of London describing users’ perception of the space as places.