Linked Data makes possible the discovery of interesting connections between semantic entities that belong to di erent domains. This paper presents CityMUS, a web application that gives to the user the experience of a walk in the city with the most suitable soundtrack, on the base of the urban context. The application relies on a recommender system that searches for paths in a knowledge graph between nearby places and music composers, making use of a combination of DBpedia and domain-speci c datasets.
musical work can be relatively poor in DBpedia (e.g. classical music often misses
information about the composer) and we therefore rely instead on the data about
their composers for nding connections. The goal is therefore to nd paths in
the DBpedia graph that connect entities that are also described in rich datasets
such as 3cixty [
Datasets Selection and Interlinking. The 3cixty knowledge base [
The DOREMUS knowledge base [
X occ(ri)
j N j i
where ri corresponds to the ith resource of a path of length N and occ(ri) to
the number of its occurrences in all found paths. Given deepmax as the biggest
path depth (in our case, 7) and len(artist; poi) as the considered path length, we
de ne the similarity between a PoI and an artist using a combination of di erent
similarity measures [
1) 1)) ) (1 k)gen
We select 5 artists for each PoI with the highest similarity score. 3
CityMUS Application and Evaluation CityMUS Mobile Application. CityMUS is a mobile web application available at https://citymus.doremus.org. The app uses the geo-location API for getting the user position. The server generates then a playlist of tracks from the artists connected to the closest 3 PoIs, with a di erent weights according to their distance. The Spotify APIs are used in order to display and play the tracks (Figure 1.a). The user can see the path of the song that is currently played (Figure 1.c) and navigate the map for discovering the songs related to other PoIs (Figure 1.b).
Evaluation. The artists interlinking is evaluated by using as ground truth the one done by ISNI5, for which the mapping is already in the DOREMUS dataset (although not used in the interlinking task). The evaluation give us a precision of 0.987, a recall of 0.940 and a f-measure of 0.963. For PoIs interlinking, we consider as correct the ones in a 250 meters distance from the one retrieved through the Google Maps API (70% of the total).
About the path, we are interested to select speci c paths, which means maximising the gen function. We prune in this way lots of paths that contain very common resources (i.e. classes or reference to Nice as a city). In particular, the paths we select for CityMUS contain resources that occur in average in just 0.001% of all paths after the pruning, while the highest number of occurrences appears in the 0.04% of them.
Future developments will involve the interlinking of di erent nodes in the musical metadata (i.e. the work itself, the genre, the historical period) that can reveal more interesting connections. A user evaluation is also planned, by adding the possibility to rate the appropiateness of a song. 5 International Standard Name Identi er, http://www.isni.org
Acknowledgments. This work has been partially supported by the French National Research Agency (ANR) within the DOREMUS Project, under grant number ANR-14-CE24-0020.