Spatial Locality

Spatial locality of user interest is about determining whether people in the same geographical region tend to receive similar notification content. To carry out our analysis we take the following approach. We define a notification message to be locally shared if at least two users in the same cluster receive the notification. We compare the degree of sharing using geographical clustering and four random clusterings. In the geographical clustering case, clients in the same city are clustered together. In the random clustering case, clients are clustered randomly with the cluster size being the same as in geographical clustering. We obtained the geographical location of users using a registration database which contains zip code information for each user. The zip code information is not clean -- some users supplied invalid zip codes; we filter out all the zip codes that are not 5 digits. 14% of the users supplied such invalid zip codes. In the remaining entries, it is still possible to have zip codes that do not match the actual user location, but the fraction is likely to be small. Furthermore, when computing the degree of local sharing, we exclude the cities to which fewer than 100 notification messages were sent over the course of the week. As shown in Figure 7, clients residing in the same city have significantly more sharing in notification content compared to the clients picked at random. We also compared geographical clustering with three other random clusterings and observed similar results. The higher degree of sharing in notification messages for clients in the same geographical region indicates that localized services are popular for notification services. For example, people living in New York are interested in receiving notification messages about weather or events in New York. The geographical locality in notification content implies that placing servers (i.e., either notification server replicas or servers in an overlay network that provide application-level multicast) close to popular geographical clusters can be useful in reducing network load.

  

Figure 7: Compare the local sharing between random clients and clients that are geographically close together.