In-class Exercise 7
Hierarchical clustering
proxmat <- dist(shan_ict, method = "euclidean") # euclidean can be used as a parameter for shiny
hclust_ward <- hclust(proxmat, method = "ward.D")
groups <- as.factor(cutree(hclust_ward, k=6))Don’t hardcode k. Instead the user should be the one to decide based on the graph.
Append to the geospatial data
shan_sf_cluster <- cbind(shan_sf, as.matrix(groups)) %>% rename(CLUSTER=as.matrix.groups.) %>% select(-c(3:4, 7:9)) %>% rename(TS = TS.x)The dendrogram
plot(hclust_ward, cex = 0.6)
rect.hclust(hclust_ward, k = 6, border = 2:5)Cluster map
qtm(shan_sf_cluster, "CLUSTER")SKATER uses distance to determine relationships. You then use your cluster data to add weights to each edge, and then partition the graph into individual trees for nodes with highest dissimilarity. But each point will require at least a single edge.
If you have polygons it uses centroids.
Spatially constrained clustering: SKATER method
shan.nb <- poly2nb(shan_sf) # spdep supports sf in its newest version now (as_Spatial is no longer necessary)
summary(shan.nb)plot(st_geometry(shan_sf), border = grey(.5))
pts <- st_coordinates(st_centroid(shan_sf))
plot(shan.nb, pts, col = "blue", add = TRUE)Computing minimum spanning tree
Calculating edge costs
lcosts <- nbcosts(shan.nb, shan_ict)Incorporating these costs into a weights object
shan.w <- nb2listw(shan.nb, lcosts, style = "B") # (use "B" because you want to know neighbour or not neighbour, do not use this as argument)
summary(shan.w)Computing MST
shan.mst <- mstree(shan.w)Visualising MST
plot(st_geometry(shan_sf), border = gray(.5))
plot.mst(shan.mst, pts, col = "blue", cex.lab = 0.7, cex.circles = 0.005, add = TRUE)Plotting SKATER tree
skater.clust6 <- skater(edges = shan.mst[,1:2], data = shan_ict, method = "euclidean", ncuts = 5)plot(st_geometry(shan_sf), border = gray(.5))
plot.mst(skater.clust6, pts, cex.lab = 0.7, groups.colors = c("red", "green", "blue", "brown", "pink"), cex.circles = 0.005, add = TRUE)Visualising the clusters in choropleth map
groups_mat <- as.matrix(skater.clust6$groups)
shan_sf_spatialcluster <- cbind(shan_sf_cluster, as.factor(groups_mat)) %>% rename(`skater_CLUSTER`=`as.factor.groups_mat.`) # use as.factor to convert to numerical form so it can be sorted
qtm(shan_sf_spatialcluster, "skater_CLUSTER")Spatially constrained clustering: ClustGEO method
cr <- choicealpha(proxmat, distmat, range.alpha = seq(0, 1, 0.1), K = 6, graph = TRUE)Saving output
clustG <- hclustgeo(proxmat, distmat, alpha = 0.2)
groups <- as.factor(cutree(clustG, k = 6))
shan_sf_clustGeo <- cbind(shan_sf, as.matrix(groups)) %>% rename(`clustGeo`=`as.matrix.groups.`)Visualising
qtm(shan_sf_clustGeo, "clustGeo")Computing spatial distance matrix
dist <- st_distance(shan_sf, shan_sf)
distmat <- as.dist(dist)Interactive cluster plot
ggparcoord(data = shan_sf_clustGeo, columns = c(17:21), scale = "globalminmax", alphaLines = 0.2, boxplot = TRUE, title = "") + facet_grid(~clustGeo) + theme(axis.text.x = element_text(angle = 30))The reason to compare multiple cluster maps side by side is to check for whether a variable is correlated with another.