官术网_书友最值得收藏!

Clusters as Neighborhoods

Until now, we have explored the concept of likeness being described as a function of Euclidean distance – data points that are closer to any one point can be seen as similar, while those that are further away in Euclidean space can be seen as dissimilar. This notion is seen once again in the DBSCAN algorithm. As alluded to by the lengthy name, the DBSCAN approach expands upon basic distance metric evaluation by also incorporating the notion of density. If there are clumps of data points that all exist in the same area as one another, they can be seen as members of the same cluster:

Figure 3.1: Neighbors have a direct connection to clusters

In the preceding figure, we can see four neighborhoods. The density-based approach has a number of benefits when compared to the past approaches we've covered that focus exclusively on distance. If you were just focusing on distance as a clustering threshold, then you may find your clustering makes little sense if faced with a sparse feature space with outliers. Both k-means and hierarchical clustering will automatically group together all data points in the space until no points are left.

While hierarchical clustering does provide a path around this issue somewhat, since you can dictate where clusters are formed using a dendrogram post-clustering run, k-means is the most susceptible to failure as it is the simplest approach to clustering. These pitfalls are less evident when we begin evaluating neighborhood approaches to clustering. In the following dendrogram, you can see an example of the pitfall where all data points are grouped together. Clearly, as you travel down the dendrogram, there is a lot of potential variation that gets grouped together since every point needs to be a member of a cluster. This is less of an issue with neighborhood-based clustering:

Figure 3.2: Example dendrogram

By incorporating the notion of neighbor density in DBSCAN, we can leave outliers out of clusters if we choose to, based on the hyperparameters we choose at runtime. Only the data points that have close neighbors will be seen as members within the same cluster, and those that are farther away can be left as unclustered outliers.

主站蜘蛛池模板: 观塘区| 渝北区| 四川省| 石泉县| 金阳县| 溧水县| 台州市| 镇原县| 玛曲县| 威远县| 博爱县| 普兰县| 临沭县| 台江县| 宁德市| 墨竹工卡县| 朝阳区| 江达县| 舞阳县| 仁布县| 京山县| 略阳县| 东乌珠穆沁旗| 谢通门县| 通辽市| 大丰市| 襄樊市| 盐边县| 确山县| 井冈山市| 太白县| 清镇市| 石棉县| 冕宁县| 莱阳市| 肇庆市| 双辽市| 台南县| 湘阴县| 丰县| 榆林市|