![\"\"](\"https:\/\/miro.medium.com\/v2\/resize:fit:700\/1*1Z2sqA_uyI7E3vvUTz9uUw.jpeg\")
<\/figure>Machine learning is a subset of artificial intelligence that employs statistical algorithms and other methods to visualize, analyze and forecast data. Generally, machine learning is broken down into two subsequent categories based on certain properties of the data used: supervised<\/strong> and unsupervised<\/strong>.<\/p>\n Supervised learning algorithms refer to those that require training datasets with labels attached to the features. Unsupervised learning algorithms, on the other hand, make use of mathematical formulas that find patterns in unlabelled data, often by identifying clusters of data with similar (but independent) features. In this article, we will discuss different clustering algorithms and how to evaluate their results. Let\u2019s get started!<\/p>\n Clustering (sometimes referred to as cluster analysis<\/strong>) is an unsupervised machine learning technique used to identify and group similar data points within a larger, unlabelled dataset. It refers to the process of finding a structure or pattern inside an otherwise unstructured dataset.<\/p>\n The four main types of clustering include:<\/p>\n In this article we will explore K-means clustering, hierarchical clustering, and DBSCAN, as these are some of the most common (and effective) methods used currently, but it\u2019s good to be aware that there are other methods out there as well.<\/p>\n K-means clustering is an unsupervised machine learning algorithm that groups unlabeled data into K-means clustering starts by taking In order for K-means clustering to be effective, however, it is imperative to first determine the optimal value for Hierarchical clustering is another type of unsupervised clustering algorithm, in which we create a hierarchy of clusters <\/strong>in the form of a tree, also referred to as a dendrogram<\/a>. <\/strong>Hierarchical clustering also automatically finds patterns by dividing data into There are two main approaches to hierarchical clustering: agglomerative<\/strong><\/a> and divisive<\/strong><\/a>. In agglomerative clustering, we consider each data point as a single cluster, and then combine these clusters until we are left with one group (the full dataset). Divisive hierarchical clustering, on the other hand, begins with the whole dataset (considered as one single cluster), which is then partitioned into less similar clusters until each individual data point becomes its own unique cluster.<\/p>\n DBSCAN<\/a> stands for density-based spatial clustering of application with noise<\/strong>. DBSCAN clustering works upon a simple assumption that a data point belongs to a cluster if it is closer to many<\/em> data points of that cluster, rather than any single point. It requires two parameters for dividing data into groups: One of the biggest advantages of DBSCAN clustering is that it is very robust to outliers and doesn\u2019t require information about the cluster size for training.<\/p>\n Struggling to track and reproduce complex experiment parameters? Artifacts are just one of the many tools in the Comet toolbox to help ease model management. Read our PetCam scenario to learn more<\/a>.<\/p><\/blockquote>\n<\/div>\n<\/div>\n<\/div>\n\n\n\n We are going to make use of the K-means clustering algorithm to cluster different iris flower species into clusters, using the famous iris dataset. To determine the correct number of clusters we will make use of the elbow method<\/strong><\/a>. The dataset we are using is located in the sklearn dataset class.<\/p>\n The elbow method has given us an optimal value of Now that we have our labels and predictions let\u2019s evaluate this model to find out well it performed!<\/p>\nWhat is Clustering?<\/h1>\n
\n
1. K-Means Clustering<\/h2>\n
k<\/code> number clusters, where k<\/code> is a user-defined integer. K-means is an iterative algorithm that makes use of cluster centroids to divide the data in a way that groups similar data into groups.<\/p>\nk<\/code> random points, and marks these points as centroids<\/a> of k<\/code> clusters. It then calculates the Euclidean distance<\/a> for each remaining data point from each of those centroids and assigns each data point to its closest cluster (based on the Euclidean distance from the centroid). Once a new point is added to the cluster, it recalculates the centroid by taking the mean of all the vectors inside the group, and then recursively calculates distance again. Then, the new centroid is recalculated, and this is repeated until all data points are assigned to a cluster.<\/p>\nk<\/code>. There are various techniques for doing so, but one of the most effective is through simple data visualization. We will also cover a couple of other methods, however, in examples later on.<\/p>\n![\"\"](\"https:\/\/miro.medium.com\/v2\/resize:fit:700\/1*_xGJdgLJapu9OPTcIjvgbw.png\")
<\/figure>2. Hierarchical Clustering<\/h2>\n
n<\/code> clusters. However, in this case, there is no need to define the number of clusters (the algorithm will determine this for you).<\/p>\n![\"\"](\"https:\/\/miro.medium.com\/v2\/resize:fit:700\/1*6lTcB15YXA05dIizH9QyjQ.png\")
<\/figure>3. DBSCAN Clustering<\/h2>\n
epsilon<\/code> and min_points<\/code>. epsilon<\/code> specifies how close one point should be to another in order to consider it part of the cluster, while min_points<\/code> determines the minimum number of data points required to form a cluster.<\/p>\n![\"\"](\"https:\/\/miro.medium.com\/v2\/resize:fit:400\/0*jONOcwF8O9hzf8nc.png\")
<\/figure>Building a Clustering Model<\/h1>\n
from sklearn import datasets\ndf = datasets.load_iris()\nX = df['data']\ny = df['target'] # not needed for clustering<\/span><\/pre>\nX<\/code> will contain information about sepal width, sepal length, and petal width whereas y<\/code> will contain information regarding the type of flower species. We will only use X<\/code>, and try to divide the dataset into different flower species clusters using K-means instead. Below, we use the elbow method to find the value of k<\/code>.<\/p>\nfrom sklearn.cluster import KMeans<\/span>wcss = []\nfor i in range(1, 11):\n kmeans = KMeans(n_clusters = i, init = 'k-means++',\n random_state = 42)\n kmeans.fit(X)\n wcss.append(kmeans.inertia_)\nplt.plot(range(1, 11), wcss)\nplt.title('The Elbow Method')\nplt.xlabel('Number of clusters')\nplt.ylabel('WCSS')\nplt.show()<\/span><\/pre>\n![\"\"](\"https:\/\/miro.medium.com\/v2\/resize:fit:491\/1*XZKP9FfwUY9EpsLKKtaSBA.png\")
<\/figure>k<\/code> that is 3. Let\u2019s use this value to build a model.<\/p>\nfrom sklearn.cluster import KMeans\nKmean = KMeans(n_clusters=3)\nKmean.fit(X)<\/span>## Predictions\ny_pred = Kmean.predict(X)<\/span><\/pre>\nEvaluation Metrics For Clustering-Based Models<\/h1>\n
1. Silhouette Score<\/strong><\/h2>\n
\n
k<\/code> (see here for more<\/a>).<\/li>\nk<\/code> is incorrect.<\/li>\nfrom sklearn.metrics import silhouette_score\nsilhouette_score(X,y_pred)\n--------------------------------\n0.55<\/span>\"\"\"\nsilhouette score is 0.55 which is acceptable and shows clusters are not overlapping.\n\"\"\"<\/span><\/pre>\n2. Calinski Harabaz Index<\/strong><\/h2>\n
\n
from sklearn.metrics import calinski_harabasz_score\ncalinski_harabasz_score(X,y_pred)\n---------------------------------------------\n561.62775<\/span><\/pre>\n3. Davies Bouldin index<\/strong><\/h2>\n