In DBSCAN, each point is initially classified as either a core, border, or noise point. A core point has at least 'minPts' neighbors within the 'eps' radius, forming the core of a cluster . Border points lie within the 'eps' radius of core points but don't meet the 'minPts' threshold themselves, aiding in connecting clusters but not forming their cores. Noise points are those which are neither core nor border points, effectively being isolated from clusters . DBSCAN examines each unvisited point, marking it as a core point if it meets the neighbor condition, expanding clusters by connecting density-reachable points, and subsequently adjusting points' classifications . This methodology enables it to delineate well-separated clusters and exclude noise effectively .

DBSCAN can perform poorly in situations where there are clusters of varying densities or in high-dimensional spaces. In cases of varying densities, DBSCAN may struggle to differentiate between dense cluster borders and noise, as it applies a uniform density threshold across the dataset . In high-dimensional spaces, the distance metric may become less meaningful due to the curse of dimensionality, making it difficult for DBSCAN to accurately determine the neighborhood (eps distance) of points .

Internal measures such as cohesion and separation provide crucial insights into the quality of clusters formed by DBSCAN. Cohesion, typically measured by the within-cluster sum of squares (WSS), reflects how closely related the data points within a cluster are, indicating cluster compactness . Separation, measured by the between-cluster sum of squares (BSS), assesses how distinct a cluster is from other clusters. For DBSCAN, these metrics can indicate how well the density-based clustering maintains distinct, cohesive groupings, with optimal clustering reflected in high separation and low cohesion scores . Using these measures allows for the tuning of DBSCAN parameters to optimize clustering performance according to the desired balance of cohesion and separation .

Changing the 'eps' parameter in DBSCAN affects the radius of the region considered when determining the neighborhood of points, while 'minPts' determines the minimum number of neighbors required to form a cluster . Increasing 'eps' can merge smaller clusters into larger ones, or include noise points within clusters, whereas a smaller 'eps' may fragment existing clusters. Similarly, increasing 'minPts' may result in fewer, larger clusters by raising the density requirement for core points, whereas decreasing it can lead to more, but possibly less significant, clusters. Choosing these parameters impacts DBSCAN's sensitivity to cluster size and noise, necessitating careful selection based on the dataset's characteristics .

DBSCAN differs from k-means in that it does not require a predefined number of clusters; instead, it identifies clusters based on the density of data points. This makes DBSCAN particularly effective for identifying clusters of arbitrary shapes and handling noise, as it can effectively isolate noise points that do not belong to any cluster . Conversely, k-means assumes spherical clusters and can struggle with datasets that contain noise or clusters of varying shapes and densities, requiring prior knowledge of the number of clusters .

The critical hyperparameters of DBSCAN are 'eps' and 'minPts.' The 'eps' parameter defines the radius within which to search for neighboring data points, while 'minPts' specifies the minimum number of points required to form a dense region, making it a core point . These parameters directly influence the clustering process by determining the size and shape of clusters: smaller 'eps' values can result in more, smaller clusters, while larger 'eps' values can lead to fewer, larger clusters and noise .

DBSCAN classifies a point as a core point if it has at least 'minPts' other points within its 'eps' neighborhood. A border point falls within the 'eps' distance of a core point but does not satisfy the 'minPts' condition itself. A noise point is neither a core nor a border point, as it lacks sufficient neighboring points . These classifications are based on local density, allowing DBSCAN to structure clusters around dense regions while leaving sparse areas as noise .

To handle DBSCAN's limitations with datasets featuring varying densities, one strategy is to employ an adaptive version of DBSCAN, where 'eps' is dynamically adjusted based on local point densities. This can mitigate issues of uniform density requirements across diverse clusters . For high-dimensional data, dimensionality reduction techniques such as PCA (Principal Component Analysis) can be applied prior to clustering with DBSCAN to reduce the noise and make the distance metric more meaningful . Additionally, using alternative distance metrics that are better suited to high-dimensional spaces, such as cosine similarity, can improve DBSCAN's performance by providing more relevant distance calculations in such contexts .

The silhouette coefficient combines cohesion and separation into a single metric, allowing for an assessment of how well-separated a cluster is from others while also considering how compact the cluster is . For each point, it measures the average distance to points within the same cluster (cohesion) versus the minimum average distance to points in another cluster (separation). Values close to 1 indicate well-defined, distinct clusters, whereas values close to -1 may indicate misclassified points or overlapping clusters. This makes it particularly useful for evaluating clustering algorithms like DBSCAN, which can create clusters with varying shapes and densities .

In DBSCAN, core points meet the minimum density requirement and form the cluster's interior, while border points are within the neighborhood boundary of a core point but do not meet the density requirement themselves . Noise points are neither core nor border points and are thus excluded from any cluster. This categorization allows DBSCAN to effectively handle datasets with noise, as it isolates sparse regions of lower density as noise points, maintaining the integrity of dense clusters .