Motivating Clustering for Customer Segmentation
-
Clustering refers to the process of grouping similar observations together
- Roughly, we can think of clustering as the process of grouping observations that are correlated with each other
- Typically, clustering is used for exploratory data analysis
-
One common use-case is customer segmentation
- Where, similar customer profiles can be identified and grouped together into clusters (or segments)
-
Clusters are grouped together based on a similarity metric that accounts for:
- Demographic attributes
- Behavioral attributes
- Financial attributes
- Each segment has a centroid representing its geometric center
- Roughly, we can think of the centroid as the average of the profile
Describing Uses for Segmentation
- For example, one segment may represent power shopper, who is older and higher income
- Whereas, another segment may represent high engager, who is young and typically uses social media
-
As a result, segmentation allows us to do the following:
- Understand who our customers are
- Summarize those established customer profiles
-
Create segments that can be used as a feature for other predictive models
- E.g. creating labels representing power shoppers, app engagers, brand loyalists, etc.
-
Generalizing sparse data
- E.g. using topic modeling to produce topics from words in articles
Challenges of Clustering Projects
-
Segmentation projects involve the following:
- Good understanding of customers and expectations
- Thorough experimentation
- Executable marketing strategies
-
Segmentation projects involve the following challenges:
- Thorough, unsupervised experimentation
-
Optimization function is ambiguous
- Not as deterministic as classification models
- Dependent on the chosen clustering method and parameters
- Difficult to interpret why certain customers are included in one cluster over another cluster
Illustrating Common Clustering Methods
-
K-means algorithm:
- Better for more spherical clustering data
- Very efficient
- Can specify number of clusters
- Each observation gets assigned to a cluster
-
DBSCAN:
- Better for density-based clustering data
- Not very efficient
- Can't specify number of clusters
- Not every observation is assigned to a cluster
-
Gaussian Mixture Modeling:
- Better for elliptical clustering data
- Less efficient
- Can specify number of clusters
- Each observation gets assigned to a cluster
-
Hierarchical:
- Better for tree-based clustering data
- Efficient
- Can specify number of clusters
- Each observation gets assigned to a cluster
References
Previous
Next