How I Used K-means clustering at Invoicemate

The data executed for the K-means model comprised the seller-level information (number of locations, number of employees, number of invoices, etc.). Additionally, it included the product usage trend (invoice submissions, transactions, registration, and creating/editing/deleting items). Data was included from both ‘the invoicing’ and ‘invoice financing’ verticals with a platform usage time of at least 30 days.

The numerical data was accumulated in the following way: 

• Average over the lifetime (in invoicing and invoice financing modules ) of a seller

• The sum of the seller’s first 30-day period (for all motives and purposes, their trial period)

• The maximum amounts over the lifetime (in a free trial or paid subscription state) of a seller.

Using the Elbow Method to determine the number of clusters

What should be the number of clusters when deploying the K-means? I used the Elbow method to serve that purpose. It is the percentage of variance explained as a function of the number of clusters. According to the Elbow method, the number (k) of clusters at which adding another cluster (k+1) only results in a small marginal gain in the percentage of variance explained.

While Figure 2 above showed that 3 clusters would likely be ideal, I considered both three and four clusters for my model. After comparing variance for individual features in a three and four-cluster model, I ultimately decided four was the most representative.

I cleaned my dataset by removing strongly correlated features (such features make it difficult to interpret the model’s results), and the features that were free of variance between clusters. 

Interpreting the Results

One important requirement of K-means clustering is that data needs to be sorted in a set of clusters. Nevertheless, after executing the model and getting the clusters, I used categorical data (industry, business size) for a better contextual understanding of clusters. It helped me understand the reason for data segregation in particular patterns. 

My team used Python (pandas) to filter the data, construct the model, and accumulate the final feature set of clustering data, cluster assignment, and demographic data. The finalized dataset was imported into Tableau for better visualization. 

From then onwards, I studied both the numerical and categorical characteristics of each cluster to figure out the ‘seller archetype’ from each cluster.

Final Results

These are the results of the ‘Invoice Financing’ data (example data) 

A major chunk of Invoice Financing sellers are in Cluster 1, while other clusters represent small percentages. For an improved understanding of how these clusters are segregated, I segmented the dataset into the aforementioned demographic data.

The below examples depict the reasons for variations in the clusters in terms of some demographics. 

1. Size of Business

The Cluster 3 holds larger sellers. This exercise helped develop product features 

for such a complex user base. 

2. Invoicing Conversion

It is pretty apparent Clusters 3 and 4 stand a fair chance to convert to ‘Invoice

Financing’ after trial. So, in the context of feature value, they are getting the most 

value from a free trial of Invoice Financing. I got 47% higher conversion through

targeted marketing campaigns for this cohort of sellers. 

Sellers in clusters 3 and 4 show the highest conversion rates for ‘Invoice Financing.’

3. Feature Interaction

I wanted to understand how the cluster interacted with different features of the 

platform. It was revealed that the cluster 2 sellers interacted with the items daily.

4. Average Invoice Submission Per Day 

Cluster 3 took the lead when it came to average invoice submission per day. 

Archetype Sellers

With the above example data, we can highlight key qualities observed in each cluster when explaining these mock findings to business stakeholders: With the help of the above sample data, I highlighted the important traits observed in each cluster for presenting them to the leadership and stakeholders. 

1. Cluster 1 contains smaller retailers distinguished by their interest in using Invoice Financing. However, they had the lowest levels of feature interaction and conversion.

2. Cluster 2 had more than average levels of interactions with features, which is commonplace for a garments store (common vertical in this cluster).