Single-layer Territory Management. Optimizing territories considering only clients¶

In territory management, a territory is a customer group or geographic area over which either an individual salesperson or a sales team has responsibility. These territories are usually defined based on geography, sales potential, number of clients or a combination of these factors.

The main complexity in territory management is to create areas that are balanced with regards to more than one factor that usually behave very differently. There is no one-size-fits-all solution, and if the balance is off, sales management is likely to leave someone within their organization unhappy or leave money on the table. This is why it is very important to identify and understand all the components and requirements of your use case to apply the most appropriate technique.

We can differentiate between two main use cases: when the location of sales reps is important (usually because they have to travel to visit their clients) and when it is not (travel rarely occurs). The first case is clearly more complex than the latter.

In this notebook we will use two different techniques to solve territory management problems when only the location of clients needs to be considered, i.e., we will have a single layer of data consisting of client locations. We will prove the value Spatial Data Science techniques by showing their additional value compared to traditional techniques.

Use case description¶

A pharma lab is interested in balancing their sales territories in the state of Texas based on the number of current and potential clients per territory.

Their clients are mainly offices and clinics of medical doctors.

They are interested in creating 5 balanced territories.

We will use the following two datasets from CARTO's Data Observatory:

• Points of Interest (POIs). In particular, office and clinic of medical doctors POIs. We will use Pitney Bowes POI-Consumer dataset.
• Texas boundary geometry. We'll use Who's on First GeoJSON - Global dataset.

Note the POI dataset is premium and a subscription is needed to access this data.

0. Setup¶

We'll start by importing all packages we'll use.

In order to be able to use the Data Observatory via CARTOframes, you need to set your CARTO account credentials first.

Please, visit the Authentication guide for further detail.

0.1. Functions¶

The following function creates an H3 polyfill of the polygon and at the resolution indicated.

The function below is used throughout the analysis to check is clusters are balanced based on different metrics.

The function arguments are:

• cluster_names so that we can provide descriptive names to clusters
• areas_df is the GeoDataFrame
• groupby is the column with the cluster to which each cell belongs to
• **kaggregations for the different metrics we'd like to calculate

1. Download and visualize data¶

Next, we will download the data described in the usecase using CARTOframes.

Note in this notebook some prior knowledge on how to explore and download data from the Data Observatory is assumed. If this is your first time exploring and downloading data from the Data Observatory, take a look at CARTOframes Guides and the Data Observatory examples and discover how easy it is to get started!

1.1 Texas boundary geometry¶

We are interested in the geometry of the state of Texas. We'll download it from Who's on First GeoJSON - Global dataset.

1.2. Client locations¶

We'll download all POIs in Texas classified as "OFFICES AND CLINICS OF MEDICAL DOCTORS" from Pitney Bowes POI-Consumer dataset.

Note this is a premium dataset and a subscription is required.

1.3 Visualize data¶

2. Discretize space. H3 grid¶

A fundamental step in territory management is to discretize space. Territory management algorithms are computationally complex and hence it is crucial to leverage the spatial component to reduce complexity. We can do this by working at an aggregated level instead of considering each client location independently.

We first need to identify the smallest spatial aggregation that makes sense for your business, our geographic support. This can be census block groups, zip codes or counties, or you can be interested in using a standard grid, in which case it would ideally be a hierarchical spatial index such as Quadkey grid or H3 grid.

In this notebook we will use an H3 grid of resolution 4. We can easily discretize space by performing a polyfill of the Texas boundary polygon.

Note a buffer has been applied because H3 will fill the polygon with all hexagons of resolution 4 whose centroid lies within the polygon to be filled and we want to make sure the whole territory is covered.

2.1. Aggregate client info at the grid level.¶

We will aggregate clients by:

• Counting the number of clients per cell
• Calculating the median rating of clients within eack cell

3. Territory Optimization¶

Once we have our data aggregated, it's time to start working on building balanced territories.

We will explore two different techniques:

• KMeans clustering.
• Max-p spatial clustering algorithm.

3.1 Approach 1. KMeans clustering¶

This is a very well known and broadly used technique. However, this technique doesn't allow you to incorporate balancing criteria and usually generates low quality results for territory management.

3.1.1. Calculate clusters¶

3.1.2. Label grid cells¶

We assign each cell to the cluster it belongs to.

In order to have comparable results with other techniques, we identify 5 representative cells, each of which being far enough from each other to make sure they are in different clusters. This way, we can compare how each cluster changes.

We create a string variable with a more descriptive name for our visualization

3.1.3. Visualize and analyze results¶

KMeans calculates nice compact clusters (see map). However, we get very unbalanced clusters with Cluster_3 having 10 times more clients than Cluster_1, as can be seen in the chart below.

3.2 Approach 2. Max-p¶

Let's now try to balance the number of clients per cluster while maintaining connected clusters as compact as possible.

We will use Pysal's implementation of the Max-p algorithm. Max-p is a spatial clustering algorithm that calculates spatially connected clusters, with similar properties, while balancing one criterion, or mixed or criteria.

3.2.1. Weights. Adjacency matrix¶

The first thing we need to do is calculate the adjacency matrix which will tell the algorithm which cells are contiguous.

We will use Rook weights which considers two polygons to be contiguous if they share one edge.

3.2.2 Balancing criteria¶

We would like to balance clusters based on total number of clients. Normally we are not looking for a perfect balance, especially when dealing with more than one multiple criteria, and a balance tolerance is introduced. In our case, we will consider a tolerance of 20%, which means that we allow clusters to be as much as 20% below the perfect balance.

Note we are only using the number of clients to balance, but this dataset also has the number of employees per client and you might even have other data you might be interested in using. The good news is Max-p allows you to do that.

3.2.3. Similarity criteria¶

Max-p also allows you to set similarity criteria. These are variables that you want to have a similar behavior within clusters. For example, you might be interested in having clusters with similar demographic or socioeconomic characteristics.

In this case, we don't have any specific criteria, so we will use the gris cell centroid coordinates as similarity criteria in order to get clusters as compact as possible. You can try removing these or only adding one of the coordinates to clearly see whats the effect of these similarity criteria.

3.2.4. Calculate clusters¶

3.2.5. Label grid cells¶

We assign each cell to the cluster it belongs to and rename clusters based on the representative cells we mentioned in the KMeans section.

3.2.6. Visualize and analyze results¶

We can see how clusters now are less compact than they were with KMeans, but now clusters are balanced, with all of them satisfying the minimum requirement of 4256 clients per cluster.