Purchase Behavior Variables

Firstly, we built a clustering model by only using variables are related to 'Purchase behavior'. Purchasing behavior can be identified based on these attributes: the number of brands purchased, brand loyalty, the number of transactions, the number of runs purchasing same brand, volume of product and average price. We built clustering models by changing some parameters such as centers, nstart and iter.max in Kmeans model. It is shown that when we changed parameters of k-means model, there is no significant difference between models. Our baseline model is created by selecting 25 random sets and using 12 iterations.

Then, we checked characteristics of clusters to understand differences between clusters. Households size of cluster 3 is higher than other clusters. This increases the consumption such as number of brands, transactions and volume. Moreover, households in cluster 3 have higher brand loyalty than other clusters (generally consume products of same brand) and their affluence index is lower than others. Education level is greatest in cluster 1 and smallest in cluster 3.

In light of this information, marketing strategies must vary from cluster to cluster. For example, if you aim to reach people who have lower economic class and higher brand loyalty, you should consider households in cluster 3 and shape your marketing planning regarding patterns of cluster 3.

Next, we drew cluster plot by using the fviz package. It can be clearly seen that this clustering model is not optimal since clusters are heavily overlapping. This model was not able to segment households successfully in the intersection area.

Then, we applied both the elbow and silhouette method to decide the number of clusters in the model. According to elbow method, we can say that best k value is 6. Alternatively, the silhouette method determined the number of clusters as 4. As a result, we selected the number of clusters as 4 since if we increase the number of clusters, the scope of the business cannot be easily managed by marketing teams.

This plot shows the clusters when we apply 4 different clusters regarding elbow and silhoutte model. This model is slightly better than previous model, but still not best because clusters are overlapping a decent amount.

Basis for Purchase Variables

Secondly, we applied k-means clustering by using different variables. Basis of purchase variables obtains percent of volume purchased not on promotion, on promo code 6 and other than 6, proposition of beauty, health and baby products.

The graph below indicates that basis for purchase variables are not sufficient to segment households in consumption of consumer goods using 3 clusters. Clusters overlapped and so variance between clusters is small.

Then, we tried to learn behavior differences between clusters. Households in cluster 2 have higher brand loyalty (77%) than other clusters (meaning they generally consume products of same brand) and their affluence index is lowest among all others. Also, social economic status of cluster 2 is the lower than others (almost 3.4). People in cluster 3 are more educated than others. Overall, households in cluster 1 and 3 shows similar patterns in consumption.

Based on these insights, marketing strategies must vary from cluster to cluster. For instance, if you work on launching products which have medium price, you should target households in cluster 2 and build your marketing strategies matching with characteristics of cluster 2.

Then, we applied both elbow and silhouette methods to decide the number of clusters in the model. According to the elbow method, we can say that the best k value is 7. Besides of this, silhouette method determined the number of clusters as 8. As a result, instead of using 7 or 8 for the number of clusters, we determined the number of clusters as 4 since this help to manage marketing plans systematically.

This plot shows the clustering model when we ran 4 different clusters considering elbow and silhoutte above. This model is slightly better than previous model using only 3 clusters, but still not best because the clusters are overlapping.

Combined Variables

Lastly, we applied k-means clustering by using combined variables in part a and part b. Combined variables includes both purchase behavior and basis for purchase variables. The tables below give information about clustering model when we selected k as 3.

The graph below indicates that combined variables are slightly better than previos models, yet it is still not sufficient to segment households. Clusters overlapped (intersections are not easy to identify) and so variance between clusters is small.

Then, we checked characteristics of clusters whether we see vital differences between clusters. Household sizes of clusters are similar to each other. Moreover, households in cluster 1 have higher brand loyalty than other clusters (hence transaction of brand runs is also highest) and their affluence index is lower than others. Order of educational level is cluster 3 > cluster 2 > cluster 1. People in cluster 3 are more educated than others. Overall, households in cluster 1 and 3 shows similar pattern in consumption.

Based on this information, marketing strategies must vary from cluster to cluster. For example, if you launch premium products into the market, you have to reach people who have higher affluence index and education level which corresponds to cluster 3. For cluster 3, purchasing power is higher and not price oriented, but rather quality oriented.

Then, we applied both elbow and silhouette method to do benchmarking among clusters. According to the elbow method, we can say that best k value is 6. Alternatively, silhouette method determined the number of clusters as 9. As a result, we determined the number of clusters as 6 due to the business implications of having too many clusters.

This plot shows the clustering model when we apply 6 different clusters regarding benchmarking analysis above. Within clusters SSQ is higher and between clusters SSQ is lower than the previous model. This model is slightly better than previous models, but still not best because clusters are overlapping.

Purchase Behavior Variables

Several issues occur when using the k-means clustering. For example, the k-means algorithm is very sensitive to outliers and noise since the mean statistic is sensitive to these occurrences. In order to address these issues, we chose to explore k-medoids clustering as a second clustering technique.

First, we will apply this clustering using the purchasing behavior. We started by using 3 clusters again. Here is a plot of the clusters using this choice.

This plot does not vary much visually from the k-means algorithm. Once again, though, we wanted to verify what the optimal number of clusters actually is. Therefore, we chose to use both the elbow and silhouette methods to check for the optimal number of clusters. The elbow method does not demonstrate a clear elbow, but the silhouette method suggests 4 clusters to be optimal. Because of this, we will run the k-medoids again using 4 clusters instead of 3.

The graph below indicates that k medoids model works slightly better than k mean regarding basis for purchase variables, yet there is a still overlapping so distance between clusters is small.

According to table, cluster 2 has higher affluence index and lower brand loyalty than other clusters. On the other hand, households in cluster 3 is the most loyal customers in this market and brand runs metric is the lowest among all households. Cluster 1 and Cluster 4 have similar consumption patterns with slight differences. Household size of cluster 4 is the highest; hence total consumption vary significantly from other clusters.

Basis for Purchase Variables

Now, just like with k means, we will cluster on the basis for purchase variables. Once again, three clusters will be used as a baseline. Here is the result of the three clusters:

The silhouette method clearly indicates that 2 clusters would be optimal in the case of clustering on basis for purchase variables.

Running the pam algorithm with only two clusters yields clusters of very different sizes, as is visible below.

This table shows differences between clusters. For example, if we want to launch premium soap, we should try to understand cluster 2 and cluster 4. Nonetheless, when we want to increase sales of low priced products, we have to focus cluster 1 and cluster 3.

Combined Variables

Lastly, we will impliment the k medoids algorithm on the combined variables. First, we will start with the baseline of three clusters.

The three clusters are visualized below. The above table shows that the size differences in clusters are large. Instead, we will search for the optimal cluster size in an attempt to balance this.

The silhouette method indicates the optimal number of clusters as 8. This is a large number of clusters, though, given the business implications - 8 separate marketing plans would be very intensive to develop and implement. Therefore, we will select 4 as the best number of clusters given the result of the elbow method. 4 clusters also do not look too bad in the silhouette graph.

The clusters are somewhat closer in size than just using 3 groups, though cluster 3 is still a lot smaller than the others. In the table below, we can see the differences between the 4 clusters on some of the key variables. Cluster three has higher brand loyalty than the other clusters. The other three, larger clusters vary on some important variables such as brand runs, volume, and household size. Cluster 4 has the lowest household size, and purchases the smallest volume.

Purchase Behavior Variables

We implemented agglomerative hierarchical clustering by using different distance techniques such as weighted, complete and Ward's method.

We chose the clustering based on Ward's method rather than complete method. The sizes of clusters based on complete and weighted measures vary significantly. Most of households are in cluster 1 (86%). However, clustering with Ward's method creates more balanced clusters. We can check the agglomerative coefficient, which measures the amount of clustering structure found (values closer to 1 suggest strong clustering structure). We can clearly say that Ward's method is better than others on the basis of this dataset.

We checked for the optimal number of clusters using the hierarchical method.

This determines 5 clusters to be good based on both the elbow and silhouette methods. However, we determined the number of clusters as 4 since when we increase the number of clusters, size of several clusters is smaller. This does not add any significant value to business, because adding value of focusing some households in small clusters is not worth it.

When we check features of the households, we can say that cluster 2 has much higher affluence index but less brand loyalty, meaning households in cluster 2 purchase a wider variety of brands in the market. On the other hand, cluster 1 has higher brand loyalty in this market. The average household size of cluster 4 is much higher than others, which is why volume of consumption is also higher.

The first table shows cluster distribution of observations visually and the second one indicates the dendogram of the hierarchical clustering method which was built by using Ward's method with 4 clusters.

Basis for Purchase Variables

Secondly, we applied hierarchical clustering by using different variables. Basis of purchase variables obtains percent of volume purchased not on promotion, on promo code 6 and other than 6, proposition of beauty, health and baby products.

We chose the clustering based on Ward's method rather than the complete method. The sizes of the clusters based on complete and weighted measure vary significantly. Most of households are in cluster 1 (99%). However, clustering with Ward's method creates more balanced clusters. We can clearly say that Ward's method is better than the others on the basis of this dataset.

Using these clustering variables, we again checked for the optimal clusters using both silhouette and elbow methods. The elbow and silhouette methods indicate 7-8 clusters as optimal, but this is a large amount from a business perspective.

Instead, we decided to try smaller numbers of clusters and compare. In the first table, cluster 1 dominates other clusters, thus, 4 clusters are better than others. On the other hand, when we increase the number of clusters, some clusters are smaller and that is not sufficient for market segmentation. As a result, we determined the number of clusters as 4 since the clusters are more balanced than others. Concentrating on feasible customer segments is much better.

The first table shows the cluster distribution of observations, and the clusters are overlapping. The features were not able to seperate observations properly. The second one indicates the dendogram of the hierarchical clustering method which was built by using Ward's method with 4 clusters.

The following table indicates the differences between the clusters. Cluster 3 shows significant differences from the other clusters (higher brand loyalty and less affluence index). Nonetheless, clusters 1 and 4 show similar consumption behaviors.

Combined Variables

Lastly, we applied hierarchical clustering by using combined variables in part a and part b. Combined variables includes both purchase behavior and basis for purchase variables.

We used Ward's method to measure distance between points (Ward performs well) and cut the tree by checking dendrogram. Overall, we can use 5 clusters to segment households and then concentrate on characteristics of these clusters.

We also checked the elbow and silhouette method for the combined clustering variables. Elbow doesn't give a super distinctive result in this case. Silhouette indicates that we should use 2 clusters, but this is very few, so we still feel a few more are better.

First table shows cluster distribution of observations, and it is evident that clusters are overlapping, meaning the features were not able to seperate observations properly. The second one indicates dendrogram of hierarchical clustering method which was built by using Ward's method with 4 clusters.

This table shows differences between clusters based on combined variables. It can be clearly seen that behavior of cluster 1 and 2 are totally different. Cluster 2 has the lowest affluence index and the highest brand loyalty.(focus this group for low priced products). For premium products, you have to consider households in cluster 1 and cluster 3.