5.5 Model Comparison Testing
5.5.1 Comparing the proposed algorithm with the approach detailed
detailed by Reutterer et al. in [140]
The approach detailed in [140] was run on the same itemsets as provided in Table 5.13. The experimental process was similar to the methodology detailed in [140], where the top 5% and the bottom 20% of customers, based on transaction size, were initially pruned. This was followed by an exclusion of high frequency items, the twelve most frequent items for each store, which as per [140], results in skewed clustering. Following this, eleven clusters were created for each of the four itemsets detailed in Table 5.13, with the most popular cluster for each antecedent being selected for treatment, i.e. marketing promotions. The comparison testing was conducted in two parts. The first part compared the effectiveness of the clustering approach, in particular its ability to reduce “false positives” and “false negatives”, whilst the second part compared the model’s ability to offer differentiated treatment to customers, with the focus on attracting new customers, and preventing revenue erosion.
5.5.1.1 Comparison of the effectiveness of the clustering approach
From Table 5.14 it can be seen that following the approach detailed in [140] results in a much smaller customer base being selected, compared with the relevant cluster sizes in Tables 5.9 and 5.10, which in itself is not a problem, provided that every customer is a highly likely candidate to act upon the treatment. However, in most cases only a fraction of the total target customers are known purchasers of the an- tecedent, and in some cases this fraction is as low as 42%, resulting in a very large “false positive” potential which is considered bad for marketing as it angers customers [116][145]. Further, a very large proportion of potential, good, target customers are left untreated as they fall in other clusters, as evidenced by the difference in the
number of target customers between Tables 5.9 and 5.10, and Table 5.14). This is a missed opportunity, and results in an increase in “false negatives”. For example, in Store 13, itemset (213,163), 1521 customers will be targeted of which 58% will be falsely targeted as they have no history of purchasing the antecedent. At the same time, 5124 customers (approximately 89% of known purchasers of the antecedent, as identified by the algorithm proposed in this study) are placed into other, non-targeted clusters. These customers will not be targeted for the itemset, and this will result in a missed opportunity for the grocery retailer. Consequently, the model proposed in this study has a more effective clustering approach than the approach outlined in [140], as it is more effective both in terms of Test 1, minimising “false positives”, and Test 2, minimising “false negatives” as detailed in Section 3.3.6 .
Target Itemset Target Customers Known Purchasers “False Positive” Potential “False Negative” Potential Store 9, (156,277) (2012) 2331 1008 57% 6464 Store 13, (213,163) (2012) 1521 646 58% 5124 Store 9, (57,88) (2013) 2394 1253 48% 7073 Store 21, (78,209) (2013) 632 299 53% 1343
Table 5.14: Target Clusters and Customers based on the approach proposed in [140]
5.5.1.2 Ability to attract new customers and offer a differentiated treat- ment approach
The “Known Purchasers” detailed in Table 5.14 were mapped into their original four loyalty groups, as detailed in Tables 5.11 and 5.12. The results of the mapping is presented in Table 5.15. It can be seen, from Table 5.15, that in general, the overall
trends of the loyalty splits from the approach taken in [140] are similar to the results obtained for the proposed algorithm in this study. Indeed, there are some subtle differences in some itemsets, and this is expected as the approach detailed in [140] prunes the lowest 20% and highest 5% of transactions, based on size, thereby reducing the pool of loyal customers who typically have higher transaction sizes. It should be noted that both approaches prune at the lower transaction level, hence no immediate comparisons can be drawn in this regard.
However, unlike the algorithm proposed in this study, the approach detailed in [140] does not segregate customers within these clusters any further. Consequently, all cus- tomers within a cluster receive the same treatment. This approach is likely to drive down prices. Being able to attract new customers, while retaining revenue spend from loyal customers is a challenge for many retailers as noted in [43], [66], [72], and [109]. Consequently the approach detailed in [140] has limitations in this regard as it could lead to unnecessary price reductions due to loyal customers being offered discounts that the store does not need to offer. Given this, and the fact that the algorithm proposed in this study offers customised treatment through a two-step clustering pro- cess, it can be concluded that the algorithm proposed in this study performs better than the approach proposed in [140] in terms of Test 3, ability to offer customised treatment, and avoid targeting customers that are loyal. It should be noted that both models are equally good at attracting new customers, given the high volume of “switchers” in the cluster pool.
Comparing the models for Test 4, ability to enhance the frequency of the target itemset could not be done given that there is a large proportion of customers that are potentially “false positive” in the results obtained from the approach detailed in [140]. Given that both models demonstrate benefits in increased frequency of purchasing, it
Store 9, (156,277) (2012) Store 13, (213,163) (2012) Store 9, (57,88) (2013) Store 21, (78,209) (2013) Target Customer Splits This Study Switchers 62% 62% 69% 82% Drop-Out 3% 9% 2% 4% Light Touch 27% 27% 22% 12% Loyal 8% 2% 7% 2% Target Customer Splits Reutterer et al. [140] Switchers 61% 57% 71% 84% Drop-Out 4% 9% 3% 5% Light Touch 28% 30% 22% 10% Loyal 7% 4% 4% 1%
Table 5.15: Model Comparison Testing - Loyalty Splits
is noted that both models pass Test 4.