5.3.1
Data Sources
The 2012 and 2013 consumer scanner panel dataset obtained from [90] was used as the basis for the experiments. In general, the data sets contained over 32,000 unique customers and over 51 million individual scanned items, across 21 stores in the UK. Items were grouped at the product category level, and given an anonymised, unique identifier. For example, all milk products, including whole milk, goat’s milk, etcetera were classified under the product category “milk”. In total, there were 286 product categories, hence the anonymised datasets used in this study comprised of 286 unique items. All store formats of the same store were also combined, e.g. internet, express, garage shop etcetera. Three stores were chosen, a “Big Four” store, (Store 9), a hard discounter, (Store 13), and a high-end grocery retailer, (Store 21).
For completeness, the model was also tested on large, dense datasets to evaluate its processing performance. These datasets were created using a synthetic transaction database generator as detailed in [83], with the largest dataset consisting of 5 million transactions, 100 unique items, 5 frequent itemsets, frequency density of 0.5, and a maximum basket size of 50 items. It should be noted that this dataset is similar in transaction volume, but considerably denser, to that of the UK’s largest grocery retailer’s daily activity, and represents approximately 28% of the UK’s grocery retail market share [131].
5.3.2
Identifying target itemsets
Frequent itemsets were mined using the Apriori algorithm as detailed in [7] for all itemsets with an initial minimum support of 0.05, and minimum confidence of 0.1. These “weak” constraints were placed on the dataset to prune highly infrequent, and
poorly associated itemsets. It should be noted that labelling these conditions as “weak” is justifiable as typical grocery retail datasets, like the datasets in this study, have highly associated items with confidence greater than 30%, and support of over 45%. Highly frequent single items typically have supports of greater than 60% [90].
Computer programs were written using R software to mine and analyse the data, whilst Microsoft Excel was used to compute the mt values and perform simulations using the techniques outlined in Chapter 4. Note that this study does not perform inter-store comparisons, hence the size of each selected store is not important. Indeed, intra-store comparisons are made, and whilst it is likely that some goods are purchased more frequently in-store or online than others, it is assumed that the pricing policy is the same across all formats and that customers have full choice in selecting a format that best suits them. These assumptions are considered fair given the prominence of multi-channel shopping, the emphasis that stores place on consistency across all channels, and the growing adoption of the “customer is king” mentality by UK retailers [47][157].
The data for each store, S, was analysed and target itemsets were identified based on the criteria that there exists two items, A and C, which are frequent, with minsup = 0.1, but their combination (A, C) is not frequent. Further (A, C) is the optimal combination to target, from all identified targets, based on the mt value obtained using Equation (4.15).
5.3.3
Identifying target customers
Target customers for each targeted itemset, (A, C), in each store, S, were identified based on the customer’s purchase history and household size, using the FCM algorithm detailed in Chapter 4. Customer clusters were then classified based on the criteria outlined in Table 4.1. To eliminate “false positives”, customers had to have visited
the store at least twice in the year, or the period under consideration, and purchased the antecedent, A, and the consequent, C, at least once during that period.
5.3.4
Simulating the impacts of the proposed model
An ergodic Markov model was created using Microsoft Excel to simulate the impact of marketing interventions on the shopping behaviour of the identified target customers. The model was based on the concepts outlined in Section 4.4. Two campaigns, one “conservative” and one “aggressive”, were used, with their corresponding proportion vectors given in Tables 5.1 and 5.2, respectively. Note that the proportion vectors have a time period of 1 week. This is due to the assumption that all transactions made by a customer during a one week period may be considered to be one transaction, as detailed in Chapter 4. The following example is used to illustrate the concept of the proportion vector: from Table 5.1, 99% of all “switchers” in week 1 will remain “switchers” in week 2, while the other 1% will be elevated to “drop-outs”. Then in week 2, 99% of all customers who are “switchers” will go on to remain as “switchers” in week 3, with the other 1% being elevated to “drop-outs”. This process will recur whilst the treatment plan is in place.
Future
Current
Leave Alone Light Touch Drop Out Switchers
Leave Alone 1 0 0 0
Light Touch 0.01 0.99 0 0
Drop Out 0 0.01 0.99 0
Switcher 0 0 0.01 0.99
Table 5.1: Conservative Marketing Campaign
It is clear that from Tables 5.1 and 5.2 that the general approach of the simulation campaigns, and indeed marketing, is to nudge customers towards increasing loyalty. Consequently, the proportion vectors in Tables 5.1 and 5.2 assume that the target customer’s loyalty is always enhanced after each marketing campaign and not dimin-
Future
Current
Leave Alone Light Touch Drop Out Switchers
Leave Alone 1 0 0 0
Light Touch 0.02 0.98 0 0
Drop Out 0 0.02 0.98 0
Switcher 0 0 0.02 0.98
Table 5.2: Aggressive Marketing Campaign
ished. This is a reasonable assumption to make, especially under the condition of zero “false positives”, which this technique aims to achieve, and is demonstrated later on this chapter (see Section 5.4.2.1). In general, marketing initiatives provide incentives to customers to purchase more, and if those incentives are correctly targeted, e.g. providing incentives for goods that customers want to buy, then these incentives are most likely going to be acted upon [120][126][141]. This notion of “nudging customers towards loyalty” is discussed further in Section 5.4.3.
5.3.5
Comparative Testing of the proposed algorithm
The performance of the proposed algorithm was tested using the principles outlined in Chapter 3. The proposed model was compared with the approach detailed in [140] against the four tests detailed in Section 3.3.6. Further detail, together with the results and discussion of the performance testing of the algorithms are provided in Section 5.5. The proposed model was also compared against “reality”, and tested using other data sets. The results and discussion of these tests are provided in Sections 5.5.2 and 5.6 respectively. Finally, and for completeness, the proposed model was also compared with alternative approaches, i.e. targeting “top sellers”, and the results and discussion of this test is detailed in Section 5.5.3.