TL;DR
This paper introduces a generalized PROFSET model to optimize product selection in large retail supermarkets by considering cross-selling profits and category management constraints. It addresses the limitations of previous models that struggled with large transaction sizes and lacked flexibility for real-world retail strategies.
What Problem Does This Paper Address?
Retailers are increasingly leveraging data mining to gain a competitive edge. While association rule mining is effective for identifying product relationships, translating these rules into concrete marketing decisions, like product selection for promotions or shelf placement, remains a challenge. Existing models, like the original PROFSET, were designed for smaller transaction datasets (e.g., convenience stores) and couldn't effectively handle the larger baskets typical of supermarkets. Furthermore, these models often overlooked the strategic importance of category management, where retailers group products into categories for management and marketing purposes. This paper aims to bridge this gap by proposing a generalized model that can handle large supermarket transaction data and incorporate retailer-defined category restrictions, leading to more realistic and profitable product selection strategies.
Methods
The core of the paper is the generalization of the PROFSET model. The key methodological advancements include:
- Handling Large Baskets: The original PROFSET model approximated profit allocation by assuming a direct correspondence between a transaction and a frequent itemset. This is problematic with large baskets where a transaction can contain multiple frequent itemsets. The generalized model addresses this by introducing a probabilistic approach to allocate transaction profit. When a transaction doesn't perfectly match a frequent set, it uses the support of maximal frequent subsets within that transaction to probabilistically determine the most likely "purchase intention" and allocates profit accordingly. This is done iteratively until the transaction is fully accounted for.
- Category Management Restrictions: The model now incorporates category management principles. Retailers can define constraints on product categories, such as specifying the number of items allowed per category or the total number of items from a category to be included in the optimal set. This allows for more strategic product selection that aligns with business goals beyond just individual product profitability.
- Optimization Framework: The problem is framed as a mixed-integer programming problem. The objective is to maximize total profit from cross-selling effects while adhering to constraints related to the total number of items, the relationship between selected itemsets and individual products, and category-specific limits.
Datasets
The empirical study utilizes a dataset from a Belgian supermarket chain, comprising 18,182 market baskets collected over one month. This dataset includes:
- 9,965 different products.
- 281 product categories.
- An average market basket size of 10.6 items.
- Data from 3,381 loyalty card customers.
The paper does not introduce new benchmark datasets but uses this real-world data to demonstrate the model's effectiveness.
Key Insights and Ideas
- Profitability is more than individual sales: The model emphasizes that the true value of a product lies not just in its own profit but in its potential to drive sales of other products through cross-selling. This is crucial for optimizing product assortments.
- Probabilistic profit allocation for large baskets: The novel approach to allocating profit from large transactions using maximal frequent subsets and their support provides a more realistic way to capture cross-selling potential when direct matches are not available.
- Category management is a strategic imperative: Incorporating category constraints allows retailers to align product selection with broader business strategies, such as managing shelf space, promoting specific product image components (e.g., "fresh," "profitable"), or adhering to promotional leaflet limitations.
- "Eye-catchers" can be counter-intuitive: The study highlights that the best "eye-catcher" product (a product intended to draw customers and encourage further purchases) might not be the one with the highest individual profit, but rather one that has strong cross-selling potential with other products.
Key Results
The empirical study on Belgian supermarket data yielded promising results:
- Improved Product Selection: For 24.7% of product categories, the generalized PROFSET model selected a different product than the one with the highest individual profit ranking. This indicates that the model effectively identified products with significant cross-selling opportunities.
- Significant Cross-Selling Improvements: For categories like "candy bars," the model achieved a 588% improvement in cross-selling profits by selecting a product that, while low in individual profit, had strong cross-selling links.
- Validation of Model's Logic: The association rules discovered during the mining phase supported the model's selections, showing strong co-occurrence patterns that justify the chosen "eye-catcher" products.
- Handling of Abstract Product Names: The model can handle abstract product names (e.g., "cheese product") that represent collections of items without individual barcodes, which is common in supermarket data.
Limitations and Future Work
- Limitations:
- The original PROFSET model's limitation in handling large baskets is addressed, but the probabilistic allocation still relies on assumptions about customer purchase intentions.
- The paper notes that for categories with no frequent products, the model defaults to selecting the product with the highest individual profit. This might not always be optimal for cross-selling.
- Future Work:
- The authors plan to test the model in practice with a real-world experiment in cooperation with the Belgian supermarket chain.
- Further research will explore the impact of promotion coupons on optimal product set composition.
- Investigating the effect of value price reductions on cross-selling profitability is also planned.
Summary updated