Online Retail Network Analysis
Tim VULIC
27 June 2024
1. Introduction
1.1. Context
The Online Retail dataset was published by the UCI Machine Learning Repository (Chen, 2015) and aims to capture sales form an online retail store based in the UK. The dataset lists all transactions that occurred between December 1, 2010 and December 9, 2011, where 18,532 orders took place, while mentioning all items by quantity and unit prices for each order and the country of origin.
The online store would like to further expand in the European market, and the manages ahve been especially interested in Switzerland. The managers want to explore following aspects:
average order values and the average number of items per order for customers in the both countries,
how do the product co-purchase networks differ and
what insights can be drawn from the community structures within these co-purchase networks of both countries.
1.2. Descriptive statistics
As visible on the table below, the number of sales in Switzerland is relatively small compared to the UK:
## # A tibble: 3 × 2
## Country Number_of_orders
## <chr> <int>
## 1 Total 18532
## 2 United Kingdom 16646
## 3 Switzerland 51While Switzerland only represented 51 orders over the time period, the table below suggests that Swiss consumers place drastically bigger orders, both in terms of money and items.
## # A tibble: 2 × 3
## Country Average_Amount Average_Items
## <chr> <dbl> <dbl>
## 1 UK 439. 256.
## 2 Switzerland 1107. 590.To simplify the overall computation, only the 30 most bought items in each country were considered in the analysis and comparisons.
2. Centrality Measures
2.1. UK-Market
UK Network - Number of nodes: 29
UK Network - Number of edges: 786 The UK-market is composed of 29 nodes and 787 edges. Each node in the networks presented below represents a product, and each edge means that the connected products were bought together in at least once over the time period.
The top nodes by degree centrality, indicating the number of direct connections each product has with others, include “jumbo bag red retrospot”, “lunch bag red retrospot”, “jumbo bag baroque black white”, “pack of 72 retrospot cake cases”, and “pack of 60 pink paisly cake cases”, each with a degree of 56, which indicates that these products are frequently purchased together with a wide variety of other items.
Closeness centrality, in the context of a retail analysis indicates which products are well-integrated into the shopping patterns of customers, which is valuable when it comes to cross-selling and promotional strategies.
## UK - Top nodes by degree centrality:## JUMBO BAG RED RETROSPOT LUNCH BAG RED RETROSPOT
## 56 56
## JUMBO BAG BAROQUE BLACK WHITE PACK OF 72 RETROSPOT CAKE CASES
## 56 56
## PACK OF 60 PINK PAISLEY CAKE CASES
## 56## UK - Top nodes by betweenness centrality:## JUMBO BAG RED RETROSPOT LUNCH BAG RED RETROSPOT
## 0.7276584 0.7276584
## JUMBO BAG BAROQUE BLACK WHITE PACK OF 72 RETROSPOT CAKE CASES
## 0.7276584 0.7276584
## PACK OF 60 PINK PAISLEY CAKE CASES
## 0.7276584## UK - Top nodes by closeness centrality:## JUMBO BAG RED RETROSPOT LUNCH BAG RED RETROSPOT
## 0.03571429 0.03571429
## JUMBO BAG BAROQUE BLACK WHITE PACK OF 72 RETROSPOT CAKE CASES
## 0.03571429 0.03571429
## PACK OF 60 PINK PAISLEY CAKE CASES
## 0.03571429The same top products are mentioned, with a closeness centrality value of 0.0357. This suggests that these items have the shortest average distance to all other nodes, enabling them to reach other products more efficiently and influencing a wide range of purchasing decisions.
Overall, these results suggest that products like “jumbo bag red retrospot” and “lunch bag red retrospot” are key players in the UK co-purchase network market.
2.2. Swiss Market
Switzerland Network - Number of nodes: 30
Switzerland Network - Number of edges: 718 The Switzerland product co-purchase network is composed of 30 nodes and 718 edges. The Uk network, in comparison, has a slightly higher number of edges, which may suggest a more diverse number of product co-purchases.
The top nodes by degree centrality include “round snack boxes set of 4 woodland”, “pack of 72 retrospot cake cases”, “plasters in tin woodland animals”, “woodland charlotte bag”, and “red retrospot charlotte bag”, with degrees ranging from 56 to 58. This high degree centrality indicates that these products are frequently purchased together with a wide variety of other items.
The top products by betweenness centrality in the Swiss network are “round snack boxes set of 4 woodland”, “pack of 72 retrospot cake cases”, “red retrospot bowl”, “red toadstool led night light”, and “blue polkadot bowl”, with values ranging from 3.96 to 6.68. These products act as significant bridges within the network, facilitating diverse purchase combinations and connecting various product clusters.
Closeness centrality also places the same top products at the forefront: “round snack boxes set of 4 woodland”, “pack of 72 retrospot cake cases”, “plasters in tin woodland animals”, “woodland charlotte bag”, and “red retrospot charlotte bag” have closeness centrality values ranging from 0.0333 to 0.0345. This suggests that these items have the shortest average distance to all other nodes, enabling them to reach other products more efficiently and influencing a wide range of purchasing decisions.
Overall, these results suggest that products like “round snack boxes set of4 woodland” and “pack of 72 retrospot cake cases” are key players in the Switzerland co-purchase network market.
## Switzerland - Top nodes by degree centrality:## ROUND SNACK BOXES SET OF4 WOODLAND PACK OF 72 RETROSPOT CAKE CASES
## 58 58
## PLASTERS IN TIN WOODLAND ANIMALS WOODLAND CHARLOTTE BAG
## 56 56
## RED RETROSPOT CHARLOTTE BAG
## 56## Switzerland - Top nodes by betweenness centrality:## ROUND SNACK BOXES SET OF4 WOODLAND PACK OF 72 RETROSPOT CAKE CASES
## 6.675568 6.675568
## BLUE POLKADOT BOWL RED RETROSPOT BOWL
## 5.213606 5.213606
## RED TOADSTOOL LED NIGHT LIGHT
## 3.962903## Switzerland - Top nodes by closeness centrality:## ROUND SNACK BOXES SET OF4 WOODLAND PACK OF 72 RETROSPOT CAKE CASES
## 0.03448276 0.03448276
## PLASTERS IN TIN WOODLAND ANIMALS WOODLAND CHARLOTTE BAG
## 0.03333333 0.03333333
## RED RETROSPOT CHARLOTTE BAG
## 0.033333333. Community detection
Continuing on the analysis and comparison of the UK and Swiss markets, the Louvain method - developed by Blondel et al. (2008), has been used to detect potential “communities” of products. The aim of a community detection process is to identify groups of nodes that are more densely connected to each other than to the rest of the network. So, in the context of online sales, the objective is to higlight potential “clusters” of products.
3.1. UK-Market
The visual representation of the UK product co-purchase network with communities below shows only one color, indicating that all nodes (products) belong to a single community. This suggests that there are no distinct clusters of products that are more frequently purchased together within separate sub-groups. Instead, the network operates as one cohesive unit.
The UK product co-purchase network’s single-community structure indicates a highly interconnected market where customers purchase a broad range of products together.
This uniformity in purchasing behavior suggests that cross-selling opportunities are extensive and not limited to specific product clusters. For the retailer, this insight is valuable to design comprehensive marketing campaigns and simplify inventory management, ensuring that a diverse array of products is always available to meet customer demand.
3.2. Swiss market
The visual representation of the Switzerland product co-purchase network with communities reveals two distinct clusters of products, indicated by the two different colors (blue and orange). This suggests that there are two main groups of products that are more frequently bought together within each group than with products outside their group.
The presence of two distinct communities indicates that Swiss customers tend to purchase products in two separate clusters. Each cluster represents a set of products commonly bought together, reflecting specific buying patterns and preferences in the Swiss market.
3.2.1. Community 1 (Blue):
This community seems to focus on household items, children’s products, and decorative items. These products might be associated with family-oriented or home decoration themes.
Some of these products include:
girls alphabet iron on patches
pack of 12 woodland tissues
easter tin bunny in circus parade
red toadstool led night light
pack of 60 spaceboy cake cases
plasters in tin woodland animals
3.2.2. Community 2 (Orange):
This community appears to include more seasonal and festive, gift-related items, focusing on picnic bags, decorative ornaments, and festive products.
Some of these products include:
woodland charlotte bag
assorted colour bird ornament
hand warmer owl design
scandinavian reds ribbons roll wrap
balloon water bomb pack of 35
woodland stickers
## Warning: `get.data.frame()` was deprecated in igraph 2.0.0.
## ℹ Please use `as_data_frame()` instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.4. Conclusion and implications
4.1. UK Market
The presence of a single community suggests that the products in the UK market are highly interconnected. Customers tend to purchase a wide variety of items together, without any clear sub-group separations. In other words, customers mix and match across the entire inventory.
With a single community, there are extensive potential cross-selling opportunities. The retailer can recommend almost any product combination since the network does not show strong, isolated clusters. Thus, marketing campaigns can target a wide audience with diverse product recommendations. Bundle offers and promotions can include a variety of items in the UK market.
4.2. Swiss market
Considering the community analysis above, the retailer cold develop specific marketing strategies targeting each community differently. Community 1 could be targeted around specific holidays and other seasonal events (e.g., back-to-school offers) and through partnering with social media influencers that target families (mostly young mothers, as society is still facing stigma). Community 2 could be targeted by partnering with young ‘lifestyle’ influencers, especially on TikTok, which has been increasingly driving sales among Gen-Z. Research by Ortiz et al. (2023) highlights how media and brand engagement through TikTok significantly influence purchase intent in this demographic.
Overall, the network analysis may help the retailer analyze which bundle offers could be developed to foster cross-sales.
When it comes to logistics and inventory management, these findings suggest that the retailer could open a small warehouse with these specific item ranges in Switzerland, so that these orders are shipped more rapidly instead of having to be shipped out from the UK.
4.3. Future market extensions and logistics
Given the rather small number of orders originating from Switzerland (yet), it may also be interesting to assess the overall potential of other countries. Looking at the map below, one can easily notice that the majority of international orders come from the Schengen area, especially from Germany (457 orders) and France (389 orders) over the given time frame.
Reflecting on the recent Brexit, the trade relationship between the EU and the UK is now more complex and involves more regulatory checks and customs procedures than it did when the UK was part of the EU’s Single Market (House of Commons Library, n.d.).
Thus, for both market expansion potential and easier shipping conditions for customers, it may be interesting to discuss the potential benefit of implementing a warehouse (or more) in the EU. While there are also orders from countries on other continents, their absolute value remains anecdotal based on the given dataset.
Orders by country - world map
5. References
Blondel, V., Guillaume, J.-L., Lambiotte, R., & Lefebvre, E. (2008). Fast Unfolding of Communities in Large Networks. Journal of Statistical Mechanics Theory and Experiment, 2008. https://doi.org/10.1088/1742-5468/2008/10/P10008
Chen,Daqing. (2015). Online Retail. UCI Machine Learning Repository. Retrieved June 19, 2024, https://doi.org/10.24432/C5BW33.
House of Commons Library. (2024). UK-EU relationship after Brexit. UK Parliament. Retrieved June 25, 2024, from https://commonslibrary.parliament.uk/brexit/uk-eu-relationship-after-brexit/
Ortiz, J.A.F., De Los M. Santos Corrada, M., Lopez, E., & Merced, L.A. (2023). Don’t make ads, make TikTok’s: Media and brand engagement through Gen Z’s use of TikTok and its significance in purchase intent. Journal of Brand Management, 30(5), 535–549. https://doi.org/10.1057/s41262-023-00330-z. Retrieved on June 27, 2024, from https://link.springer.com/article/10.1057/s41262-023-00330-z