CMS: Honda, Yamaha, and now also Squadra!

Fast Delivery, Extensive Range, and Easy Online Shopping. This is what CMS represents for thousands of motorcycle enthusiasts and hobbyists worldwide who have relied on CMS for over 25 years for their parts needs. Whether it’s a new spark plug to get back on the road or fenders for summer cross-riding, there is no shortage of parts with CMS’s stock of over 4 million products. Known as ‘Consolidated Motor Spares,’ CMS has grown to become a market leader and supplier of world-renowned brands such as Honda, Suzuki, and Kawasaki.

Challenge  

Each motorcycle is unique, and so is each ‘motorhead’ seeking the right part for their bike. Some spend every weekend tinkering with a Yamaha that once belonged to their father, while others aim to maximize performance by transforming their Kawasaki into a MotoGP-worthy powerhouse. Such diverse stories and motives naturally require a wide variety of parts. Today, CMS’s inventory consists of more than 4,300,000 products, each with its own characteristics and specifications that customers use to determine if it’s the right fit for their motorcycle. Proper classification of these products is crucial for CMS to make them easily searchable for everyone. Although the necessary information for classification—such as brand, images, and inclusion in specific schematics—was available and correctly recorded, the sheer volume of 4,300,000 items made manual classification an impossible task. “Companies often face data issues in the form of missing attributes and characteristics. In this case, the opposite was true; there was so much data available that manual classification was not an option,” said Martijn Schilpzand, Data Scientist at Squadra Machine Learning Company.

What are the consequences of not classifying products correctly? Primarily, it affects discoverability. When products do not appear through filters, they are less likely to be purchased than when customers can finely filter and select products based on their needs, motorcycle, and budget. A classification algorithm emerged as a potential solution, but this brought a new challenge: the need for training data, which was not yet available. “Some of CMS’s data was already classified, but a large portion was not. Therefore, we developed an application that allows for rapid and bulk labeling of products. This generates more labels and classifications for training our algorithm, with the ultimate goal of enabling it to classify parts autonomously,” said Developer Nick Minkels about the application he developed to facilitate the training and realization of an autonomous classification algorithm.

Solution  

As you may have noticed, the solution for CMS consists of two parts: a classification algorithm and an application to provide the algorithm with training data by quickly and in bulk classifying products. The overall project was a convergence of Development and Data Science, a dynamic collaboration that Martijn and Nick fondly recall. “It was an iterative collaboration, something I found very enjoyable. When Martijn encountered a classification issue, I could immediately incorporate a solution into the application. After some back-and-forth tweaking, a new feature would emerge, which Martijn could then use for the algorithm. Repeat this cycle a few times, and you end up with a fully functional application, as we have now,” said Nick about the iterative process that enabled him to equip Martijn with the tools needed to train and improve the algorithm.

This project also advanced towards ‘Multi-model’ classification by incorporating various types of data in product classification, such as schematics and images. To make this concept more digestible, Martijn explained: “We don’t focus solely on images or text to train our algorithm; instead, we use all available contextual information about a part. This can include images, colors, texts, brands, and so on. This use of data from multiple sources is known as multi-level classification.”The diversity of data used for training generally results in better-performing algorithms compared to single-level classifiers, which are trained on just one type of data, such as images or text alone.

Result  

Thanks to the software developed by Martijn and Nick, CMS has classified over half a million products to date. Moreover, this classification has enhanced the navigation structure of their website. When customers can navigate more efficiently through the vast array of parts, products are purchased more quickly. Additionally, this improvement enhances the overall customer journey by optimizing the final touchpoints through the software.