Case study: RECOMMENDATION ENGINE FOR SKODA AUTO
How can the recommendation engine be used in the automotive industry?
Companies such as Netflix, YouTube or Amazon collect user data and analyze user behavior to be able to recommend content that fits their profile and improve the user experience. Trask and SKODA partners has been developing and improving such a recommendation engine for SKODA for the past two years. The aim is to offer customers the best possible car configurations and most relevant optional equipment to purchase with their chosen car.
"The car configuration website (cc.skoda-auto.cz) is divided by country and language. For users who are not familiar with the models and options that SKODA offers, the car configurator serves as a tool to explore the range and configure the user’s dream car from the comfort of their home."
Nguyen Tuan Anh, Data Scientist, Trask
1. How the solution works
The project can be split into two separate units because the technical solution behind the recommender system consists of two very different technical problems.
- Gathering all available data, identifying patterns and trends, transforming and cleaning it
To solve this problem, Model Training is used, which uses various data transformation techniques, statistical methods and machine learning models to perform pattern analysis and model training. This allows us to efficiently consolidate millions of data points collected from different sources through ETL processes into a final set. This dataset is a fraction of the size of the original data and represents the best possible vehicle configurations for a given market. The type of data that Model Training processes ranges from user browsing data, to emissions data, to past sales data to identify the latest customer behavior patterns. To ensure that recommendations are based on the latest trends, the underlying models are updated daily.
- Leveraging information from the data and using it to make real-time recommendations
"The user selects features they would like the car to have and indicates their budget. Our system then returns four different configurations that best match the criteria. It also diversifies the selection of car configurations in order to show you the as wide variety of options as possible.”
Nguyen Tuan Anh, Data Scientist, Trask
The entire recommendation operation, which involves complex calculations, has to be performed under 500ms. In our case, this is handled by Model Serving, which takes the data produced by Model Training and uses it to recommend different vehicle configurations to customers online. It does this based on their pre-determined preferences using vehicle and equipment scores.
Trask's Data Science team is responsible for the entire Model Training portion and the diversification process in Model Serving.
"One of the most challenging problems we have encountered is unifying data from multiple sources into one dataset. You need to understand every attribute of the data."
Martin Szlauer, Full Stack Developer, Trask
2. Engine Features
- Because we have found that customers in different countries have different tastes, one of the most important features of our recommender system is that the data is analyzed separately by market and therefore can be adjusted by market preferences.
- There commendation engine is able to adapt the diversification process so there are not recommended only the four most popular cars to customers, because they could be, for example, very similar configurations of the SUPERB. Our recommendation is more diverse, right for the concrete customer.
"Our wish is that SKODA's recommendation system will work in future not only for car models or extras equipment. It would be interesting to work with more SKODA product catalogues to provide additional valuable recommendations for customers."
Luboš Louženský, Product Owner, Recommendation Engine
3. Long-lasting cooperation
What unites us with SKODA AUTO is our shared passion for the most advanced technology and modern way of developing software. That's why our teams work in an agile environment relying on automation through CI/CD (Continuous Integration and Continuous Delivery), which allows us to continuously improve the recommendation engine. We have made improvements not only on the quality of the recommendations, but also on the online user experience.
"The use of the recommendation engine is not limited to the automotive industry or streaming sites, for example. We believe this technology can be used in many other industries - banking, insurance or telecommunications."
Andrej Svitek, Tech Lead, Trask
