Use Cases for Online Portal Recommendations Using Data Products GmbH is Germany’s largest vehicle market and, as part of the eBay Classifieds Group, belongs to one of the world’s biggest online companies. Logo is a marketplace for the buying and selling of vehicles. Every month, the web platform draws 13.5 million visitors who can choose from the more than 1.6 million vehicles on offer. Each visit to the platform creates a stream of data which contains information about the demand for particular vehicles, the quality of the vehicles for sale, and user requirements. wants to use this data to continuously improve the user experience for both vehicle sellers and purchasers. Each function which is added to the platform as a result of usage information and which provides additional added value for the user is called a “data product”.

We use this term to illustrate the fact that data can be used to provide major benefits to the web platform’s users. This strategy is the basis for’s collaboration with inovex on new data products, one of which provides user-specific vehicle recommendations. This project’s main aim is to support users in finding the perfect vehicle and to help them do so more quickly and easily.

User-Specific Recommendations in the Purchasing Process

When they first approach the platform, many users have only a very rough idea of the perfect vehicle for them. In order to gain an overview, they initially look at a large number of different vehicles (this is known as the “orientation” phase of the process). Vehicle model recommendations at this stage help users gain the most comprehensive overview possible of all the models which could potentially suit them.

As users become clearer about their own requirements and what is on offer, their search queries and behaviour become increasingly specific (the “specification” phase of the process). The ads they view become less diverse and their choices are narrowed down to a model (or small range of models). Users “favourite” vehicles, saving the ads in order to contact the sellers. Recommendations on users’ favourites lists show them additional relevant ads which they have not previously seen. In order to facilitate this, the system uses the user preferences reflected in the items in the users’ favourites lists. Once users begin contacting sellers, they have entered the “realisation” phase of the purchase process.

Recommendations bring:

User Benefits:

At every stage in the process, targeted recommendations improve the selection process and aid users in their decision-making, as well as in finding suitable vehicles for their needs. Users are inspired by the recommendations, and the provider is involved, advising and assisting customers. The user experience is significantly improved, with users finding the best vehicle for them more quickly.

Business Advantages:

The improvements for users also mean objective business benefits for Clearly measurable benefits include an increase in clicks and a reduction in the bounce rate. The targeted recommendations cause users to spend longer on the site. An overview of the various KPIs and how they are affected by the use of intelligent algorithms is available at

Recommender Use Cases for

We have worked with to develop and implement several use cases. These are either currently in use on the platform or will be shortly:

1. Recommendations directly on the homepage

Darstellung Startseite

Users are offered initial recommendations on the site?s homepage. These serve as an alternative to entering the site via the Search function. Users who have already visited are recognised and directed straight to available vehicles, which are selected based on the users’ previous interactions with the marketplace.

2. The favourites list for finding similar vehicles

Recommendations are also made in the “car park”, the user favourites list. Based on the ads they have saved there, users are shown similar vehicles for sale. A new or more comprehensive search is created for the user using the suggestions.

Depiction of car park feature on

3. Display of vehicles similar to favourites

Depiction of suggestion feature on

If a vehicle that a user has saved to his or her favourites list is no longer available, he or she is shown equivalent vehicles. These recommendations are based on the user’s favourites list and take into account the searches he or she has performed. The vehicles the user is shown are of the same model as the vehicle that is no longer available, with as many of the same features as possible. This use case does not take into account previous search queries for other models.

Implementation and Future Plans

We enhanced the existing collaborative filtering infrastructure (Mahout and related technologies were already in use) by adding components for content-based filtering using Elasticsearch. This enhanced setup enabled us to implement a hybrid recommender and to support the evaluation of various implementation scenarios using live A/B testing.

All the data products were developed and rolled out using agile processes. These are used in order to provide continuous added value through iterative improvements. The aim was (and is) to use information provided by the hybrid recommender to define potential interesting use cases and to move these as quickly as possible through the development process to the live platform.

Parallel to this, we also identified potential for optimisation within the infrastructure. These improvements will enable us to handle paradigm shifts in technology (like making recommendations using deep learning approaches), as well as to make incremental improvements.

These two goals were pursued in parallel and an appropriate deep learning approach was implemented based on Google?s ?Wide & Deep Learning for Recommender Systems’ paper. The results of the evaluations were integrated directly in the implementation of the wide and deep learning system.

The Results of the Improvements

Graph: 60 percent growth

The results of the improvements swiftly became apparent on the platform, even as the initial recommendations were going online. The “favourites recommendations” resulted in a 64% increase in clicks, while click-through rates increased by 60%.

Strategic Optimisation of the User Experience Through Data Products

In creating these use cases, was able to draw upon inovex’s many years of experience in developing data products. Another data product we previously implemented for and with automatically estimates a vehicle’s sales value based on its features. Rather than viewing the individual data products as separate from one another, we consider them components of an integrated portfolio. Our common technological infrastructure and agile development processes enable us to rapidly and flexibly create a great variety of data products. In doing so, we take upon ourselves the strategic challenges of product management, which requires us to constantly assess the fitness of a product for the market and to determine the appropriate business model for a particular data product.

Although we can testify to the smooth collaboration with inovex, we have been particularly pleased to receive positive user feedback which confirms the success of the recommender system.

Benjamin Eckart

Senior Manager Data and Trust & Safety,
Technology Stack
  • Mahout
  • Elasticsearch
  • Cassandra
  • Hive
  • Python
  • Kafka

Get in touch!

Florian Wilhelm

Head of Data Science, Contact for Data Management & Analytics