- \n
- Collaborative filtering<\/b>: This approach is based on the assumption that two different users who liked the same products in the past will probably like the same ones now and in the future. Such models recommend products liked or viewed by a user with similar interests. A great advantage of this method is that it does not require domain knowledge about the products and also gives users the chance to explore products that they explored before. However, such models cannot recommend products newly added to a company’s catalog and require frequent retraining in order to include them in their recommendation lists.<\/li>\n<\/ul>\n
- \n
- Hybrid<\/b>: This approach combines both content based and collaborative filtering and adapts the recommendations.<\/li>\n<\/ul>\n
Requirements<\/h2>\n
Which approach should a company choose though? Before deciding, it is important to assess the concept of Recommendation Engines from three main perspectives:<\/p>\n
Business Objective: What shall be improved?<\/h3>\n
Identifying the business objective <\/b>is the first prerequisite that should be fulfilled in order to make sure that the recommendations will ultimately achieve the desired outcome for a business.<\/p>\n
For example if the objective is to increase the revenue of a wide variety of products, then a Recommendation Engine could be optimized to suggest to users more diverse products . On the other hand, if the objective is to improve the user experience, the recommendations could be based for example on products that fit well together or products that are liked or viewed by users with similar interests.<\/p>\n
Serving: Touchpoints and Timing<\/h3>\n
Another important aspect is to understand how a Recommendation Engine is intended to be used<\/b>. For example, will the recommendations be presented to users in real time as soon as they visit a website, an app or even a store in person? Or will the strategy be instead to serve recommendations periodically via an email, display and video advertising campaign? Will the users interact with it for example by giving feedback about the recommendations?<\/p>\n
Such decisions will definitely have an impact not only on the Machine Learning method but also on the overall architecture of the Recommendation Engine and thus, they should be addressed at a very early stage of the development phase.<\/p>\n
Data Assessment<\/h3>\n
Developing a Recommendation Engine that is based on Machine Learning techniques requires large volumes of historical user and product data<\/b>. Such data is fed to Machine Learning models to optimize their ability to make predictions on recommendations.<\/p>\n
As a result of this, poor quality data can lead to poor recommendations (concept known in machine learning as “Garbage in, garbage out”). A good practise is to analyze the available data and assess whether it is relevant and sufficient for building a Recommendation Engine. Knowing for example if the product catalog contains many items, how often the product catalog data is updated as well as if there is data capturing users feedback, will definately simplify the decision making process.<\/p>\n
Even if the data is available, having the right capabilities to identify the business needs (Business Understanding) as well as to process the data and implement the solution (Cloud & Data Science) is the key to delivering a Recommendation Engine that will eventually create value for the company and its customers.<\/p>\n
![\"Recommendation](\"https:\/\/e-dialog.group\/wp-content\/uploads\/2025\/08\/e-dialog-Blogartikel-Recommendation-Engines-2.png\")
(c) e-dialog<\/figcaption><\/figure>\n Google Cloud Platform Solutions for Recommendation Engines<\/h2>\n
Currently, Google Cloud Platform offers four possible ways to leverage state of the art Machine Learning techniques and deploy Recommendation Engines in the cloud.<\/p>\n
- \n
- BigQuery ML<\/a><\/li>\n
- Recommendations AI<\/a><\/li>\n
- AI Platform<\/a><\/li>\n
- Self built & customized with Tensorflow<\/a> or other similar programming frameworks<\/li>\n<\/ul>\n
Using BigQuery ML for Recommendations<\/h3>\n
BigQuery ML<\/b> is a great tool for low cost and fast implementations<\/b> of Recommendation Engines, however currently supports only one specific recommendation type<\/b> ( recommends to users only products which are liked or viewed by other users with similar interests). The models are created using SQL <\/b>and can be accessed in BigQuery via a Web User Interface (Rest API, command line tools and external tools such as jupyter notebooks are also supported). When it comes to performance tests<\/b> (such as A\/B testing), a separate implementation needs to be integrated with this solution.<\/p>\n
Recommendations AI – Reco AI<\/h3>\n
![\"\"](\"https:\/\/e-dialog.group\/wp-content\/uploads\/2025\/08\/e-dialog-Blogartikel-Recommendation-Engines-3.png\")
(c) e-dialog<\/figcaption><\/figure>\n On the other hand, Recommendations AI <\/b>offers more of an end to end<\/b> solution<\/b>, supporting a wider range of recommendation types<\/b> while providing a User Interface <\/b>which simplifies the process of monitoring and testing the performance <\/b>of recommendations. Further advantages include its ability to take into account the users phase in the customer journey and that it can be activated in multiple touchpoints and channels. It definitely minimizes the effort that is required to build a Recommendation Engine, since it does not require expertise in ML or domain knowledge. Keep in mind though that it comes with higher runtime costs and less flexibility for customization<\/b>.<\/p>\n
AI Platform<\/h3>\n
AI Platform<\/b> is a development platform which allows quick and cost-efficiently deployment <\/b>of ML projects. It definitely requires a certain level of ML expertise and domain knowledge<\/b> for preparing, building, validating and deploying ML models. When it comes to Recommendation Engines, AI platform offers the possibility to use one of its built-in algorithms (Wide & Deep<\/a>) which has been used in the past for increasing app acquisitions in Google Play.<\/p>\n
Self built & customized Recommendations Engine<\/h3>\n
Self Customized<\/b> Recommendation Engines on the other hand offer the most flexibility<\/b> but require ML, domain expertise<\/b> as well as higher development and maintenance efforts.<\/b><\/p>\n
Below you can see a quick summary of what each solution can offer:<\/p>\n
\n\n
\n <\/td>\n BigQuery ML<\/b><\/td>\n Recommendations AI<\/b><\/td>\n AI Platform<\/b><\/td>\n Self Built & Customized<\/b><\/td>\n<\/tr>\n \n Recommendation Type<\/b><\/td>\n “Products liked by users with similar interests”<\/td>\n “Other products you may like”<\/td>\n “Most relevant products”<\/td>\n Depends on the selected method, for example:\n “Similar products”,<\/p>\n
“Products liked by users with similar interests”<\/p><\/td>\n<\/tr>\n
\n “Frequently bought together”<\/td>\n<\/tr>\n \n “Recommended for you”<\/td>\n<\/tr>\n \n “Recently viewed”<\/td>\n<\/tr>\n \n Method<\/b><\/td>\n Matrix factorization<\/td>\n Not disclosed<\/td>\n Wide & Deep<\/td>\n Hybrid<\/td>\n<\/tr>\n \n Data Source<\/b><\/td>\n Product ids<\/td>\n Product catalog<\/td>\n Product catalog<\/td>\n Product catalog\n User events<\/p><\/td>\n<\/tr>\n
\n User ids<\/td>\n User events from a wide range of data sources<\/td>\n User events from a wide range of data sources<\/td>\n User events from a wide range of data sources<\/td>\n<\/tr>\n \n Implicit or explicit feedback<\/td>\n Optionally other relevant data sources<\/td>\n Optionally other relevant data sources<\/td>\n Optionally other relevant data sources<\/td>\n<\/tr>\n \n Domain Knowledge <\/b><\/td>\n Low<\/td>\n Low<\/td>\n High<\/td>\n High<\/td>\n<\/tr>\n \n Development Effort<\/b><\/td>\n Medium<\/td>\n Low<\/td>\n Medium<\/td>\n High<\/td>\n<\/tr>\n \n Runtime Cost<\/b><\/td>\n Medium<\/td>\n High<\/td>\n Medium<\/td>\n Low<\/td>\n<\/tr>\n \n Configuration UI<\/b><\/td>\n No<\/td>\n Yes<\/td>\n No<\/td>\n No<\/td>\n<\/tr>\n \n Customization<\/b><\/td>\n Low<\/td>\n Low<\/td>\n Medium<\/td>\n High<\/td>\n<\/tr>\n \n Performance Measuring<\/b><\/td>\n Can be customized<\/td>\n Incorporated in the UI<\/td>\n Can be customized<\/td>\n Can be customized<\/td>\n<\/tr>\n \n A\/B testing <\/b><\/td>\n Can be customized<\/td>\n Implicit<\/td>\n Can be customized<\/td>\n Can be customized<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n Are ready made Recommendation Engines from well known suppliers the right solution?<\/h2>\n
When searching for Recommendation Engines you will find a plethora of well known vendors offering out of the box software. Even though many of them offer great solutions there are several points that one must be aware of:<\/p>\n
- Recommendations AI<\/a><\/li>\n
- BigQuery ML<\/a><\/li>\n
- Hybrid<\/b>: This approach combines both content based and collaborative filtering and adapts the recommendations.<\/li>\n<\/ul>\n