Recombee in 2020: New Features and Improvements
Welcome to our annual blog post about the things we accomplished during the year, including a little bit of shameless bragging about some important newly-added features. We know that this year has been quite challenging for many people, including ourselves. However, today we want to focus entirely on the positive (no pun included) side of the year and the stuff we are the proudest of.
We acquired new tech hires from the beginning of the year, and they have become irreplaceable reinforcements to the teams. The business development team expanded, enabling Recombee to improve the service we are providing to our new and existing clients. We have also doubled our AI/ML team to stay ahead of competitors not only in the performance of recommendation algorithms but also to speed up research and development of new features requested by our customers.
HTML Widgets Integration
As much as we believe that API integration of Recombee into your sites and applications has always been straightforward for programmers, we still felt a need to introduce an integration method that is more accessible for everybody, be it a programmer or a non-programmer. Following this philosophy, we have worked on the widget feature.
We have extended the Recombee Admin Interface by a UI based (WYSIWYG) editor, much like Word, where you define/click the whole widget and its behavior. Each widget can be easily used just by deploying a generated code tag to your site.
Once you deploy the widget code tag, you can tweak and modify widget behavior in our editor and deploy all the changes with a single click. No additional coding is needed!
You can optimize widget behavior depending on the screen or site element size so that you can use a single widget configuration on both mobile and web versions of your site.
Our UI/UX team worked countless hours to bring the most polished and streamlined version of the UI widget editor that we believe you will enjoy using. We invite you to give us feedback on existing features and possibly request new features, as we care about your experience using our product.
Personalized Full-Text Search
You can probably imagine that when a programmer searches for “python”, and when a reptiles enthusiast searches for “python”, they both expect very different results. And that is why we introduced a new API endpoint for personalized full-text search!
Traditional search solutions take into account only the user’s search query, but our search solution also takes into account additional user data such as interaction history and properties, and can therefore return exactly what the user is looking for.
Another benefit is that you can combine the search capabilities with other Recombee features, such as Filters and Boosters, for applying your business rules (e.g., slightly preferring new content or products with higher margin).
If you like to learn more about the algorithms behind personalized search, please refer to our recent blog post.
Improvements in Scalability: We Can Now Handle Traffic of ANY Site
In 2020 we started to work with clients that require a significantly higher number of recommendation requests than we used to process in the previous years.
Although Recombee is designed from the very beginning to be a very scalable solution, we had to introduce several new improvements to handle dozens of thousands of recommendation requests per second and similar amounts of interactions.
What is even more challenging is that these recommendation requests need to be served immediately, providing results within tens of milliseconds. Within this limited time, complex computations to calculate the result need to be performed. Our engine is simultaneously updated in a real-time fashion with incoming user interactions, and models are being re-trained on the fly.
All these computations are performed on a server infrastructure that has to be robust and self-healing because Recombee provides strict availability guarantees to many enterprise customers.
We believe that it is not just bragging when we say that we are now capable of providing recommendations for any website or app in the world - of course while maintaining the quality of the recommendations and providing just the right content or products for each individual user.
Infrastructure on the US West Coast
We have a fast-growing customer base in the US, and establishing a second server cluster in North America was yet another step to boost this trend. We can now serve our US customers with minimal network latencies, as we operate infrastructures on both the East and the West Coast.
Besides the two North American infrastructures, we also offer clusters in Europe and Australia.
Improvements in Recommendation Models
Last year, we added several new algorithms such as real-time deep learning models, contextual bandits or deep reinforcement learning to our portfolio. In 2020 we have been improving the existing models and work on end-to-end recommendation models.
Our research of next basket prediction involves large scale evaluation of different architectures that are computationally expensive. Thanks to our partnership with NVIDIA, we can utilize advanced GPU clusters for research purposes.
We are also participating in an interdisciplinary research team (that includes also UX specialists and journalists) which is working on analyzing and modeling the impact of editors and the recommender engine on the audience and studying the effects in behavioral data (user interactions). We work with several customers, where Recombee personalizes everything, including the front page. Hence, including the preferences of editors, curators and studying the impact on the audience is of crucial importance. The goal is to research and develop a powerful recsys human-machine interface enabling our customers (not exclusively in media) to understand better which content makes their users happy in all niches. Also, we aim to visualize how different recommendation strategies and rules that our customers use to adjust the recommender system impact their end-users.
We believe this is an important step towards explainability in recommender systems and the reduction of various biases that have negative impact on the end-users. Let us know if you would like to work with us.
See You In 2021!
Our team is working on new disruptive features that promise to make our customers more successful with their products. Some features, such as next basket prediction, will open new opportunities and markets for them. We also work hard to democratize our AI technology, so you do not need to be a programmer to deploy and manage Recombee to enhance your product. We wish you a Happy New Year and stay safe!
Next Articles
Recombee and Kentico Xperience: Guide to One-On-One Personalization
Recombee expanded its integration options - and now is available at the Kentico Xperience platform! Analyzing different types of personalization, we look into why Kentiko chose our AI-powered recommendation engine over manual segmentation.
Deep Learning for Recommender Systems: Next Basket Prediction and Sequential Product Recommendation
Accurate “next basket prediction” will be enabling next generation e-commerce — predictive shopping and logistics. In this blogpost, we will discuss the deep learning technology behind next basket...
How Interdisciplinary Collaboration Can Accelerate AI Innovation
In a world where innovation is the new standard, Recombee uses the power of interdisciplinary collaboration to stay at the cutting edge of innovation. Partnering up with the leading player in the food industry (Bofrost) and academia (FIT CTU), allowed Recombee to hold a student competition to create AI which can shape the future of the food industry.