Recombee in 2019: New Features and Improvements
This year was really huge for us. First and most importantly, we have gained trust in our service from many new clients all over the world! We all in Recombee are very grateful and will continue to build service you all love to use and can rely on.
Because we keep in touch with our clients and listen to their needs, we gained new experience in more areas and have collected a lot of valuable feedback. That’s why we worked on new features very hard in 2019 — so hard that we almost forgot to write a blog post about them :)
We would like to reduce this debt by the following short summary of some of the new features.
Recommendation Logic
Our service can be used in several places on your website or application. For example, you may like to show one or more recommendation rows on a homepage, on a product/content detail, inside an email or push notification, etc. Every such place we call the Scenario. So “homepage”, “detail-page”, and “email” may all be scenarios and you may require different settings for them.
For each scenario, you can now assign a Logic, which is a model ensemble that we pre-configured for a particular use-case. It represents the best configuration of our service for specific needs. Some of the logics are universal, such as recombee:recently-viewed and recombee:popular for scenarios showing recently viewed items or popular items. Others are more domain-specific, such as ecommerce:cross-sell logic useful for shopping cart recommendations on e-commerce sites, or video:watch-next for recommending content at the end of playback in video streaming systems.
We internally used the concept of “logic” for several years by our data science team for custom tuning of the service for premium customers. However, we are now confident enough to release this feature and several logics for everyone to use out-of-the-box. Logic can be configured through the API calls, or set in the Admin UI for each scenario — read on!
New Admin UI
After a few months of hard work (and countless cups of coffee), we released a completely new web interface (we simply call it the Admin UI). We are very excited about the new design. The Admin UI is now much more user-friendly and allows you to manage almost every aspect of our service.
It is also much nicer under the hood, which allowed us to introduce many new, useful, and advanced features. One of the most important are the options to easily configure the recommendations based on your current needs, without changing your code.
Configuration of Scenarios in the Admin UI
Prior to the release of this feature, every change (such as changing the mentioned Logic or business rules) for a particular scenario required adjustment of the API parameters. This is quite easy using our SDKs that we provide for several programming languages, but implementing the change typically requires allocation of a software developer for this task.
This is not the case anymore, since both the Logic and the Business Rules can be now set in just a few clicks in our Admin UI. No programming background is required for using our Admin UI, so the behavior can be easily adjusted by your product or content managers.
There are predefined Filters (e.g. allow only recent items, allow only items from particular category etc.) and Boosters (e.g. upsell, boost nearby items etc.) in our Recombee Library, that can be immediately used. You can also create your custom rules that suit your business needs. Take a look at our documentation for more details.
If you are not using Recombee yet and you want to try out these new features, you can create a free instant account at recombee.com
Australian Infrastructure
We are on another continent! This year we built new infrastructure in Australia to better cover the region. Besides Australia itself, the infrastructure lowers latency also for New Zealand and part of East Asia. Australia is the third beside our largest infrastructure in Europe with hundreds of bare-metal servers, and fast-growing infrastructure in North America, which we operate since last year.
Improvements in Recommendation Models
Offering an easy integration and user-friendly UI is definitely important if we want to be the leading recommender as a service. However even more important is having cutting edge recommendation models at our backend. After all, the quality of the recommendations is what drives the conversions and ROI. Our data science team was therefore also very active, namely in the following areas of R&D:
AutoML
- We improve our AutoML system responsible for tuning ensembles of recommender models by the component that is able to better estimate online performance from offline data (extending our research published at the conference RECSYS 2018 in Vancouver).
Real-time deep learning models
- Deep learning algorithms need to be compatible with real-time philosophy of your system so our implementation is capable of online learning and process sophisticated real-time queries instantly.
Contextual bandits
- We have added several variants of contextual bandits, we support ensembling these models with our portfolio of collaborative filtering and context based algorithms.
- These models are quite powerful in news and media domains, but work great in any domain with large traffic and need to discover emerging items quickly.
Deep reinforcement learning
- By utilizing deep embeddings from images and text descriptions of items and historical interactions, we are able to train deep recommenders on sparse rewards so the recommendations maximize not only short term goals such as conversion rate, but also longer term engagement of users and other criteria.
Metalearning
- Hundreds of databases in our infrastructure contain rich portfolio of image and text repositories uploaded as item attributes. Utilizing approaches such as meta learning with implicit gradient, we are able to adjust deep embeddings of items with few shot learning to every database.
- Note that all databases are strictly separated in our system and we never enrich user profiles or use any interaction data from different databases to improve the recommendation. We only improve our algorithms to better understand the content that is being recommended.
2020: Boldly Go Where No Man Has Gone Before!
We hope that you like these recent advancements and the way our product is heading.
Looking at the features that are being currently developed and the features in our roadmap for 2020, the next year should be at least as exciting as 2019. Stay tuned!
