AI Recommender
A smart recommendation system that predicts which TV programs and movies might be of interest to the user. It is designed to prolong user engagement on the platform and to increase their levels of satisfaction with the smart TV experience. See our solution in detail.
The hybrid system: harnesses the best of both worlds
Our AI recommendation system uses both collaborative and content-based filtering. This means that the end user benefits from their own data, the data fetched from other users with similar interests, and the product metadata.
Data gathering
The data is generated from two sources: operators responsible for product modifications and users interacting with content on their smart TV devices. Each type of interaction is treated differently, and their varying levels of importance are reflected in the recommendation system. The data, including the newest product modifications and user interactions, is gathered daily and processed by the recommendation system.
What data exactly
The users' actions fetched to the recommendation system are divided into two subgroups. They can be player-related, such as stopping the video, generating a bookmark, or watching only some periods of the video. They can also be content-related, for instance adding the product to the favorites list, having free or purchased access to the product, renting the product, etc. Moreover, the timing of these actions is also taken into account.
Predictions
Predictions are stored in the form of vectors in the ElasticSearch database. This elegant solution ensures system scalability, allowing predictions to be generated for each customer with minimal resource use, even for millions of users. In addition, it also stores product metadata and gives operators the option to use efficient queries for filtering predictions.
Eye-catching recommendations
Recommendations designed to keep the user engaged
Create as many sections with recommendations as you want. Freely adjust their order and composition. The sections may be general or focused on very specific genres or other metadata information. The system ensures that only the most relevant content is shown to the end user.
Diverse data preprocessing techniques and ML algorithms
The system uses some of the most efficient data preprocessing techniques and Machine Learning algorithms.
During the preprocessing phase, decisions are made regarding the significance of each user's actions, caching the data, and other customizations essential to our solution. Some of the ML algorithms used in our application are:
- ALS (Alternating Least Squares) - a highly efficient algorithm designed for large-scale collaborative filtering with implicit feedback. It not only factorizes large sparse matrices, but in our case also returns recommendations for the selected user based on the preferences of other users.
- SVD (Singular Value Decomposition) - another efficient factorization algorithm. Being highly flexible, it is used in both content-based and collaborative filtering. Although it handles both sparse and non-sparse matrices, in sparse matrices the missing values must be filled. Both algorithms work very well with parallel processing.
Solving the cold-start problem
The cold-start problem refers to the challenge of generating recommendations for new users who have little to no history of interacting with content on the platform. In this situation, our system will generate a basic recommendation vector based on active users' preferences. The basic set of predictions will then be displayed to operators on the Dashboard and shown to the new users on their respective platforms.
Need to know more?
Customizable filters on Dashboard
Easy customization just by moving a few blocks around
Our database solution provides operators with the option to set up highly customizable filters in order to generate different sections with recommendations. They can decide whether to include different genres, actors, adult content, age ratings, etc., in the given section. The filter works on top of the recommendations already generated by the recommendation system for the user.
Legality and adult content protections
All user data is anonymized with identifiers before being fetched to the recommendation system. The product metadata is obtained legally. While the system generates recommendations for adult content like it does for any other content, the younger audience is protected. Adult content will be displayed only in parts of the platform intended for the adult audience and will be hidden behind the CA PIN. This ensures a safe experience for both underage viewers and their parents.
Build with us Your own Product
Unlock your product's potential with Proexe. Our expertise in Android and iOS development ensures market reach and streamlined processes. Using HTML, CSS, React JS, and PHP Laravel, we craft visually stunning experiences. With Google Cloud Platform, security and scalability are guaranteed. Plus, our agile approach means fast, flexible development.
