The Innovation

This patent application concerns a system for controlling a recommender designed to provide updated predictions of user preferences for a large product set, such as in a Video on Demand (VOD) catalog. It addresses the challenge of balancing the need to retrain the recommender system with the consumption of system resources, such as bandwidth and computational power. The application proposes managing the type and amount of training data to drive the recommender system’s performance towards a desired level, which may not be the optimal level but is achievable for most or all users​​.

Technical Contribution

The board examined the technical aspects of claim 10 of the main request. The claim involves automatically controlling the performance of a recommender system in a communications system, including a client device for users. Document D1 was cited as disclosing a similar system, indicating that part of feature (A) in claim 10 lacks novelty. However, the board noted that the act of recommending products is not generally recognized as having a technical character. The application’s purpose was argued to limit resource use rather than the recommendations themselves​​.

Distinguishing features C to H in claim 10 over document D1 include comparing a predetermined reference performance metric of the recommender system with a received measured performance metric from a previous iteration to determine a difference value. The method employs a closed-loop control algorithm to generate control parameters, thereby controlling the recommender system to align the measured performance metric with the predetermined reference metric. This process involves training data derived from user interaction data, with a positive correlation between the amount of training data and the subsequent iteration’s measured performance metric​​.

The technical effect of these features is to minimize the use of network bandwidth and storage needed for training data in the communications system, including both the client device and the recommender system. This effect is achieved by limiting the amount of training data through the measured performance metric’s convergence towards a predetermined performance level. The objective technical problem addressed is reducing network bandwidth usage and storage in a communications system comprising a client device and a recommender system​​.

Key Findings

• Automatically controlling the performance of a recommender system – non-technical
• Comparing performance metrics and determining a difference value – technical
• Using a closed-loop control algorithm for generating control parameters – technical
• Controlling the recommender system based on derived training data – technical
• Training data derived from user interaction data – technical
• Providing derived training data to the recommender system – technical

Read the full decision here: T 0183/21 Decision