Key points
- "The application relates to fraud detection in digital transaction data. A machine learning system is to be configured to produce, for a proposed transaction received as input, [an output] value [...] indicative of whether the proposed transaction represents an "anomaly"
- "Neither a machine learning model (or "system") nor the training (be it supervised or not) of such a model has per se a technical character. In the present case, the trained machine learning model serves the purpose of determining the likelihood that a transaction is an anomalous one, which is not a technical purpose. Consequently, achieving this purpose at a high degree of accuracy is not a technical problem"
- The claimed method comprised generating an "augmented labelled training set" and training a machine learning system with that data set.
- The Board: "Determining whether a (financial) transaction presents an "anomaly" is not per se a technical problem. Hence, it is not apparent that the trained supervised machine learning system, when used to measure a degree of anomaly for a given transaction, produces any "further technical effect". In particular, no technical effect is implied by the fact that the system has been trained with the augmented, labelled training set. Consequently, the "trained supervised machined learning system" has no technical character and, hence, its production does not represent or imply a technical effect."
- "In its preliminary opinion (point 13), the board additionally expressed doubts that the claimed method actually solved the problem formulated by the appellant (see point 5.2 above) over the whole scope of the claim, but this issue was ultimately left open."
- I'm still looking forward to a published decision where there is no doubt at that point, i.e., that the claimed method provides a better machine learning system (e.g., more hardware-efficient or energy-efficient).
EPO
The link to the decision is provided after the jump.