T 0874/19 (Classifying resources using a deep network/GOOGLE) 06-07-2022

The application

1. The application relates to classifying search engine resources as a spam resource (or as belonging to the "spam" category) or not a spam resource (or as belonging to the "not spam" category).

2. Figure 2 reproduced below illustrates this resource classification system 200:

FORMULA/TABLE/GRAPHIC

Main request - inventive step

3. The board considers document D1 an appropriate starting point for assessing inventive step.

4. In document D1, by combining the idea of deep learning and large-margin discriminative learning, a new kNN classification and supervised dimensionality reduction method called "DNet-kNN" is proposed. It encompasses a non-linear feature transformation to directly achieve the goal of large-margin kNN classification, which is based on a deep encoder network with four hidden layers pre-trained with restricted Boltzmann machines (RBMs) (page 358, left-hand side, last paragraph; page 359, right-hand side, first paragraph; page 360, section "4 Large-margin kNN classification using deep neural networks", first paragraph and Figure 2).

5. Document D1 discloses that "kNN" is one of the most popular classification methods due to its simplicity and reasonable effectiveness. It has been shown to have good performance for classifying many types of data (section "1 Introduction", page 357, left-hand column). The method of document D1 uses non-linear transformations so that each data point stays closer to its nearest neighbours having the same class as it than to any other data in the non-linearly transformed feature space (page 358, left-hand side, first paragraph).

6. The method of document D1 is applied to classify newsgroup text data with binary features. This newsgroup text data constitutes a data set that contains binary occurrences for 100 words. The binary feature vectors have been used to classify the postings into four categories, which are "computer", "recreation", "science" and "talks" (section 5.3 on page 365). The board agrees with the examining division that an "embedding function" according to claim 1 is implicitly disclosed by this passage of document D1 since the newsgroup text data has to be transformed into numeric values at some stage.

7. Thus, document D1 discloses a system comprising, in the language of claim 1 of the main request:

a deep network implemented in one or more computers that defines a plurality of layers of non-linear operations, wherein the deep network comprises:

an embedding function layer configured to:

receive an input comprising a plurality of

features of a resource, wherein each feature is a value of a respective attribute of the resource, and

process each of the features using a respective

embedding function to generate one or more numeric values, and

one or more neural network layers configured to:

receive the numeric values, and

process the numeric values to generate an

alternative representation of the features of the resource, wherein processing the numeric values comprises applying one or more non-linear transformations to the numeric values; and

a classifier configured to process the alternative representation of the input to generate a plurality of categories (the part in italics being different from the definition of the classifier in claim 1)

8. The following are thus the distinguishing features of claim 1 having regard to document D1.

The classifier according to claim 1 is configured to

process the alternative representation of the input to generate a respective category score (226) for each category in a pre-determined set of categories, wherein each of the respective category scores measure (sic) a predicted likelihood that the resource belongs to the corresponding category (the underlined part being the distinguishing features).

9. The board notes that the application describes that the categories might be:

- a "spam" category and a "not spam" category (description of the application as originally filed, page 5, lines 11 to 14)

- a "not spam" category and a category for each type of spam, such as "content spam", "link spam" and "cloaking spam", etc. (page 5, lines 29 to 34)

- a category for each resource type of a group of resource types, including "news resources", "blog resources", "forum resources", "shopping resources", "product resources" and "political resources" (page 6, lines 3 to 12)

The system generates a score for each category. A score is a prediction of how likely it is that the resource belongs to the corresponding category (page 5, lines 6 to 10; page 5, line 34 to page 6, line 2 and page 11, line 31, to page 12, line 2). For example, the generated score can be a predicted likelihood that the resource is a "spam" resource (page 5, lines 26 to 28).

10. However, it is apparent neither what a further technical effect of the distinguishing features could be, nor what objective technical problem the subject-matter of claim 1 would solve.

In its statement of grounds of appeal, the appellant argued that following decision T 1227/05, the claimed system could at least be regarded as simulating a hardware circuit that classifies inputs and thus had a technical purpose. However, the board considers that decision T 1227/05 cannot be followed as argued by the appellant in view of recent decision G 1/19, Reasons 133.

In the absence of any technical effect beyond its straightforward implementation in one or more computers, the subject-matter of claim 1 of the main request does not involve an inventive step (Article 56 EPC).

First auxiliary request - inventive step

11. Claim 1 of the first auxiliary request differs from claim 1 of the main request in that the system is specified as being a system "for detecting spam" and that the following text has been added at the end of the claim:

"and wherein the pre-determined set of categories includes a search engine spam category, and the category score for the resource measures a predicted likelihood that the resource is a search engine spam resource".

12. A "spam" category is one example of a category considered by the board as falling within the scope of claim 1 of the main request.

13. Since the board doubts that the meaning of the expression "search engine spam resource" is clear, it interprets this expression in light of the description for the assessment of inventive step. The application designates a search engine spam resource as a resource provided to a search system that has been manipulated by a spammer to give the resource a high search engine ranking as a response to one or more queries which the resource would not legitimately have. For example, content in a resource may be made to appear particularly relevant to a specific geographic area, and so be highly ranked for queries directed to that area, when in fact the content refers to a business, for example, that has no place of business in the area. Search engine spam can include other forms of erroneous information as well (page 5, lines 15 to 23).

14. The set of features added by the first auxiliary request relates to the categorisation of the resource as "spam" and appears to be merely of a non-technical cognitive nature. A "spam" resource could at most comprise intrinsic features related to the inherent result of its manipulation by a spammer. In any case, the board doubts that such a manipulation is technical. Even if it were, the result of the manipulation, i.e. the "spam(med) resource", would not necessarily be. Furthermore, claim 1 does not define how the alternative representation is processed to generate a predicted likelihood that the resource is a search engine spam resource. Generating a likelihood appears instead to be a mathematical feature, and not technical either.

15. Therefore, the subject-matter of claim 1 of the first auxiliary request does not involve an inventive step (Article 56 EPC).

Second auxiliary request - inventive step

16. Claim 1 of the second auxiliary request differs from claim 1 of the first auxiliary request in that

"wherein each of the embedding functions is specific to features of a respective feature type, and wherein each of the embedding functions receives a feature of the respective type and applies a transformation to the feature that maps the feature into a numeric representation in accordance with a set of embedding function parameters,"

is inserted after

"process each of the features (220) using a respective embedding function to generate one or more numeric values (222)".

17. These additional features in this added text are not disclosed by document D1.

18. The effect of these additional distinguishing features (having regard to document D1) is to generate a numeric representation (i.e. one or more numeric values) according to the respective type of the resource features.

19. In a similar case (T 872/19), the appellant argued, for the first auxiliary request comprising the same additional features, that the objective problem to be solved was to provide a more accurate representation of the resource features by generating an alternative representation of these resource features "in a new way".

In decision T 872/19, Reasons 22.2, the board noted that providing a more accurate "alternative representation" of the resource features, and therefore refined relevance scores (the "category scores" in the current case), was not a technical effect since the relevance scores ("category scores" here) did not constitute a technical feature. The board considers that the same reasoning also applies to claim 1 of the second auxiliary request in the case at hand.

20. In the absence of any technical effect beyond its mere implementation in one or more computers, the subject-matter of claim 1 of the second auxiliary request cannot be considered to involve an inventive step (Article 56 EPC).

Conclusion

21. Since none of the requests is allowable, the appeal is to be dismissed.