SYSTEM AND METHOD FOR PREDICTING INTELLECTUAL PROPERTY INFRINGEMENT
A computer-implemented method including determining a feature-embedding vector for a listing item based on textual feature data and imagery feature data for the listing item. The method also can include determining, via a machine learning module, an intellectual property infringement prediction associated with a genuine item based on a feature-embedding vector for the genuine item and the feature-embedding vector for the listing item. Furthermore, the method can include upon determining that the intellectual property infringement prediction is positive, causing a take-down of the listing item from a retailer platform. Other embodiments are described.
1 . A system comprising one or more processors and one or more non-transitory computer-readable media storing computing instructions that, when run on the one or more processors, cause the one or more processors to perform operations comprising:
determining a feature-embedding vector for a listing item based on textual feature data and imagery feature data for the listing item;
determining, via a machine learning module, an intellectual property infringement prediction associated with a genuine item based on a feature-embedding vector for the genuine item and the feature-embedding vector for the listing item; and
upon determining that the intellectual property infringement prediction is positive, causing a take-down of the listing item from a retailer platform.
2 . The system in claim 1 , wherein the operations further comprise:
training the machine learning module based on a single training dataset comprising respective training items of each intellectual-property-infringed brand of multiple brands.
3 . The system in claim 1 , wherein:
determining the feature-embedding vector for the listing item further comprises:
extracting one or more textual embeddings from the textual feature data for the listing item;
extracting one or more imagery embeddings from the imagery feature data for the listing item; and
generating the feature-embedding vector for the listing item based on the one or more textual embeddings and the one or more imagery embeddings.
4 . The system in claim 1 , wherein the operations further comprise:
training the machine learning module based on a training dataset comprising genuine items, positive training items associated with the genuine items, and unlabeled training items.
5 . The system in claim 4 , wherein:
the unlabeled training items comprise unlabeled intellectual-property-infringing items and unlabeled intellectual-property-noninfringing items.
6 . The system in claim 1 , wherein the operations further comprise:
sampling, from unlabeled items, unlabeled brand training items of a brand training dataset for an intellectual-property-infringed brand, wherein:
the brand training dataset further comprises genuine brand items for the intellectual-property-infringed brand and positive brand training items associated with the genuine brand items; and
after sampling the unlabeled brand training items, training the machine learning module based at least in part on the brand training dataset.
7 . The system in claim 6 , wherein:
a quantity of the unlabeled brand training items is proportional to a quantity of the positive brand training items.
8 . The system in claim 6 , wherein:
the unlabeled brand training items for the intellectual-property-infringed brand comprise respective type items sampled based at least in part on a respective type of each of the respective type items and major item types for the intellectual-property-infringed brand.
9 . The system in claim 8 , wherein:
the respective type items of the unlabeled brand training items are further sampled based on a respective item-type percentage for each type of the major item types.
10 . The system in claim 9 , wherein:
the respective type items of the unlabeled brand training items for the intellectual-property-infringed brand are sampled uniformly from each brand of multiple brands for the unlabeled items; and
the multiple brands comprise the intellectual-property-infringed brand.
11 . A computer-implemented method comprising:
determining a feature-embedding vector for a listing item based on textual feature data and imagery feature data for the listing item;
determining, via a machine learning module, an intellectual property infringement prediction associated with a genuine item based on a feature-embedding vector for the genuine item and the feature-embedding vector for the listing item; and
upon determining that the intellectual property infringement prediction is positive, causing a take-down of the listing item from a retailer platform.
12 . The computer-implemented method in claim 11 further comprising:
training the machine learning module based on a single training dataset comprising respective training items of each intellectual-property-infringed brand of multiple brands.
13 . The computer-implemented method in claim 11 , wherein:
determining the feature-embedding vector for the listing item further comprises:
extracting one or more textual embeddings from the textual feature data for the listing item;
extracting one or more imagery embeddings from the imagery feature data for the listing item; and
generating the feature-embedding vector for the listing item based on the one or more textual embeddings and the one or more imagery embeddings.
14 . The computer-implemented method in claim 11 further comprising:
training the machine learning module based on a training dataset comprising genuine items, positive training items associated with the genuine items, and unlabeled training items.
15 . The computer-implemented method in claim 14 , wherein:
the unlabeled training items comprise unlabeled intellectual-property-infringing items and unlabeled intellectual-property-noninfringing items.
16 . The computer-implemented method in claim 11 further comprising:
sampling, from unlabeled items, unlabeled brand training items of a brand training dataset for an intellectual-property-infringed brand, wherein:
the brand training dataset further comprises genuine brand items for the intellectual-property-infringed brand and positive brand training items associated with the genuine brand items; and
after sampling the unlabeled brand training items, training the machine learning module based at least in part on the brand training dataset.
17 . The computer-implemented method in claim 16 , wherein:
a quantity of the unlabeled brand training items is proportional to a quantity of the positive brand training items.
18 . The computer-implemented method in claim 16 , wherein:
the unlabeled brand training items for the intellectual-property-infringed brand comprise respective type items sampled based at least in part on a respective type of each of the respective type items and major item types for the intellectual-property-infringed brand.
19 . The computer-implemented method in claim 18 , wherein:
the respective type items of the unlabeled brand training items are further sampled based on a respective item-type percentage for each type of the major item types.
20 . The computer-implemented method in claim 19 , wherein:
the respective type items of the unlabeled brand training items for the intellectual-property-infringed brand are sampled uniformly from each brand of multiple brands for the unlabeled items; and
the multiple brands comprise the intellectual-property-infringed brand.