Trained multi-label support vector machine running a one-vs-the-rest classifier

1. A system, comprising:

a multi-label document classifier, comprising:

a trained multi-label support vector machine (SVM) running a one-vs-the-rest classifier, wherein the trained multi-label SVM running the one-vs-the-rest classifier is trained to label input documents with one or more labels of a plurality of labels comprising at least fifty (50) labels, and the trained multi-label SVM is configured with trained parameters that are learned from training the trained multi-label SVM running the one-vs-the-rest classifier on:

document features of training documents each belonging to one or more of the plurality of labels, and

hyperplane determinations on labels in the plurality of labels, wherein the trained parameters include distributions of distances between the at least fifty labels and the hyperplanes, and the trained parameters are stored on a memory of the multi-label document classifier for use in applying the trained multi-label SVM running the one-vs-the-rest classifier;

a feature generator that creates the document features representing features of words in the input documents and the training documents; and

a harvester that:

harvests labels of the plurality of labels based on distances between the hyperplanes for the plurality of labels and the document features determined based on applying the trained multi-label SVM to the input documents, and

assigns the harvested labels to the input documents.

2. The system of claim 1 , further comprising:

a document database comprising the input documents, wherein the multi-label document classifier is configured to apply the trained multi-label support vector machine running the one-vs-the-rest classifier to the input documents in the document database to label the input documents.

3. The system of claim 1 , wherein the document features represent frequencies or semantics of words in a document.

4. The system of claim 1 , wherein the document features include frequency features based on term frequency-inverse document frequency (TF-IDF).

5. The system of claim 1 , wherein the document features include semantic features based on embedding in a multi-dimensional vector space using Word2Vec.

6. The system of claim 1 , wherein the document features include semantic features based on embedding in a multi-dimensional vector space using global vectors for word representation (GloVe).

7. The system of claim 1 , further comprising:

a hyperparameter tuning component configured to select hyperparameters of the trained multi-label SVM running the one-vs-the-rest classifier across regularization, class weight, and loss function in a predetermined search range such that an at-least-one score is at or within ten percent of maximum attainable over the predetermined search range.

8. The system of claim 7 , wherein the at-least-one score is a ratio of count of documents with at least one pairwise match between inferred labels and ground truth labels to a total number of documents with at least one ground truth label.

9. The system of claim 1 , wherein one label of the plurality of labels is parked domain, for documents posted on parked domains, the system further comprising:

a website identifier configured to identify parked domains by:

crawling uniform resource locators (URLs) that are within a predetermined edit distance of selected URL names; and

a document crawler configured to collect documents posted on the parked domains by:

determining for at least some of the crawled URLs that URL resolution is referred to an authoritative nameserver that appears in a list of parked domain nameservers identified as dedicated to parked domains,

collecting the documents posted on the crawled URLs that are referred to the parked domain nameservers,

labeling the collected documents as collected from the parked domains, and

storing the documents and parked domain labels for use in training.

10. The system of claim 1 , wherein the multi-label document classifier is configured to:

label a first document with first tier labels based on applying the trained multi-label SVM running the one-vs-the-rest classifier to the first document, wherein applying the trained multi-label SVM running the one-vs-the-rest classifier to the first document comprises:

creating, by the feature generator, document features representing features of words in the first document;

applying classification parameters of the trained multi-label SVM for the plurality of labels to the document features to determine distances between the hyperplanes for the plurality of labels and the document features;

harvesting, by the harvester, positive labels of the plurality of labels having positive distances as first tier labels;

harvesting, by the harvester, negative labels of the plurality of labels having negative distances and a strong separation from a distribution of the negative distances as first tier labels; and

applying, by the harvester, the first tier labels to the first document.

11. A system, comprising:

one or more processors;

a first memory, functionally coupled with the one or more processors, the first memory storing:

a trained multi-label support vector machine (SVM) running a one-vs-the-rest classifier, wherein the trained multi-label SVM running the one-vs-the-rest classifier is trained to label input documents with one or more labels of a plurality of labels comprising at least fifty (50) labels, and the trained multi-label SVM is configured with trained parameters that are learned from training the trained multi-label SVM running the one-vs-the-rest classifier on:

document features of first documents each belonging to one or more of the plurality of labels, and

hyperplane determinations on label classes in the plurality of labels, wherein the trained parameters include distributions of distances between the at least fifty labels and the hyperplanes, and the trained parameters are stored in the first memory for use in applying the trained multi-label SVM running the one-vs-the-rest classifier; and

a second memory, having stored thereon instructions that, upon execution by the one or more processors, cause the one or more processors to label second documents with one or more of the plurality of labels based on applying the trained multi-label SVM running the one-vs-the-rest classifier to the second documents, wherein the instructions to label the second documents comprises instructions that, upon execution by the one or more processors, cause the one or more processors to:

create the document features of the second documents;

determine distances between the hyperplanes for the plurality of labels and the document features using the trained multi-label SVM; and

harvest labels of the plurality of labels based on the distances between the hyperplanes for the plurality of labels and the document features.

12. The system of claim 11 , wherein the document features represent frequencies or semantics of words in a document.

13. The system of claim 11 , wherein the document features include frequency features based on term frequency-inverse document frequency (TF-IDF).

14. The system of claim 11 , wherein the document features include semantic features based on embedding in a multi-dimensional vector space using Word2Vec.

15. The system of claim 11 , wherein the document features include semantic features based on embedding in a multi-dimensional vector space using global vectors for word representation (GloVe).

16. The system of claim 11 , wherein the trained parameters are further learned from selecting hyperparameters of the trained multi-label SVM running the one-vs-the-rest classifier across regularization, class weight, and loss function in a predetermined search range such that an at-least-one score is at or within ten percent of maximum attainable over the predetermined search range.

17. The system of claim 16 , wherein the at-least-one score is a ratio of count of the first documents with at least one pairwise match between inferred labels and ground truth labels to a total number of documents with at least one ground truth label.

18. The system of claim 11 , wherein one label of the plurality of labels is parked domain, for documents posted on parked domains, and wherein the instructions include further instructions that, upon execution by the one or more processors, cause the one or more processors to:

identify parked domains and collect third documents posted on the parked domains by:

crawling uniform resource locators (URLs) that are within a predetermined edit distance of selected URL names,

determining for at least some of the crawled URLs that URL resolution is referred to an authoritative nameserver that appears in a list of parked domain nameservers identified as dedicated to parked domains, and

collecting the third documents posted on the crawled URLs that are referred to the parked domain nameservers;

label the third documents as collected from the parked domains; and

store the documents and parked domain labels for use in training.

19. The system of claim 11 , wherein the instructions to label the second documents comprise further instructions that, upon execution by the one or more processors, cause the one or more processors to:

label each document of the second documents with first tier labels based on applying the trained multi-label SVM running the one-vs-the-rest classifier to the respective document, wherein applying the trained multi-label SVM running the one-vs-the-rest classifier to the respective document comprises:

creating the document features representing features of words in the respective document;

applying classification parameters of the trained multi-label SVM for the plurality of labels to the document features to determine distances between the hyperplanes for the plurality of labels and the document features;

harvesting positive labels of the plurality of labels having positive distances as first tier labels;

harvesting negative labels of the plurality of labels having negative distances and a strong separation from a distribution of the negative distances as first tier labels; and

applying the first tier labels to the respective document.