Microsoft Technology Licensing, LLC

Sumit Srivastava

Suhit Sinha

Ananth Sankar

Grace W Tang

Rushi P Bhatt

Mohit Wadhwa

Shilpi Agrawal

Srinivasa Madhava Phaneendra Angara

Hitesh Manwani

Ashish Tripathy

Anirban Biswas

Nishka Krishnappa Saligrama

Techniques herein balance the need for flexibility with the need accuracy, using a reactive approach to viral spam detection. After content (e.g., a social media platform news feed or timeline post) is created, interaction activity (e.g., content views) with the content is monitored. Based on the monitoring of the interactivity activity, it is determined whether a reactive viral spam analysis condition is satisfied for the content (e.g., because the number of content views exceeds a threshold). In response to determining that the reactive viral spam analysis condition is satisfied, a determination is made whether the content is or is not viral spam. If the content is determined to be viral spam, then it may be reported or flagged for further action (e.g., take down after manual confirmation).

Reactive viral spam detection

Embodiments of copypasta filtering system technologies cluster digital content items into copypasta clusters, extract a first feature set from the digital content items in the copypasta clusters, apply a first set of filters to the first feature set, and based on output of the first set of filters, divide the copypasta clusters into first intent copypasta clusters and possible second intent copypasta clusters. A second feature set is extracted from the digital content items in the possible second intent copypasta clusters. A second set of filters is applied to the second feature set. Based on output of the second set of filters, second intent copypasta clusters are created.

Intent-based copypasta filtering

Techniques for selecting models using greedy search on validation metrics are disclosed herein. A system generates corresponding predictions for a validation dataset using a plurality of prediction models. The system selects one of the prediction models for inclusion in an ensemble set based on the selected prediction model generating more correct predictions for the validation dataset than the other prediction models, and then removes the selected prediction model from the plurality of prediction models to form a reduced plurality of prediction models. The system identifies remaining data records of the validation dataset for which the selected prediction model generated an incorrect prediction, and then selects an additional prediction model from the reduced plurality of prediction models for inclusion in the ensemble set based on a determination that the additional prediction model generated more correct predictions for the remaining data records than the other prediction models in the reduced plurality of prediction models.

Model selection using greedy search

Methods, systems, and computer programs are presented for detecting near duplicates and partial matches of videos. One method includes an operation for receiving a video containing frames. For each frame, keypoints are determined within the frame. For each keypoint, a horizontal gradient vector is calculated based on a horizontal gradient at the keypoint and a vertical gradient vector is calculated based on a vertical gradient at the keypoint. The horizontal and vertical gradients are binary vectors. Further, a keypoint description is generated for each keypoint based on the horizontal gradient vector and the vertical gradient vector. Further, the frames are matched to frames of videos in a video library based on the keypoint descriptions of the keypoints in the frame in the videos in the video library. Further, a determination is made if the video has near duplicates in the video library based on the matching.

Partial-video near-duplicate detection

Technologies for generating a multi-modal representation of an image based on the image content are provided. The disclosed techniques include receiving an image, to be classified, that comprises one or more embedded text characters. The one or more embedded text characters are identified from the image and a first machine learning model is used to generate a text vector that represents a numerical representation of the one or more embedded text characters. A second machine learning model is used to generate an image vector that represents a numerical representation of the graphical portion of the image. The text vector and the image vector are used as input to generate a multi-modal vector that contains information from both the text vector and the image vector. The image may be classified into one of a plurality of image classifications based upon the information in the multi-modal vector.

Growing community of inventors

Bengaluru, India

Sumit Srivastava