Microsoft Technology Licensing, LLC

Sap Se

Jesper B Lind

Christopher A Meek

Darko Kirovski

Michael Stark

Suchitra Sundararaman

Everaldo Aguiar

Patrice Y Simard

Geoffrey John Hulten

Kenneth D Ray

David Earl Heckerman

Carl Myers Kadie

Eric David Brill

Benjamin Livshits

Asela J Gunawardana

C James MacLennan

Vijay Gajjala

Marc A Smith

Gennady Medvinsky

Lev Sigal

Jody D Biggs

Yuri Arshavski

Duncan L Davenport

Anton Ioffe

Adithya Kumar

Juliy Broyda

Wensi Xi

Evan Adkins

Anupam Dewan

Yongjian Bi

Shivani Patel

Ravi Sharma

The present disclosure involves systems, software, and computer implemented methods for transaction auditing. One example method includes training at least one machine learning model to determine features that can be used to determine whether an image is an authentic image of a document or an automatically generated document image, using a training set of authentic images and a training set of automatically generated document images. A request to classify an image as either an authentic image of a document or an automatically generated document image is received. The machine learning model(s) are used to classify the image as either an authentic image of a document or an automatically generated document image, based on features included in the image that are identified by the machine learning model(s). A classification of the image is provided. The machine learning model(s) are updated based on the image and the classification of the image.

Anomaly and fraud detection with fake event detection using machine learning

Some embodiments provide a non-transitory machine-readable medium that stores a program. The program receives from a client device an image and a user identifier associated with a user of the client device. Based on the user identifier, the program further retrieves a set of historical transactions associated with the user, the program also uses a first machine learning model to predict a string based on the image. The program further uses a second machine learning model to predict a set of digits based on the image. Based on the set of historical transaction data, the string, and the set of digits, the program also determines a data value for a field in a form.

Automated determination of data values for form fields

The present disclosure involves systems, software, and computer implemented methods for transaction auditing. One example method includes determining valid pixel-based pattern(s) that are included in valid reference images. Fraudulent pixel-based pattern(s) that are included in fraudulent reference images are determined. A request to classify an image is received. A determination is made as to whether pixel values in the image match a valid pixel-based pattern or a fraudulent pixel-based pattern. In response to determining that the pixel values match a valid pixel-based pattern, a likelihood of classifying the first image as a valid image is increased. In response to determining that the pixel values match a fraudulent pixel-based pattern, a likelihood that the image as a fraudulent image is increased. The image is classified in response to the request as either a valid image or a fraudulent image based on the likelihoods.

Anomaly and fraud detection with fake event detection using pixel intensity testing

The present disclosure involves systems, software, and computer implemented methods for transaction auditing. One example method includes receiving text submitted with a request. Token values are automatically extracted for selected token types from the text using a machine learning model. Tokens are identified in the text, and for each token, features and a token type of the token are determined, along with a score that indicates a likelihood that the token has the token type. A token value is extracted for the token from the text. Extracted token values are compared to request data values. A request value for the selected token type is identified in the data values for each selected token type. The extracted token value for each selected token type is compared to the request value. An audit alert is generated in response to determining that an extracted token value does not match a corresponding request value.

Transaction auditing using token extraction and model matching

Embodiments of the present disclosure pertain to matching transactional data. In one embodiment, the present disclosure includes a computer implemented method comprising receiving transactional data for a first user and second data for the first user, selecting transactional data records for the first user from a data store of transactional data for a plurality of users, determining a plurality of similarities between fields of the transactional data and second data, determining a likelihood of a match between a transactional data field and a second data field based on the plurality of similarities using a machine learning model, and in accordance therewith, identifying one record in the transactional data records for the first user that generates said likelihood of the match above a first threshold, and replacing values second data fields with corresponding values in the one record.

Systems and methods for matching transactional data

The present disclosure involves systems, software, and computer implemented methods for transaction auditing. One example method includes receiving receipt data associated with an entity. Policy questions associated with the entity are associated with at least one policy question answer that corresponds to a conformance or a violation of a policy selected by the entity. For each policy question, a machine learning policy model is identified for the policy question that includes, for each policy question answer, receipt data features that correspond to the policy question answer. The machine learning policy model is used to automatically determine a selected policy question answer to the policy question by comparing features of extracted tokens to respective receipt data features of the policy question answers that are included in the machine learning policy model. In response to determining that the selected policy question answer corresponds to a policy violation, an audit alert is generated.

Transaction policy audit

Embodiments of the present disclosure pertain to extracting data corresponding to particular data types using machine learning algorithms. In one embodiment, a method includes receiving an image in a backend system, sending the image to an optical character recognition (OCR) component, and in accordance therewith, receiving a plurality of characters recognized in the image. The character set is matched against known values to generate candidate character strings. The character set is processed by one or more machine learning algorithms to produce features. For each candidate character string, the features are then processed by a random forest model to determine a final character string.

Machine learning data extraction algorithms

Embodiments of the present disclosure pertain to systems and method for extracting data from an image. In one embodiment, a method of extracting data from an image comprises receiving, from an optical character recognition (OCR) system, OCR text in response to sending an image to the OCR system. The OCR text comprises a plurality of lines of text. Each line of text is classified as either a line item or not a line item using a machine learning algorithm, and a plurality of data fields are extracted from each line of text classified as a line item.

Systems and methods for extracting data from an image

Various embodiments are disclosed that relate to the automated identification of one or more computer program functions for potentially placing on a remote computing device in a split-computational computing environment. For example, one disclosed embodiment provides, on a computing device, a method of determining a factorable portion of code to locate remotely from other portions of the code of a program to hinder unauthorized use and/or distribution of the program. The method includes, on a computing device, receiving an input of a representation of the code of the program, performing analysis on the representation of the code, the analysis comprising one or more of static analysis and dynamic analysis, and based upon the analysis of the code, outputting a list of one or more functions determined from the analysis to be candidates for locating remotely.

Identifying factorable code

A human interactive puzzle (HIP) authorization architecture where keyed and animated puzzles are executed by HIP players which are distinct and obfuscated to the point where breaking a single player is a relatively costly operation. A key is created in response to a request for a service, a HIP player is created based on the key, and a small installation executable is created that expands during installation to produce a computationally expensive data structure on the client relative to verification of the solution at the server. Thus, copying of the player or relay of the puzzle to a third system requires more time than allowed to receive the solution at the server.

Keyed human interactive proof players

Human Interaction Proofs ('HIPs', sometimes referred to as 'captchas'), may be generated automatically. An captcha specification language may be defined, which allows a captcha scheme to be defined in terms of how symbols are to be chosen and drawn, and how those symbols are obscured. The language may provide mechanisms to specify the various ways in which to obscure symbols. New captcha schemes may be generated from existing specifications, by using genetic algorithms that combine features from existing captcha schemes that have been successful. Moreover, the likelihood that a captcha scheme has been broken by attackers may be estimated by collecting data on the time that it takes existing captcha schemes to be broken, and using regression to estimate the time to breakage as a function of either the captcha's features or its measured quality.

Automatic construction of human interaction proof engines

A 'Quality-Based Ad Pricer' prices ads as a function of how users respond to a particular page and/or domain to which they are sent by an online advertisement. User experience is improved by ensuring that advertisements that are more relevant to a user are priced less than an ads which are less relevant to the user. In particular, a quality factor for each ad is determined as a property of the advertiser's site based on measured user behaviors with respect to that site. This quality factor is then used in ranking, selecting, and pricing ads in an automated online auction. Further, while ad aggregators are not excluded from the ad market by the pricing rules of the Quality-Based Ad Pricer, these rules ensure that there is a 'level playing field' such that ads of merchants are not excluded by the ads of ad aggregators.

Quality based pricing and ranking for online ads

The present invention provides systems and methods that rank search results. Such ranking typically includes determining a relevance of individual search results via one or more feature-based relevance functions. These functions can be tailored to users and/or applications, and typically are based on scoped information (e.g., lexical), digital artifact author related attributes, digital artifact source repository attributes, and/or relationships between features, for example. In addition, relevance functions can be generated via training sets (e.g., machine learning) or initial guesses that are iteratively refined over time. Upon determining relevance, search results can be ordered with respect to one another, based on respective relevances. Additionally, thresholding can be utilized to mitigate returning results likely to be non-relevant to the query, user and/or application.

Systems and methods that rank search results

The present invention provides collaborative filtering systems and methods employing statistical smoothing to provide quickly creatable models that can efficiently predict probability that a user likes an item and/or similarities between items. Smoothing is accomplished by utilizing statistical methods such as support cutoff, single and multiple prior on counts, and prior on measure of association and the like. By improving model-based collaborative filtering with such techniques, performance is increased with regard to product-to-product recommendations. The present invention also provides improvements over systems based on dependency nets (DN) in both areas of quality of recommendations and speed of model creation. It can also be complementary to DN to improve the value of an existing collaborative filtering system's overall efficiency. It is also employable with low frequency user preference data.

Growing community of inventors

Seattle, WA, United States of America

Jesper B Lind