Abstract: Certain aspects of the present disclosure provide techniques for training and using a machine learning model to extract relevant textual content for custom fields in a software application from freeform text samples. An example method generally includes generating, via a natural language processing pipeline, a training data set from a data set of freeform text samples and field entries for a plurality of custom fields defined in a software application. A first machine learning model is trained to identify custom fields for which relevant data is included in freeform text. A second machine learning model is trained to extract content from the freeform text into one or more custom fields of the plurality of custom fields defined in the software application and identified by the first machine learning model as custom fields for which relevant data is included in the freeform text.

Abstract: Techniques for detecting fraud may include mapping routing numbers of one or more financial institutions with geolocation data of the financial institutions; obtaining a geolocation of a user based on the user's internet protocol (IP) address; obtaining a first user input from the user indicating a first financial institution; generating a match score for each of the one or more financial institutions that indicates a level of match between the first user input and the respective financial institution; boosting the match score for each financial institution based on its location with respect to the geolocation of the user; generating a list of financial institutions having the boosted match score above a threshold; obtaining a second user input from the user indicating at least one second financial institution; and presenting search results to the user based on the second user input, wherein the search results are boosted.

Abstract: A processor may receive an image of a user and an avatar representing the user within a computing environment. The processor may generate a score for the avatar on the basis of its resemblance to the image using a machine learning (ML) process. The processor may configure at least one option for action by the user within the computing environment according to the score.

Abstract: Methods and systems for efficiently generating tagged training data for machine learning models. In conventional systems, all of the raw data (e.g., each sentence) has to be manually tagged. Instead, the methods and systems generate a representative sample for multiple portions of raw data, e.g., a representative sentence for multiple, similar sentences. Only the representative sample is tagged and used for training, thereby realizing a significant efficiency in both tagging the data and training the machine learning models.

Abstract: Certain aspects of the present disclosure provide techniques and systems for screening chat attachments. A chat attachment screening system monitors a chat window of a first computing device associated with a first user during an interaction session between the first user and a second user. An upload of an attachment is detected based on the monitoring. Access to the attachment from a second computing device associated with the second user is blocked, in response to detecting the upload. Content from the attachment is identified and extracted. A type of the attachment is determined based on the content. A determination is made as to whether the second user is authorized to access the type of the attachment. An indication of the determination is presented on at least one of the first computing device or the second computing device during the interaction session.

Abstract: Embodiments disclosed herein select a content message to present to a user on a page of an application based on paralinguistic features of audio input received from the user for the application. The audio input is received via a microphone associated with a computing device. A feature extractor extracts paralinguistic features from the audio input. A predictive model determines a label indicating a measure of receptiveness to product placement (e.g., a predicted marketing outcome) based on the paralinguistic features. A content-selection component selects a content message to present to the user based on the label and based on a profile of the user.

Abstract: Embodiments disclosed herein provide language-agnostic routing prediction models. The routing prediction models input text queries in any language and generate a routing prediction for the text queries. For a language that may have sparse training text data, the models, which are machine learning models, are trained using a machine translation to a prevalent language (e.g., English) to the language having sparse training text data -with the original text corpus and the translated text corpus being an input to multi-language embedding layers. The trained machine learning model makes routing predictions for text queries for the language having sparse training text data.

Abstract: Knowledge about a user is used to determine whether one or more messages received by the user are malicious. The knowledge about the user may be based on the user's financial history such as transaction records. Particularly, a classifier model is trained on a supervised approach using a dataset containing, for example, a categorization of incoming messages (e.g., password change message), the user's aggregated transaction records, message attributes, user attributes, and corresponding classification labels. After the training, the classifier model is deployed to determine whether an incoming message is malicious.

Abstract: A processor may receive an image and determine a number of foreground pixels in the image. The processor may obtain a result of optical character recognition (OCR) processing performed on the image. The processor may identify at least one bounding box surrounding at least one portion of text in the result and overlay the at least one bounding box on the image to form a masked image. The processor may determine a number of foreground pixels in the masked image and a decrease in the number of foreground pixels in the masked image relative to the number of foreground pixels in the image. Based on the decrease, the processor may modify an aspect of the OCR processing for subsequent image processing.

Abstract: A method includes extracting attribute values of attributes from login events, filtering the attribute values based on correlation between the attributes and classes to obtain filtered attributes values, and generating a vector embedding of the filtered attributes values to obtain login vectors. The method further includes executing a sequential machine learning model on the login vectors to determine a class of the classes, and outputting the class.

Abstract: A method implements churn prevention using graphs. The method includes receiving clickstream data, which includes an event, of a user session with an application. The method further includes identifying the event as corresponding to a churn user account and mapping the event to a pair of nodes of a graph. The method further includes updating a churn user count of the pair of nodes in response to identifying the event as corresponding to the churn user account. The method further includes identifying an edge of the graph, corresponding to the pair of nodes. The method further includes updating a value of the edge using an active user count and the churn user count presenting an update responsive to the value.

Abstract: A method implements churn prediction using clickstream data. The method includes receiving clickstream data of a user and converting the clickstream data to a token list. The method further includes processing the token list with a first recurrent layer, a second recurrent layer, and an attention layer of a machine learning model to generate a churn risk. The method further includes executing a reactivation action in response to the churn risk.

Abstract: Certain aspects of the present disclosure provide techniques for generating a recommendation of third-party applications to a user by a recommendation engine. The recommendation engine includes two deep-learning models that use various data sources (e.g., user data and application data) to generate the recommendation. One deep-learning model generates a relevance score for each available third-party application. The relevance score is used to determine a relevant application(s). The other deep-learning model generates a connection score for each relevant application. The recommendation engine uses the relevance score and the connections to generate an engagement score for each relevant application to determine whether the user would use the third-party application if recommended to the user. Those relevant applications with an engagement score that meet pre-determined criteria are determined and displayed to the user in the application as a recommendation.

Abstract: Aspects of the present disclosure provide techniques for categorical anomaly detection. Embodiments include receiving values for a plurality of data categories for an entity of a plurality of entities. Embodiments include generating a feature vector for the entity based on the values, the feature vector excluding a first value for a first data category of the plurality of data categories. Embodiments include providing one or more inputs to a machine learning model based on the feature vector and determining, based on one or more outputs received from the machine learning model, one or more other entities of the plurality of entities that are grouped with the entity. Embodiments include determining that the first value is anomalous based on respective values for the first data category for the one or more other entities. Embodiments include performing one or more actions based on the determining that the first value is anomalous.

Abstract: A method identifying clusters with anomaly detection. The method includes aggregating a set of events, of a user, to generate a user vector in response to identifying an event of the set of events. The method further includes aggregating a set of user vectors to a periodic vector for a time period. The method further includes processing a set of periodic vectors to generate a periodic distance. The method further includes selecting the time period, corresponding to the periodic vector, using the periodic distance and a threshold. The method further includes processing the set of user vectors to generate clusters of user vectors, wherein the set of user vectors includes the event during the time period. The method further includes processing the clusters of user vectors to identify a selected cluster and performing an action to a set of user accounts corresponding to the selected cluster.

Abstract: In one aspect, the present disclosure relates to a method of generating business descriptions performed by a server, said method may include: receiving a plurality of invoices, each invoice being associated with a business of a plurality of businesses; extracting a plurality of texts from the plurality of invoices; embedding the plurality of texts to a vector space to obtain a plurality of invoice vectors; generating a plurality of clusters in the vector space, each cluster of the plurality of clusters comprising at least one invoice vector of the plurality of invoice vectors; generating a description for a cluster, the description for the cluster representing all invoice vectors assigned to the cluster; for each business of the plurality of businesses that has at least one invoice vector assigned to the cluster, associating the business with the description; and indexing the plurality of businesses within a database by the generated descriptions.

Abstract: A computer-implemented system and method for generating and implementing real-time optimized savings recommendations during online purchase checkout processes. The recommendations may be in the form of personalized digital nudges designed to influence the user in a manner that furthers a savings goal.

Abstract: Systems and methods for generating a custom document template are disclosed. An example method may be performed by one or more processors of a system and include retrieving a user document including a user data entry in a user data field, identifying a set of system data fields within a plurality of system documents potentially relevant to the user document, determining, for each of the set of system data fields, a weighted value indicative of a likelihood that the system data field is relevant to the user data field, identifying a most relevant system data field of the set of system data fields, the most relevant system data field having a highest weighted value of the determined weighted values, and generating a custom document template including a dynamic data region for the user data entry, the dynamic data region mapped to the most relevant system data field.

Abstract: The present disclosure provides techniques for recommending vendors using machine learning models. One example method includes generating a dependency graph based on one or more microservices in a transaction, wherein each node of the dependency graph corresponds to each microservice of the one or more microservices in the transaction, respectively, generating a first set of features using the dependency graph, predicting a first set of nodes that are likely to fail using a first machine learning model based on the first set of features, generating a second set of features using the dependency graph, predicting a second set of nodes that are likely to fail using a second machine learning model based on the second set of features, and applying one or more fixes to one or more microservices based on a combination of the first set of nodes and the second set of nodes.