Abstract: A computer-implemented method and system having computer-executable instructions stored in a memory for processing user behavior features by neural networks to identify user segments. The method includes receiving user datasets from a database along with respective user identifiers, retention labels, static user features and interactive user features associated with an online product during a time period. A first neural network processes the interactive user features to generate a time distributed concatenation representation. A second neural network is configured to generate a vector by embedding the time distributed concatenation representation and the static user features through an embedding layer. The second neural network is configured to process the vector through a plurality of layers. A cluster model is used to determine user segments based on values extracted from nodes of a second to last layer of the second neural network.

Abstract: A computer-implemented system and method for generating heterogeneous graph feature embeddings for feature learning and prediction. An application server may receive and process a plurality of feature datasets to generate a graph data structure comprising a plurality of interconnected transaction pairs. The application server processes the graph data structure to determine a first-order transaction pair corresponding to a maximum transaction frequency based on a user identifier; executes a jumping probability algorithm to process the graph data structure to determine a second-order transaction pair jumping from a first-order transaction pair; and generates a transaction sequence associated with the user identifier.

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: A computer-implemented system and method for generating heterogeneous graph feature embeddings for feature learning and prediction. An application server may receive and process a plurality of feature datasets to generate a graph data structure comprising a plurality of interconnected transaction pairs. The application server processes the graph data structure to determine a first-order transaction pair corresponding to a maximum transaction frequency based on a user identifier; executes a jumping probability algorithm to process the graph data structure to determine a second-order transaction pair jumping from a first-order transaction pair; and generates a transaction sequence associated with the user identifier.

Abstract: A method implements calibrated risk scoring and sampling. Features are extracted from a record. A risk score, associated with the record, is generated from the features using a machine learning model. The record is mapped to a risk bucket using the risk score. The risk bucket may include multiple risk bucket records. The record is selected from the risk bucket records with a sampling threshold corresponding to the risk bucket. A form prepopulated with values from the record is presenting to a client device.

Abstract: A method may include generating a vector from unstructured data included in an untransformed transaction, and determining, for the vector, a cluster ID of cluster IDs by matching the vector with a matching cluster vector of cluster vectors. The method may further include generating a query using the cluster ID and the untransformed transaction, and transforming, using the cluster IDs, untransformed transactions to transformed transactions. The transformed transactions may each include a cluster ID. The method may further include generating, using the query, a query result from features of the transformed transactions, generating a fraud score using the query result, and presenting the fraud score and the cluster ID.

Abstract: A computer-implemented system and method for generating heterogeneous graph feature embeddings for feature learning and prediction. An application server may receive and process a plurality of feature datasets to generate a graph data structure comprising a plurality of interconnected transaction pairs. The application server processes the graph data structure to determine a first-order transaction pair corresponding to a maximum transaction frequency based on a user identifier; executes a jumping probability algorithm to process the graph data structure to determine a second-order transaction pair jumping from a first-order transaction pair; and generates a transaction sequence associated with the user identifier.

Abstract: A method implements a customer recognition system. A request with an identifier of an unidentified user is received. Sparse data is generated from string information corresponding to the identifier. Preexisting identifiers are filtered to generate a list of candidate identifiers using the sparse data. The plurality of preexisting identifiers correspond to a plurality of preexisting users. A core identifier is selected by determining a match between the identifier and a preexisting identifier from the preexisting identifiers using distance information generated using the list of candidate identifiers. The core identifier is matched to the identifier using the match to identify the unidentified user as a preexisting user from the plurality of preexisting users.

Abstract: Certain aspects of the present disclosure provide techniques for detecting fraudulent transactions in a transaction processing system. An example method generally includes receiving a request to process a transaction. An input data set including a vector representing the transaction and a plurality of vectors representing historical transactions is generated. The input data set is divided into a plurality of ragged tensors corresponding to non-overlapping time segments of variable length and having a plurality of vectors associated with dates within each time segment A reduced input data set is generated by generating, for each respective ragged tensor of the plurality of ragged tensors, a respective representative vector using max pooling over vectors in the ragged tensor. A fraudulent transaction score is generated based on the reduced input data set using a fraud detection model. The transaction is processed based, at least in part, on the fraudulent transaction score.

Abstract: A data management system predicts whether users will continue using the data management system. The data management system includes an analysis model that generates user retention prediction data based on time dependent user data and static user data. The analysis model also generates recommended actions to be taken by the data management system to increase the probability of retaining the user.

Abstract: Certain aspects of the present disclosure provide techniques for detecting fraudulent transactions in a transaction processing system. An example method generally includes receiving a request to process a transaction. An input data set including a vector representing the transaction and a plurality of vectors representing historical transactions is generated. The input data set is divided into a plurality of ragged tensors corresponding to non-overlapping time segments of variable length and having a plurality of vectors associated with dates within each time segment A reduced input data set is generated by generating, for each respective ragged tensor of the plurality of ragged tensors, a respective representative vector using max pooling over vectors in the ragged tensor. A fraudulent transaction score is generated based on the reduced input data set using a fraud detection model. The transaction is processed based, at least in part, on the fraudulent transaction score.

Abstract: A computer-implemented system and method for generating heterogeneous graph feature embeddings for feature learning and prediction. An application server may receive and process a plurality of feature datasets to generate a graph data structure comprising a plurality of interconnected transaction pairs. The application server processes the graph data structure to determine a first-order transaction pair corresponding to a maximum transaction frequency based on a user identifier; executes a jumping probability algorithm to process the graph data structure to determine a second-order transaction pair jumping from a first-order transaction pair; and generates a transaction sequence associated with the user identifier.

Abstract: Aspects of the present disclosure provide techniques for behavior prediction. Embodiments include receiving activity data of a user, identifying user sessions comprising sets of time-stamped actions in the activity data, and segmenting the activity data into subsets corresponding to the user sessions. Embodiments include providing the subsets as inputs to a hierarchical attention time-series (HAT) model comprising: a first layer that determines attention scores for respective time-stamped actions in the subsets; and a second layer that determines attention scores for the subsets based on aggregations of the attention scores for the respective time-stamped actions.

Abstract: Certain aspects of the present disclosure provide techniques for detecting fraudulent transactions in a transaction processing system. An example method generally includes receiving a request to process a transaction. An input data set including a vector representing the transaction and a plurality of vectors representing historical transactions is generated. The input data set is divided into a plurality of ragged tensors corresponding to non-overlapping time segments of variable length and having a plurality of vectors associated with dates within each time segment A reduced input data set is generated by generating, for each respective ragged tensor of the plurality of ragged tensors, a respective representative vector using max pooling over vectors in the ragged tensor. A fraudulent transaction score is generated based on the reduced input data set using a fraud detection model. The transaction is processed based, at least in part, on the fraudulent transaction score.

Abstract: Certain aspects of the present disclosure provide techniques for displaying sentiment of a user text comment. One example method generally includes receiving a text comment comprising a sequence of words, providing a vector sequence representing the sequence of words to a sentiment model configured to output a sequence of sentiment scores for the vector sequence and providing cleaned text to a topic module configured to output relevance scores. The method further includes receiving, from the sentiment model, the sequence of sentiment scores for the vector sequence and receiving, from the topic module, the relevance scores for the cleaned text. The method further includes determining, final sentiment scores for each word of the sequence of words and generating a sentiment visualization for the sequence of words showing the final sentiment scores corresponding to each word of the sequence of words.

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: A computer-implemented method and system having computer-executable instructions stored in a memory for processing user behavior features by neural networks to identify user segments. The method includes receiving user datasets from a database along with respective user identifiers, retention labels, static user features and interactive user features associated with an online product during a time period. A first neural network processes the interactive user features to generate a time distributed concatenation representation. A second neural network is configured to generate a vector by embedding the time distributed concatenation representation and the static user features through an embedding layer. The second neural network is configured to process the vector through a plurality of layers. A cluster model is used to determine user segments based on values extracted from nodes of a second to last layer of the second neural network.

Abstract: Certain aspects of the present disclosure provide techniques for displaying sentiment of a user text comment. One example method generally includes receiving a text comment comprising a sequence of words, providing a vector sequence representing the sequence of words to a sentiment model configured to output a sequence of sentiment scores for the vector sequence and providing cleaned text to a topic module configured to output relevance scores. The method further includes receiving, from the sentiment model, the sequence of sentiment scores for the vector sequence and receiving, from the topic module, the relevance scores for the cleaned text. The method further includes determining, final sentiment scores for each word of the sequence of words and generating a sentiment visualization for the sequence of words showing the final sentiment scores corresponding to each word of the sequence of words.

Abstract: Certain aspects of the present disclosure provide techniques for displaying sentiment of a user text comment. One example method generally includes receiving a text comment comprising a sequence of words, providing a vector sequence representing the sequence of words to a sentiment model configured to output a sequence of sentiment scores for the vector sequence and providing cleaned text to a topic module configured to output relevance scores. The method further includes receiving, from the sentiment model, the sequence of sentiment scores for the vector sequence and receiving, from the topic module, the relevance scores for the cleaned text. The method further includes determining, final sentiment scores for each word of the sequence of words and generating a sentiment visualization for the sequence of words showing the final sentiment scores corresponding to each word of the sequence of words.

Abstract: Certain aspects of the present disclosure provide techniques for displaying sentiment of a user text comment. One example method generally includes receiving a text comment comprising a sequence of words, providing a vector sequence representing the sequence of words to a sentiment model configured to output a sequence of sentiment scores for the vector sequence and providing cleaned text to a topic module configured to output relevance scores. The method further includes receiving, from the sentiment model, the sequence of sentiment scores for the vector sequence and receiving, from the topic module, the relevance scores for the cleaned text. The method further includes determining, final sentiment scores for each word of the sequence of words and generating a sentiment visualization for the sequence of words showing the final sentiment scores corresponding to each word of the sequence of words.