Abstract: Disclosed are systems for determining uncertainty from a deep learning classification model. A system for determining uncertainty from a deep learning classification model may include at least one processor programmed or configured to determine a classification for an input based on a deep learning classification model, generate an uncertainty score for the classification, determine whether the uncertainty score satisfies a threshold, in response to determining that the uncertainty score satisfies the threshold, determine an automated action based on a decision model, and in response to determining that the uncertainty score does not satisfy the threshold, determine the automated action based on at least one predefined rule. Methods and computer program products are also disclosed.

Abstract: Methods, systems, and computer program products are provided for shared latent space-based debiasing. An example system includes at least one processor configured to: transform data from each of a target domain, which lacks protected features, and a separate source domain, which contains these features, into correlated latent representations; jointly train a cross-domain protected group estimator on the representations; and debias a downstream machine learning model an adversarial learning technique that leverages the group estimator.

Abstract: Systems, methods, and computer program products are provided for deep learning with plausible deniability. An example system includes at least one processor configured to: (i) obtain a dataset including a plurality of batches of data samples; (ii) compute a plurality of gradients of the plurality of batches; (iii) select a gradient; (iv) add noise to the selected gradient; (v) determine, based on the noised gradient and the plurality of gradients, a number of gradients that satisfy a privacy function; (vi) in response to the number of gradients that satisfy the privacy function satisfying a threshold number, training, using the noised gradient, a machine learning model; and (vii) in response to the number of gradients that satisfy the privacy function failing to satisfy the threshold number, reshuffling the dataset to generate an updated plurality of batches of data samples and returning to step (ii).

Abstract: Provided is a method for fraud prevention using deep learning and survival models. The method may include receiving, with at least one processor, transaction data associated with a plurality of transactions of at least one payment account. At least one attempted attack may be detected based on the transaction data. A fraud risk score for each subperiod of a plurality of subperiods in a time period following the at least one attempted attack may be generated based on the transaction data using a deep learning model and a survival model. The fraud risk score for each respective subperiod may be associated with a probability that a fraudulent transaction will not occur by the respective subperiod. A system and computer program product are also disclosed.

Abstract: A computer-implemented method includes, for each of a set of training dates: receiving, for each of a sequence of dates including the training date, an input data array representing values of a predetermined set of date-dependent features; receiving a target output corresponding to an evaluation of a predetermined metric at the training date; and performing an update routine including processing the input data array for each date using first layers of a neural network, processing a resulting intermediate data array using second layers of the neural network to generate a network output, and updating values of parameters of the neural network in in a direction of a negative gradient of an error between the target output and the network output. The data processing system is then arranged to generate an embedding array by processing an input data array for each of a given sequence of dates using the first layers of the neural network.

Abstract: Provided are systems for authenticating an individual using image feature templates that include at least one processor to train a first machine learning model based on a training dataset of a plurality of images of a user, generate a plurality of image feature templates using the first machine learning model, wherein each image feature template of the plurality of image feature templates is associated with a positive authentication of the identity of the user during a time interval, generate a second machine learning model based on the plurality of image feature templates, generate a predicted image feature template using the second machine learning model, determine whether to authenticate the identity of the user based on an input image of the user, and perform an action based on determining whether to authenticate the identity of the user. Methods and computer program products are also provided.

Abstract: Provided is a system for segmenting large scale datasets according to machine learning models based on transfer learning that includes at least one processor programmed or configured to train a base machine learning model using a training dataset to generate a trained machine learning model, evaluate the trained machine learning model using an evaluation dataset, wherein, when evaluating the trained machine learning model using the evaluation dataset, the at least one processor is programmed or configured to generate a confidence score for each data instance of the evaluation dataset with the trained machine learning model, augment the evaluation dataset based on the confidence score for each data instance of the evaluation dataset to generate an augmented evaluation dataset, and retrain the trained machine learning model using the augmented evaluation dataset to generate a final machine learning model. Methods and computer program products are also provided.

Abstract: Provided are systems for authenticating an individual using image feature templates that include at least one processor to train a first machine learning model based on a training dataset of a plurality of images of a user, generate a plurality of image feature templates using the first machine learning model, wherein each image feature template of the plurality of image feature templates is associated with a positive authentication of the identity of the user during a time interval, generate a second machine learning model based on the plurality of image feature templates, generate a predicted image feature template using the second machine learning model, determine whether to authenticate the identity of the user based on an input image of the user, and perform an action based on determining whether to authenticate the identity of the user. Methods and computer program products are also provided.

Abstract: Described are a system, method, and computer program product for state compression in stateful machine learning models. The method includes receiving a transaction authorization request for a transaction and loading at least one encoded state of a recurrent neural network (RNN) model from a memory. The method further includes decoding the at least one encoded state by passing each encoded state through a decoder network to provide at least one decoded state. The method further includes generating at least one updated state and an output for the transaction by inputting at least a portion of the transaction authorization request and the at least one decoded state into the RNN model. The method further includes encoding the at least one updated state by passing each updated state through an encoder network to provide at least one encoded updated state, and storing the at least one encoded updated state in the memory.

Abstract: Methods, systems, and computer program products are provided for cleaning noisy data from unlabeled datasets using autoencoders. A method includes receiving training data including noisy samples and other samples. An autoencoder network is trained based on the training data to increase a first metric based on the noisy samples and to reduce a second metric based on the other samples. Unlabeled data including unlabeled samples is received. A plurality of third outputs is generated by the autoencoder network based on the plurality of unlabeled samples. For each respective unlabeled sample, a respective third metric is determined based on the respective unlabeled sample and a respective third output, and whether to label the respective unlabeled sample as noisy or clean is determined based on the respective third metric and a threshold. Each respective unlabeled sample determined to be labeled as noisy is cleaned.

Abstract: Methods for generating fraud detection rules based on transaction data may include receiving historical transaction data, associating tags with each transaction, generating decision trees having root nodes and child nodes operably connected to the respective root nodes, determining at least one primary rule and at least one set of secondary rules associated with the at least one primary rule based on relationships between features of the transactions, assigning primary rules and sets of secondary rules to the at least one decision tree to populate the tree, extracting a plurality of rule sets including at least one primary rule and one or more secondary rules, determining an ordering of the plurality of rule sets; and determining a subset of rule sets from the ordered plurality of rule sets against which subsequently received transactions are compared against to determine if the subsequent transactions are fraudulent.

Abstract: Methods, systems, and computer program products are provided for cleaning noisy data from unlabeled datasets using autoencoders. A method includes receiving training data including noisy samples and other samples. An autoencoder network is trained based on the training data to increase a first metric based on the noisy samples and to reduce a second metric based on the other samples. Unlabeled data including unlabeled samples is received. A plurality of third outputs is generated by the autoencoder network based on the plurality of unlabeled samples. For each respective unlabeled sample, a respective third metric is determined based on the respective unlabeled sample and a respective third output, and whether to label the respective unlabeled sample as noisy or clean is determined based on the respective third metric and a threshold. Each respective unlabeled sample determined to be labeled as noisy is cleaned.

Abstract: Provided is a method for fraud prevention using deep learning and survival models. The method may include receiving, with at least one processor, transaction data associated with a plurality of transactions of at least one payment account. At least one attempted attack may be detected based on the transaction data. A fraud risk score for each subperiod of a plurality of subperiods in a time period following the at least one attempted attack may be generated based on the transaction data using a deep learning model and a survival model. The fraud risk score for each respective subperiod may be associated with a probability that a fraudulent transaction will not occur by the respective subperiod. A system and computer program product are also disclosed.

Abstract: Provided is a method for fraud prevention using deep learning and survival models. The method may include receiving, with at least one processor, transaction data associated with a plurality of transactions of at least one payment account. At least one attempted attack may be detected based on the transaction data. A fraud risk score for each subperiod of a plurality of subperiods in a time period following the at least one attempted attack may be generated based on the transaction data using a deep learning model and a survival model. The fraud risk score for each respective subperiod may be associated with a probability that a fraudulent transaction will not occur by the respective subperiod. A system and computer program product are also disclosed.

Abstract: Methods for generating fraud detection rules based on transaction data may include receiving historical transaction data, associating tags with each transaction, generating decision trees having root nodes and child nodes operably connected to the respective root nodes, determining at least one primary rule and at least one set of secondary rules associated with the at least one primary rule based on relationships between features of the transactions, assigning primary rules and sets of secondary rules to the at least one decision tree to populate the tree, extracting a plurality of rule sets including at least one primary rule and one or more secondary rules, determining an ordering of the plurality of rule sets; and determining a subset of rule sets from the ordered plurality of rule sets against which subsequently received transactions are compared against to determine if the subsequent transactions are fraudulent.

Abstract: Disclosed are systems for determining uncertainty from a deep learning classification model. A system for determining uncertainty from a deep learning classification model may include at least one processor programmed or configured to determine a classification for an input based on a deep learning classification model, generate an uncertainty score for the classification, determine whether the uncertainty score satisfies a threshold, in response to determining that the uncertainty score satisfies the threshold, determine an automated action based on a decision model, and in response to determining that the uncertainty score does not satisfy the threshold, determine the automated action based on at least one predefined rule. Methods and computer program products are also disclosed.

Abstract: Provided is a method for predicting future states based on time series data using feature engineering and/or hybrid machine learning models. The method may include receiving payment transaction data associated with a plurality of payment transactions, the plurality of payment transactions including a first subset of payment transactions associated with a first entity; determining a plurality of features based on the payment transaction data associated with the plurality of payment transactions; inputting the plurality of features into at least one machine learning model to provide at least one prediction of a net settlement position of the first entity; and communicating the at least one prediction of the net settlement position to a first entity system associated with the first entity. A system and computer program product are also disclosed.

Abstract: A system, method, and product for determining a reason for a deep learning model output that obtain training data associated with training samples and first labels for the training samples; train a first model using the training samples and the first labels, training the first model generating predictions for the training samples; train a second model using the training samples and the predictions as second labels for the training samples; extract one or more weights of the trained second model; process, using the first model, input data including features associated with at least one sample, to generate output data, the output data including at least one prediction for the at least one sample; and apply the one or more extracted weights to the features to determine one or more contributions of one or more features of the features to the at least one prediction for the at least one sample.

Abstract: A computer-implemented method includes, for each of a set of training dates: receiving, for each of a sequence of dates including the training date, an input data array representing values of a predetermined set of date-dependent features; receiving a target output corresponding to an evaluation of a predetermined metric at the training date; and performing an update routine including processing the input data array for each date using first layers of a neural network, processing a resulting intermediate data array using second layers of the neural network to generate a network output, and updating values of parameters of the neural network in in a direction of a negative gradient of an error between the target output and the network output. The data processing system is then arranged to generate an embedding array by processing an input data array for each of a given sequence of dates using the first layers of the neural network.

Abstract: Provided are methods for multi-task learning (MTL) in deep neural networks. An exemplary method may include receiving an MTL model; receiving a testing data set comprising testing data items for the MTL model, each testing data item comprising a plurality of elements, each element associated with a respective feature; grouping the features into a plurality of groups based on an impact of each feature on the tasks of the MTL model, determining an overall accuracy score and task-specific accuracy scores based on inputting the testing data to the MTL model; applying feature reduction evaluation (FRE) to provide a feature score for each feature; and adjusting the feature scores based on a respective grouping associated with the respective feature and at least one of the overall accuracy score, the task-specific accuracy scores, or any combination thereof to provide an adjusted feature score. Systems and computer program products are also disclosed.