Abstract: Methods, systems, apparatuses, devices, and computer program products are described. A flow generation service may receive a natural language input that indicates instructions for automating a task according to a first process flow. Using a large language model (LLM), the flow generation service may decompose the natural language input into a set of elements (e.g., logical actions) and connectors, where the LLM may be trained on first metadata corresponding to a second process flow that is created manually by a user. In addition, using the LLM, the flow generation service may generate second metadata corresponding to each of the set of elements based on decomposing the natural language input. The flow generation service may sequence and merge the set of elements to generate the first process flow. In some examples, the flow generation service may send, for display to a user interface of a user device, the first process flow.

Abstract: The present invention provides a deep learning-based natural language understanding method and an intelligent teaching assistant system. First, it involves constructing a knowledge database and a question database, where learning material documents are saved into the knowledge database and preprocessed natural language information is saved into the question database. The method then involves learning and understanding the natural language information in the question database, searching for related knowledge points in the knowledge database based on the understood content, selecting the best-matched learning materials corresponding to these knowledge points as samples to respond to the natural language information, and generating a record that includes the question, response, and evaluation, which is saved into the knowledge database. Finally, it generates multiple forms of responses and outputs them according to the corresponding requirements.

Abstract: System and method for transformer-based adversarial active learning system. A machine learning system includes a generator, a transformer encoder, a classifier, and a discriminator all working in combination to generate and select unlabeled data points for labeling. The system utilizes a generative adversarial network paired with an active learning framework to optimize text embedding and feature encoding according to distribution of training data.

Abstract: Systems, devices, and techniques are disclosed for data shards for distributed processing. Data sets of data for users may be received. The data sets may belong to separate groups. User identifiers in the data sets may be hashed to generate hashed identifiers for the data sets. The user identifiers in the data sets may be replaced with the hashed identifiers. The data sets may be split to generate shards. The data sets may be split into the same number of shards. Merged shards may be generated by merging the shards using a separate running process for each of the merged shards. The merged shards may be generated using shards from more than one of the two or more data sets. An operation may be performed on all of the merged shards.

Abstract: Methods, systems, apparatuses, devices, and computer program products are described. An application server or another device may receive a set of input data associated with an activity between an actor and an electronic communication message (e.g., a marketing email). From the input data, the application server may identify a set of features associated with the activity (an open rate, a click rate, etc.) and a set of source network addresses of respective, known automated scanners. The application server may input the features and source network addresses into a positive-and-unlabeled (PU) learning model, which may output a classification result that indicates a probability that the activity is associated with an automated scanner.

Abstract: Methods, systems, apparatuses, devices, and computer program products are described. A modeling service may generate a set of candidate segments using a set of cluster models and based on a seed segment and entity data. Based on respective features associated with the segments, the service may generate candidate segment fingerprints and a seed segment fingerprint, where a segment fingerprint may indicate a distribution of entities within a segment based on similarities between features associated with entities within the segment. That is, a segment fingerprint may depict how similar entities are in a candidate segment based on different features. The service may calculate similarity scores between the seed segment and the candidate segments using the segment fingerprints, and rank entities in terms of their similarity. The highest ranking entities may be identified from the candidate segments and included in a lookalike segment corresponding to the seed segment.

Abstract: A prediction system of an online system deploys one or more machine-learned architectures to generate predictions. In one embodiment, the machine-learned architecture is a stacked ensemble model. The stacked ensemble model includes a plurality of base models, where a base model is coupled to receive input data and generate a base prediction for the input data. The stacked ensemble model includes a meta model that combines the base predictions to generate a meta prediction for the input data. The prediction system also generates a reliability measure that takes advantage of the base predictions to evaluate the reliability of the meta prediction. In this manner, while the quality of individual predictions may differ from one another depending on the values of the input data, the prediction system can dynamically generate the reliability measure to account for this variation.

Abstract: A system can recommend a next action for a user. A memory can store user data corresponding to the user and can include historic interaction points. A behavior pattern can be identified based on two or more interaction points stored in the user data. An intent of the user based on the behavior pattern can be identified. The intent can be based on a previous behavior pattern of another user. Several probabilities that the user will meet one or more objectives can be determined based on the intent. The probabilities can be scored using and used to assign a policy to the first user. A next action can be recommended based on the policy and executed with respect to the user. The outcome of the recommended next action can be stored to the user data.

Abstract: Methods, systems, and devices for displaying a user interface for frequent pattern (FP) analysis are described. In some cases, data stored at a multi-tenant database server may be analyzed to understand various interactions and patterns between data attributes associated with multiple users, or determine one or more attributes associated with a characterization of an individual (e.g., a persona). The multi-tenant database server may effectively cluster and/or perform calculations on attributes of the data to understand user patterns and determine common personas. The results may then be displayed by a user interface at a user device (e.g., associated with the user).

Abstract: A system can recommend a next action for a user. A memory can store user data corresponding to the user and can include historic interaction points. A behavior pattern can be identified based on two or more interaction points stored in the user data. An intent of the user based on the behavior pattern can be identified. The intent can be based on a previous behavior pattern of another user. Several probabilities that the user will meet one or more objectives can be determined based on the intent. The probabilities can be scored using and used to assign a policy to the first user. A next action can be recommended based on the policy and executed with respect to the user. The outcome of the recommended next action can be stored to the user data.

Abstract: A prediction system of an online system deploys one or more machine-learned architectures to generate predictions. In one embodiment, the machine-learned architecture is a stacked ensemble model. The stacked ensemble model includes a plurality of base models, where a base model is coupled to receive input data and generate a base prediction for the input data. The stacked ensemble model includes a meta model that combines the base predictions to generate a meta prediction for the input data. The prediction system also generates a reliability measure that takes advantage of the base predictions to evaluate the reliability of the meta prediction. In this manner, while the quality of individual predictions may differ from one another depending on the values of the input data, the prediction system can dynamically generate the reliability measure to account for this variation.

Abstract: A data processing server may receive a set of data objects for frequent pattern (FP) analysis. The set of data objects may be analyzed using an attribute diversity technique. For the set of data attributes of the set of data objects, the server may arrange the attributes in one or more dimensions. The server may initialize a set of centroids on data points and identify mean values of nearby data points. Based on an iteration of the mean value calculation, the server may identify a set of attributes corresponding to final mean values as being groups of similarly frequent attributes. These groups of similarly frequent attributes may be analyzed using an FP analysis procedure to identify frequent patterns of data attributes.

Abstract: Methods, computer readable media, and devices for determining a hyperparameter for influencing non-local samples in machine learning are disclosed. One method may include identifying a set of local samples associated with a first entity, identifying a set of non-local samples comprising samples associated with a plurality of entities other than the first entity, assigning a local sample weight to one or more samples of the set of local samples, determining a range of non-local sample weights, determining a range of hyperparameters based on the range of non-local sample weights, determining an optimized hyperparameter based on the range of hyperparameters, assigning an optimized non-local sample weight to one or more samples of the set of non-local samples, and generating a prediction using machine learning.

Abstract: Methods, systems, and devices supporting data processing are described. In some systems, a data processing platform may support communication message analysis using machine learning. For example, a system may receive a set of communication messages (e.g., social media messages) and perform a machine learning process on the message contents and message interaction data to train a machine learned model. The system may further receive a subject line for a communication message for analysis, input the subject line into the machine learned model, and receive, as an output of the machine learned model, an engagement score based on the subject line. The engagement score may indicate an estimated probability that a user receiving the communication message opens the communication message (e.g., based on the subject line). A user—or the system—may modify the subject line based on the analysis to improve the engagement score.

Abstract: Disclosed embodiments are related to send time optimization technologies for sending messages to users. The send time optimization technologies provide personalized recommendations for sending messages to individual subscribers taking into account the delay and/or lag between the send time and the time when a subscriber engages with a sent message. A machine learning (ML) approach is used to predict the optimal send time to send messages to individual subscribers for improving message engagement. The personalized recommendations are based on unique characteristics of each user's engagement preferences and patterns, and deals with historical feedback that is generally incomplete and skewed towards a small set of send hours. The ML approach automatically discovers hidden factors underneath message and send time engagements. The ML model may be a two-layer non-linear matrix factorization model. Other embodiments may be described and/or claimed.

Abstract: A cloud platform supports a digital communication system that identifies recommended communication frequencies based on past communication data. The cloud platform may support blending of weights applied to different engagement rates. Based on the weights, the system identifies recommended frequency ranges to maximize engagement rates, including the blended engagement rate using a redistribution simulation process.

Abstract: Systems, devices, and techniques are disclosed for data shards for distributed processing. Data sets of data for users may be received. The data sets may belong to separate groups. User identifiers in the data sets may be hashed to generate hashed identifiers for the data sets. The user identifiers in the data sets may be replaced with the hashed identifiers. The data sets may be split to generate shards. The data sets may be split into the same number of shards. Merged shards may be generated by merging the shards using a separate running process for each of the merged shards. The merged shards may be generated using shards from more than one of the two or more data sets. An operation may be performed on all of the merged shards.

Abstract: Systems, devices, and techniques are disclosed for localization of machine learning models trained with global data. Data sets of event data for users may be received. The data sets may belong to separate groups. The data sets of event data may be combined to generate a global data set. A matrix factorization model may be trained using the global data set to generate a globally trained matrix factorization model. A localization group data set may be generated including event data from the global data set for users from a first of the groups. The globally trained matrix factorization model may be trained with the localization group data set to generate a localized matrix factorization model for the first of the groups.

Abstract: A system can recommend a next action for a user. A memory can store user data corresponding to the user and can include historic interaction points. A behavior pattern can be identified based on two or more interaction points stored in the user data. An intent of the user based on the behavior pattern can be identified. The intent can be based on a previous behavior pattern of another user. Several probabilities that the user will meet one or more objectives can be determined based on the intent. The probabilities can be scored using and used to assign a policy to the first user. A next action can be recommended based on the policy and executed with respect to the user. The outcome of the recommended next action can be stored to the user data.

Abstract: Methods, systems, and devices supporting epsilon (?)-closure for frequent pattern (FP) analysis are described. Some database systems may analyze data sets to determine FPs. In some cases, the FP set may include a large number of semi-redundant patterns, resulting in significant memory or processing overhead. To reduce the redundancy of these patterns, the database system may implement pre-configured or dynamic threshold occurrence differences (e.g., ? values) to test against related patterns. For example, the database system may calculate the difference between the data objects covered by a sub-pattern and a super-pattern (e.g., where the super-pattern includes all the same data attributes of the sub-pattern, plus one additional attribute). This difference may be compared to a corresponding ? value, and if the difference is less than the ? value, the database system may remove one of the patterns (e.g., the sub-pattern) from the set of valid FPs to limit redundancy.