Systems and methods for detecting user account fraud rings are disclosed. In an embodiment, a computer system may access a plurality of user accounts created within a past period. The computer system may generate a tree of user accounts by recursively identifying pairs of user accounts by beginning with a seed account for the tree and iterating through user account pairs at lower branch levels to determine whether each user account has been paired to one or more other user accounts based on respective hard link features and soft link features. If a user account has been paired to one or more other user accounts, the computer system adds the one or more other user accounts to a branch level below the user account in the tree. The user accounts of the tree may be included in a cluster. Actions can be taken against the user accounts in the cluster.

Systems and methods for detecting user account fraud rings are disclosed. In an embodiment, a computer system may access a plurality of user accounts created within a past period. The computer system may generate a tree of user accounts by recursively identifying pairs of user accounts by beginning with a seed account for the tree and iterating through user account pairs at lower branch levels to determine whether each user account has been paired to one or more other user accounts based on respective hard link features and soft link features. If a user account has been paired to one or more other user accounts, the computer system adds the one or more other user accounts to a branch level below the user account in the tree. The user accounts of the tree may be included in a cluster. Actions can be taken against the user accounts in the cluster.

Methods, apparatus, systems, and articles of manufacture for user identification via community detection are disclosed. Example instructions, when executed, cause at least one processor to at least access personally identifiable information to device links, build a device graph based on the personally identifiable information to device links, split components of the device graph into person clusters using community detection, create a snapshot including a device-to-person link lookup, and prepare a person-level impression measurement report from the snapshot.

The techniques described herein relate to computerized methods and apparatus for grouping images of objects based on semantic and visual information associated with the objects. The techniques described herein further relate to computerized methods and apparatus for training a machine learning model for object recognition.

A computer-implemented method can facilitate delivery of targeted content to user devices in situations in which historic tracking data (e.g., cookie data) is generally unavailable and/or unreliable. A p-dimensional embedding of websites can be generated based on a group of user devices for whom tracking data is available. Conversion event data that indicates indicating whether that audience member performed a conversion action can be received. A machine learning model can be trained using the conversion event data and the positions of websites appearing in the conversion event data within the p-dimensional embedding to predict a likelihood of conversion and/or a type of content to provide given a position in the p-dimensional embedding. When an indication that a user device is accessing a website is received, a position of that website in the p-dimensional embedding can be determined and targeted content can be delivered to the user device.

A computer-implemented method can facilitate delivery of targeted content to user devices in situations in which historic tracking data (e.g., cookie data) is generally unavailable and/or unreliable. A p-dimensional embedding of websites can be generated based on a group of user devices for whom tracking data is available. Conversion event data that indicates indicating whether that audience member performed a conversion action can be received. A machine learning model can be trained using the conversion event data and the positions of websites appearing in the conversion event data within the p-dimensional embedding to predict a likelihood of conversion and/or a type of content to provide given a position in the p-dimensional embedding. When an indication that a user device is accessing a website is received, a position of that website in the p-dimensional embedding can be determined and targeted content can be delivered to the user device.

A novel and useful system and method of improved power performance and lowered memory requirements for an artificial neural network based on packing memory utilizing several structured sparsity mechanisms. The invention applies to neural network (NN) processing engines adapted to implement mechanisms to search for structured sparsity in weights and activations, resulting in a considerably reduced memory usage. The sparsity guided training mechanism synthesizes and generates structured sparsity weights. A compiler mechanism within a software development kit (SDK), manipulates structured weight domain sparsity to generate a sparse set of static weights for the NN. The structured sparsity static weights are loaded into the NN after compilation and utilized by both the structured weight domain sparsity mechanism and the structured activation domain sparsity mechanism. The application of structured sparsity lowers the span of search options and creates a relatively loose coupling between the data and control planes.

An embodiment generates an initial set of training data from monitoring data. The initial set of training data is generated by combining outputs from a plurality of pretrained classifiers. The embodiment trains a new classification model using the initial set of training data to identify anomalies in monitoring data. The embodiment performs a multiple-level clustering of the data samples resulting in a plurality of clusters of sub-clusters of data samples, and generates a review list of data samples by selecting a representative data sample from each of the clusters. The embodiment receives an updated data sample from the expert review that includes a revised target classification for at least one of the data samples of the expert review list. The embodiment then trains another replacement classification model using a revised set of training data that includes the updated data sample and associated revised target classification.

An embodiment generates an initial set of training data from monitoring data. The initial set of training data is generated by combining outputs from a plurality of pretrained classifiers. The embodiment trains a new classification model using the initial set of training data to identify anomalies in monitoring data. The embodiment performs a multiple-level clustering of the data samples resulting in a plurality of clusters of sub-clusters of data samples, and generates a review list of data samples by selecting a representative data sample from each of the clusters. The embodiment receives an updated data sample from the expert review that includes a revised target classification for at least one of the data samples of the expert review list. The embodiment then trains another replacement classification model using a revised set of training data that includes the updated data sample and associated revised target classification.

A data prediction subsystem includes receives event data indicating amounts of items removed from locations over a previous period of time. For a first day of the first set of event data having zero events or an empty status indicating that the first item is not believed to be present at the first location, longitudinal and cross-sectional components are determined. An anticipated event value for the first item at the first location is determined using the longitudinal component and the cross-sectional component. Based at least in part on the anticipated event value, a prediction value is determined that corresponds to a recommended amount of the first item to request at a future time.