Abstract: The present disclosure provides a method of building a sorting model, and an application method and apparatus based on the model. The method of building a sorting model comprises: obtaining, from a search log, a query including a relationship triple and a clicked title of a search result corresponding to the query, wherein the relationship triple includes a content word pair and a relationship word of the content word pair; obtaining training data using the obtained query, the clicked title corresponding to the query, and times of click of the clicked title; using the training data to train a neural network-based sorting model, the sorting model being used to sort sentences according to the sentences' description of a relationship of the content word pair.

Abstract: System and method for detecting domain names that exhibit Domain Generation Algorithm (DGA) like behaviours from a stream of Domain Name System (DNS) records. In particular, this document describes a system comprising a deep learning classifier (DL-C) module for receiving and filtering the stream of DNS records before the filtered DNS records, which have been determined to possess domain names that exhibit DGA behaviour are provided to a series filter-classifier (SFC) module. The SFC module then groups the records into various series based on source IP, destination IP and time. For each series, it then filters away records that do not exhibit the dominant DGA characteristics of the series. Finally, for each series, it makes use of the remaining DNS records' timestamps to generate a time series of DGA occurrences and then, using this time series of occurrences, determine the number of DGA bursts throughout the time period of analysis.

Abstract: Disclosed are systems and methods for determining the best time to send an electronic communication from a sender to a recipient. In one aspect, a method is disclosed that includes selecting a time window from a series of candidate time windows based on a corresponding first value for each candidate time window, wherein each first value is representative of a likelihood of receiving an event notification within a specified first delay after the candidate time window. The method further includes selecting a time period from a plurality of time periods within the selected time window based on a corresponding second value for each time period representative of the likelihood of receiving the event notification within a specified second delay after the time period. The method further includes generating a signal indicative of a time within the selected time period at which an electronic communication should be sent.

Abstract: A set of profile parameters to characterize an unknown group of servers is computed. A set of known groups of servers is selected from a historical repository of known group of servers. A subset of known group is selected such that each known group in the subset has a corresponding similarity distance that is within a threshold similarity distance from the unknown group. A decision tree is constructed corresponding to a known group in the subset, by cognitively analyzing a usage of the set of profile parameters of the unknown group in the known group. Using the decision tree a number of problematic servers is predicted in the unknown group. When the predicted number of problematic servers does not exceed a threshold number, a post-prediction action is caused to occur on the unknown group, which causes a reduction in an actual number of problematic servers in the unknown group.

Abstract: The present disclosure provides systems and methods for on-device machine learning. In particular, the present disclosure is directed to an on-device machine learning platform and associated techniques that enable on-device prediction, training, example collection, and/or other machine learning tasks or functionality. The on-device machine learning platform can include a context provider that securely injects context features into collected training examples and/or client-provided input data used to generate predictions/inferences. Thus, the on-device machine learning platform can enable centralized training example collection, model training, and usage of machine-learned models as a service to applications or other clients.

Abstract: A method for generating data standardization rules includes receiving a training data set containing tokenized and tagged data values. A set of machine mining models is built using different learning algorithms for identifying tags and tag patterns using the training set. For each data value in a further data set: a tokenization is applied on the data value, resulting in a set of tokens. For each token of the set of tokens one or more tag candidates are determined using a lookup dictionary of tags and tokens and/or at least part of the set of machine mining models, resulting for each token of the set of tokens in a list of possible tags. Unique combinations of the sets of tags of the further data set having highest aggregated confidence values are provided for use as standardization rules.

Abstract: The invention notably relates to a computer-implemented method for updating a model of a machine learning system. The method comprises providing a first set of observations of similar events, each observation being associated with one or more variables, each variable being associated with a value, and with a target value; indexing each observation of the first set with its corresponding one or more variables and target value; receiving, on the index, a query allowing a selection of a subset of the first set of observations; returning, as a result of the query, a subset of the first set of observations; providing a second model; training the provided second model with the returned subset of the first set of observations; and loading the trained second model.

Abstract: An individual ising device connected to common buses includes neuron circuits, a memory, and a router. The memory holds connection destination information per neuron circuit. An individual item of connection destination information includes first address information identifying one of a plurality of connection destination neuron circuits of a neuron circuit and second address information identifying a first ising device including at least one of the connection destination neuron circuits, the first and second address information being correlated.

Abstract: Disclosed is an apparatus for generating an artificial-neural-network-based prediction model. The apparatus includes an input data conversion unit configured to convert input data of an L-dimensional array (L is a natural number) into normalized vector data and input the normalized vector data, a modeling unit configured to model an artificial-neural-network-based prediction model for learning the input vector data and output a value predicted through the modeling, and an adjustment unit configured to compare the value predicted by the modeling unit with an actually measured value to calculate an error value and adjust learning parameters of an artificial neural network using the error value and a back-propagation algorithm.

Abstract: There is provided a method of generating a machine learning (ML) model ensemble for computing likelihood of an entity failing to meet a target parameter, comprising: training ML-sub-models that each output sub-values for an input of raw data elements, training a principal ML model that outputs a value of an entity parameter corresponding to the target parameter for an input of the sub-values, using a training dataset including for sample entities, the ML-sub-values and corresponding entity parameters, inputting raw data elements associated with the entity into the ML-sub-models to obtain respective sub-values, in iterations: computing simulated adjustments to the sub-values to generate adjusted sub-values that are inputted into the principal ML model to obtain simulated values for the entity parameter, and generating a risk classifier that generates a likelihood of the entity failing to meet the target parameter according to an analysis of the simulated values for the entity parameter.

Abstract: Methods and systems for pruning a convolutional neural network (CNN) include calculating a sum of weights for each filter in a layer of the CNN. The filters in the layer are sorted by respective sums of weights. A set of m filters with the smallest sums of weights is filtered to decrease a computational cost of operating the CNN. The pruned CNN is retrained to repair accuracy loss that results from pruning the filters.