Amazon Technologies, Inc.

Archiman Dutta

Xianshun Chen

Avik Sinha

Nikhil Anand Navali

Deepak Kumar Nayak

Kai Chieh Liu

Lichao Wang

Shoubhik Bhattacharya

Subhadeep Chakraborty

Rahul Gupta

Dmitry Vladimir Zhiyanov

Lohith Ravi

Nathan Rose

Kai Liu

Dhinesh Kumar Dhanasekaran

Meghana Shivanand Rajamane

Zahed Patel

Multiple distinct control descriptors, each specifying an algorithm and values of one or more parameters of the algorithm, are created. A plurality of tuples, each indicating a respective record of a data set and a respective descriptor, are generated. The tuples are distributed among a plurality of compute resources such that the number of distinct descriptors indicated in the tuples received at a given resource is below a threshold. The algorithm is executed in accordance with the descriptors' parameters at individual compute resources.

Artificial intelligence system providing automated distributed training of machine learning models

Devices, systems, and methods are provided for brand matching using multi-layer machine learning. A method may include generating, based on a first embedding vector and a second embedding vector as inputs to a twin neural network, a third embedding vector and a fourth embedding vector; generating, based on the first embedding vector and the second embedding vector as inputs to a difference neural network, a difference vector indicative of a difference between the first embedding vector and the second embedding vector; generating a concatenated vector by concatenating the third embedding vector with the fourth embedding vector and the difference vector; generating, based on the concatenated vector as an input to a feedforward neural network (FFN), a score between zero and one, the score indicative of a relationship between a first entity and a second entity.

Enhanced brand matching using multi-layer machine learning

A hierarchical representation of an input data set comprising similarity scores for respective entity pairs is generated iteratively. In a particular iteration, clusters are obtained from a subset of the iteration's input entity pairs which satisfy a similarity criterion, and then spanning trees are generated for at least some of the clusters. An indication of at least a representative pair of one or more of the clusters is added to the hierarchical representation in the iteration. The hierarchical representation is used to respond to clustering requests.

Scalable hierarchical clustering

Respective labels are generated automatically for a plurality of record pairs, with a label for a given pair indicating a relationship detected between the records of the pair. One or more machine learning models are trained using the labeled record pairs. The trained versions of the models are stored.

Artificial intelligence system with unsupervised model training for entity-pair relationship analysis

Entity record pairs are extracted from a selected cluster of entity records. Attribute value pairs are obtained from the entity record pairs. Labels are assigned to the attribute value pairs based at least in part on entity-level similarity scores of the entity records from which the attribute value pairs were obtained. A machine learning model is trained, using a data set which includes at least some attribute value pairs to which the labels are assigned, to generate attribute similarity scores for pairs of attribute values.

Artificial intelligence system using unsupervised transfer learning for intra-cluster analysis

Respective text feature sets and non-text feature sets are generated corresponding to individual pairs of a plurality of record pairs. At least one text feature is based on whether a text token exists in both records of a pair. Perceptual hash values are used for non-text feature sets. A machine learning model is trained, using the text and non-text feature sets, to generate relationship scores for record pairs. The model includes a text sub-model and a non-text sub-model.

Artificial intelligence system employing multimodal learning for analyzing entity record relationships

Methods, systems, and computer-readable media for taxonomy generation with automated analysis and auditing are disclosed. A primary classification is determined for a hierarchical taxonomy of items in a marketplace. The primary classification is selected from a plurality of terms describing items in the marketplace, and the primary classification is selected based at least in part on automated analysis of the terms. A plurality of secondary classifications are determined for the hierarchical taxonomy. The secondary classifications are selected from the terms describing the items in the marketplace, and the secondary classifications are selected based at least in part on automated analysis of the terms. The hierarchical taxonomy is modified based at least in part on feedback from a plurality of users. The feedback comprises one or more terms entered by one or more of the users to filter a set of items.

Taxonomy generation with statistical analysis and auditing

Described are techniques for modifying or creating classification data used to automatically classify items in an online marketplace or catalog, based on user interaction data. For one or more classification labels that may be applied to an item, user interaction data indicative of a count of instances that the label was accessed, a length of time during which the label was accessed, counts of instances that parent and child labels were accessed, and counts of instances that the label was accessed via a search query may be determined. Based on the user interaction data, an importance score for the label may be determined. Labels having an importance score greater than or equal to a threshold value may be included in classification data and used for subsequent classification of items. Labels having an importance score less than a threshold may be excluded from the classification data.

System for classifying items based on user interactions

A technology for determining accuracy estimates for classifications used in an electronic catalog. In one example, classifications for product groupings included in an electronic catalog may be updated as a result of the classifications inaccurately representing products included in the product groupings. The electronic catalog of products may be grouped into a plurality of product groupings using classifications. Classifications of product groupings that inaccurately represent products included in the product grouping may be updated with suggested classifications. Update metrics for updates made to the grouping classifications may be collected and the update metrics may be used to calculate an accuracy estimate for the classifications used in the electronic catalog.

Classification accuracy estimation

Online service providers may operate a rendering service for generating and providing substitute web content information for rendering substitute web content instead of authentic web content. The rendering service may obtain web content information for the authentic web content in response to receiving a request for web content. The rendering service may use the web content information to generate the substitute web content information. The substitute web content information is useable by the computing device to generate substitute web content that includes one or more visual elements resembling resource objects of the authentic web content. The visual elements are rendered, as a result of processing by the computing device, as image content instead of interactive objects.

Substitute web content generation for detection and avoidance of automated agent interaction

Techniques are provided herein for utilizing a classification engine to improve a classification model. For example, a classification engine may derive a statistical model based at least in part on a synthetic data set. A misclassification may be determined based at least in part on an output of the statistical model. An audit question may be provided to an individual, the audit question being determined based at least in part on the determined misclassification. Response data related to the audit question may be received. The statistical model may be validated based at least in part on the response data.

Techniques for generating machine learning training data

Methods, systems, and computer-readable media for item classification using customer-visible attributes are disclosed. A plurality of terms are determined that describe a plurality of items in a marketplace. Individual ones of the items are classified in a hierarchical taxonomy comprising a plurality of classifications, and individual ones of the terms correspond to individual ones of the classifications. A description of a new item is received. The description of the new item comprises a plurality of customer-visible terms. One or more of the plurality of classifications in the hierarchical taxonomy are selected for the new item. The one or more classifications are selected for the new item based at least in part on automated matching of individual ones of the customer-visible terms to individual ones of the terms that correspond to individual ones of the classifications.

Item classification using visible attributes

Disclosed are various embodiments for assessing the quality of a node that comprises a collection of items containing textual data. The homogeneity of the node can be related to its quality. Highly ranked descriptive terms used in the node are identified and quality score is calculated that provides a measure of the quality of the node. Additionally, a node can be examined for outliers to improve node quality.

Growing community of inventors

Bellevue, WA, United States of America

Archiman Dutta