Amazon Technologies, Inc.

Giovanni Zappella

Cedric Philippe Archambeau

Cédric Philippe Charles Jean Ghislain Archambeau

Rodolphe Jenatton

Edward Thomas Banti

David Spike Palfrey

Samuel Theodore Sandler

Michael Brueckner

Matthias Seeger

Valerio Perrone

Lukas Stefan Balles

Iaroslav Shcherbatyi

Dmitrij Scsadej

David Salinas

Vito Bellini

Jurgen Ommen

Borys Marchenko

Marcel Ackermann

Martin Milicic

Giuseppe Di Benedetto

Techniques for concept shift detection and correction using probabilistic models and learned feature representations are described. A gaussian process model is trained using representations generated by a primary machine learning (ML) model for existing training data elements in a training memory. For a new batch of data elements, representations again generated by the primary ML model can be used as input for the gaussian process model to generate predictive distributions. When the true targets for the new data elements are not sufficiently likely according to the corresponding predictive distributions, concept shift is likely and the training memory can be purged of the existing data elements before further retraining of the primary ML model.

Concept shift detection and correction using probabilistic models and learned feature representations

A particular hyper-parameter combination (HPC) that was recommended for a first task is included in a collection of candidate HPCs evaluated for a second task. Hyper-parameter analysis iterations are conducted for the second task using the collection. In one of the iterations, the second task is executed using a first iteration-specific set of HPCs, including the particular HPC and one or more other members of the collection. One or more of the HPCs of the first iteration-specific set of HPCs are pruned to generate a second iteration-specific set of HPCs for a subsequent iteration. HPCs are selected for pruning based on a comparison of their results with the results obtained from the particular HPC that was recommended for the first task. A recommended HPC for the second task is identified based on results of the analysis iterations.

Resource-efficient techniques for repeated hyper-parameter optimization

Hyperparameters for tuning a machine learning system may be optimized using Bayesian optimization with constraints. The hyperparameter optimization may be performed for a received training set and received constraints. Respective probabilistic models for the machine learning system and constraint functions may be initialized, then hyperparameter optimization may include iteratively identifying respective values for hyperparameters using analysis of the respective models performed using an acquisition function implementing entropy search on the respective models, training the machine learning system using the identified values to determine measures of accuracy and constraint metrics, and updating the respective models using the determined measures.

Hyperparameter optimization with operational constraints

A computer system and process extract information from a literary work regarding relationships between entities (e.g., characters, locations, etc.) described or represented in the literary work, and generate a graph representing these relationships. The graph data is parsed into sub-graphs, and the subgraphs are used to generate a signature of the literary work. The respective signatures of different literary works may be compared for purposes of generating literary work recommendations for users.

Generation and use of literary work signatures reflective of entity relationships

Techniques are described herein for generating adaptive recommendations in response to a content request. The system herein detects abrupt changes and leverages the seasonality of a reward function. A collection of contextual models are utilized, each one learning about one of the unique reward stationary states. A short-term memory model is used to detect reward shifts toward stationary periods that have not occurred in the past. In this case, a new base bandit instance is initialized. In order to perform the change point detection, at each step every model gets assigned a score indicating how likely the last observation is to come from a corresponding stationary period represented by a respective model. A model is selected based on the scores. The model provides a recommendation and the system can monitor clickstream data to identify the reward for providing the recommendation.

Techniques for content selection in seasonal environments

Computer systems and associated methods are disclosed to implement a model executor that dynamically selects machine learning models for choosing sequential actions. In embodiments, the model executor executes and updates an active model to choose sequential actions. The model executor periodically initiates a recent model and updates the recent model along with the active model based on recently chosen actions and results of the active model. The model executor periodically compares respective confidence sets of the two models' parameters. If the two confidence sets are sufficiently divergent, a replacement model is selected to replace the active model. In embodiments, the replacement model may be selected from a library of past models based on their similarity with the recent model. In embodiments, past models that exceed a certain age or have not been recently used as the active model are removed from the library.

Artificial intelligence system incorporating automatic model switching based on model parameter confidence sets

A testing environment in which offline simulations can be run to identify policies and/or prediction models that result in more valuable content being included in content pages is described herein. For example, the offline simulations can be run in an application executed by an experiment device using data gathered by a production content delivery system. The simulation application can test any number of different policies and/or prediction models without impacting users that use a production content delivery system to request content.

Offline simulation system for optimizing content pages

A set of informational content elements pertaining to an item for presentation to one or more potential item consumers is identified at an artificial intelligence service. A plurality of optimization iterations are implemented. In a particular iteration, a set of content elements to be presented to a target audience in accordance with a set of presentation constraints indicated by a content source associated with the item is identified using a machine learning model, and metrics indicating the effectiveness of the content elements are analyzed.

Artificial intelligence system for automated selection and presentation of informational content

A system ranks and/or recommends literary works based on information extracted from the text of the literary works. For example, the system may use information extracted from the text of a literary work to generate a graph representing the relationships of entities in the literary work. The system may identify sub-graphs in the graph, and generate a signature based on the values associated with the various sub-graphs. The system may generate signatures of a plurality of literary works. The system may then retrieve the signature of a literary work that was highly rated by a user, and compare the retrieved signature with other generated signatures using machine-learning algorithms to select literary works to recommend to the user.

Machine learning-based literary work ranking and recommendation system

Technologies for optimized selection of content for delivery to a user that both optimizes the expected return from the delivery of the content to the user and that enables exploration of delivery of new content to users are disclosed. Content is selected for delivery to a user based on an exploitation score that defines an estimate of the feedback expected from the delivery of the content to the user and an exploration score that varies inversely with the number of times that the content has been transmitted to all users. The use of the exploration score enables the exploration of delivery of new content to users. The content might be delivered via e-mail messages, a web site, or using another mechanism.

Optimized selection and delivery of content

Data mining systems and methods are disclosed for evaluating pairwise substitutability relationships among items. For example, a pairwise similarity measure may correspond to a value quantifying the extent to which an item A is favored over an item B by a population of users. Given a base item selected by a user, the system may select a candidate item from a set of potential substitute items for the base item based on current estimates of corresponding pairwise similarities. The system may then present the candidate item to the user in a context of comparison against the base item and obtain an indication of user preference between the two. The system may then update corresponding pairwise similarities based on the indication of preference.

Data mining system for assessing pairwise item similarity

A system is disclosed that identifies early adopter users by creating a directed graph of item access information for an item category and performing a page rank type process on the item access information. This directed graph may be created in a reverse temporal order. The early adopter users can be identified as the users with nodes in the directed graph that have a threshold number or rate of incoming links directly or indirectly pointing towards the nodes. Using the early adopter users as a sample, systems herein can determine whether to recommend an item based on the popularity of the item with respect to the early adopter users. Further, systems herein can determine an inventory level to maintain for an item based on the popularity of the item with respect to the early adopter users.

Growing community of inventors

Berlin, Germany

Giovanni Zappella