Amazon Technologies, Inc.

Vijai Mohan

Srikanth Thirumalai

Grant Michael Emery

Aswath Manoharan

Mark J Tomko

Egidio Terra

Leo Parker Dirac

Oleg Rybakov

Rejith George Joseph

Roshan Harish Makhijani

Benjamin Thomas Cohen

Madhu Madhava Kurup

Lisa Jane Hinegardner

Matias Omar Gregorio Benitez

Vishwanathan Swaminathan

Xiaoxin Yin

Houssam Nassif

Andrew Dennis Hamel

Sriram Srinivasan

Mitchell Howard Goodman

Eiman Mohamed Hamdy Elnahrawy

Eric Nalisnick

A network-based enterprise or other system that makes items available for selection to users may implement selecting user interface elements for inclusion with a search result according to item category features of prior item selections. A search request for an item may be received. An item category for the item may be identified and a user interface element type selection model for the item category may be accessed to select of user interface element types for inclusion in a display of a search result in response to the search request. The user interface element type selection model for the item category may be generated based on features of previous item selections in the identified item category. Content for the selected user interface elements may be determined and a display of the search result may be provided that includes user interface elements generated according to the selected type and identified content may be included.

System, method and non-transitory computer-readable medium for selecting user interface element types for display with a search result according to item category features of prior item selections

The present disclosure is directed to reducing model size of a machine learning model with encoding. The input to a machine learning model may be encoded using a probabilistic data structure with a plurality of mapping functions into a lower dimensional space. Encoding the input to the machine learning model results in a compact machine learning model with a reduced model size. The compact machine learning model can output an encoded representation of a higher-dimensional space. Use of such a machine learning model can include decoding the output of the machine learning model into the higher dimensional space of the non-encoded input.

Encodings for reversible sparse dimensionality reduction

The present disclosure is directed to generating and using a machine learning model, such as a neural network, by augmenting another machine learning model with an additional parameter. The additional parameter may be connected to some or all nodes of an internal layer of the neural network. A machine learning model can determine a value associated with the additional parameter using non-behavior or non-event-based information. The machine learning model can be trained using non-behavior or non-event-based information and parameter values of the other machine learning model.

Generation and use of model parameters in cold-start scenarios

Techniques described herein are directed to improved artificial neural network machine learning techniques that may be employed with a recommendation system to provide predictions with improved accuracy. In some embodiments, item consumption events may be identified for a plurality of users. From these item consumption events, a set of inputs and a set of outputs may be generated according to a data split. In some embodiments, the set of outputs (and potentially the set of inputs) may include item consumption events that are weighted according to a time-decay function. Once a set of inputs and a set of outputs are identified, they may be used to train a prediction model using an artificial neural network. The prediction model may then be used to identify predictions for a specific user based on user-specific item consumption event data.

Recommendation system using improved neural network

Disclosed are various embodiments for a similarity service. Multiple samplings of user accounts are randomly selected from a pool of user accounts. Interaction history data for each of the user accounts is used to determine item similarities corresponding to each of the user account samplings. The item similarity data is aggregated to determine similar items.

Random sample aggregation similarities

Recommendations can be generated even in situations where sufficient user information is unavailable for providing personalized recommendations. Instead of generating recommendations for an item based on item type or category, a relation graph can be consulted that enables other items to be recommended that are related to the item in some way, which may be independent of the type or category of item. For example, images of models, celebrities, or everyday people wearing items of clothing, jewelry, handbags, shoes, and other such items can be received and analyzed to recognize those items and cause them to be linked in the relation graph. When generating recommendations or selecting advertisements, the relation graph can be consulted to recommend products that other people have obtained with the item from any of a number of sources, such that the recommendations may be more valuable to the user.

Object relation builder

Features are disclosed for identifying randomized latent feature language modeling, such as a recurrent neural network language modeling (RNNLM). Sequences of item identifiers may be provided as the language for training the language model where the item identifiers are the words of the language. To avoid localization bias, the sequences may be randomized prior to or during the training process to provide more accurate prediction models.

Randomized latent feature learning

The arrangement and selection of digital content to present to a user can be based upon criteria such as profitability or interest to a user. The selection can also be made to ensure that a diversity of item content is presented. The selection can utilize various rules or policies for diversity at the category level or item feature level, among other such options. In addition to selection diversity, the placement of item content displayed can satisfy various diversity criteria in order to ensure diversity of display as well.

Content selection algorithms

Features are disclosed for identifying and routing items for tagging using a latent feature model, such as a recurrent neural network language model (RNNLM). The model may be trained to identify latent features for catalog items such as movies, books, food items, beverages, and the like. Based on similarities in latent features, tags previous assigned to items may be applied to untagged items. Application may be manual or automatic. In either case, resources need to be balances to ensure efficient tagging of items. The included features help to identify and direct these limited tagging resources.

Latent feature based tag routing

According to aspects of the disclosed subject matter, a method for identifying a set of documents from a document corpus that are potential duplicates of a source document, is provided. A source document is obtained. A list of queries corresponding to the source document is identified. Each query in the identified list of queries is executed on the document corpus, wherein the execution of each query yields a corresponding results set identifying an ordered set of documents in the document corpus. For each document identified in each results set, a document score is generated for the identified document based on the identified document's ordinal position in its results set. A subset of the identified documents of the results set is selected according to the generated document scores that satisfy predetermined selection criteria. The selected subset of identified documents are stored or displayed.

Identifying potential duplicates of a document in a document corpus

A computer system and method for determining whether the subject matter described in a received document is substantially similar to the subject matter of other documents in a document corpus, such that the received document can be considered a duplicate document. After receiving a first document, a set of tokens for the first document is generated. A non-fielded relevance search on a token index is executed. The relevance search returns a set of candidate duplicate documents with scores corresponding to each candidate document. For each candidate document with a score above a threshold, filtering is performed on each candidate document to determine whether each candidate document is a true duplicate of the first document. A set of candidate documents with a score above the threshold that were not disqualified as candidate documents is then provided.

Duplicate entry detection system and method

Systems and methods for determining a set of variation-phrases from a collection of documents in a document corpus is presented. Potential variation-phrase pairs among the various documents in the document corpus are identified. The identified potential variation-phrase pairs are then added to a variation-phrase set. The potential variation-phrase pairs in the variation-phrase set are filtered to remove those potential variation-phrase pairs that do not satisfy a predetermined criteria. After filtering the variation-phrase set, the resulting variation-phrase set is stored in a data store.

Determining variation sets among product descriptions

Systems and methods for filtering tokens from a document for determining whether the document describes substantially similar subject matter compared to another document are described. In one embodiment, a first document is obtained. This document is organized into a plurality of fields, and at least some of the fields include tokens representing the subject matter described by the document. A field of this document is selected and a token from within the selected field having the highest inverse document frequency (IDF) is selected. Those tokens that have a higher IDF than the selected token are removed. Using the remaining tokens, a determination is made as to whether the first document describes substantially similar subject matter to the subject matter described by a second document. An indication is provided as to whether the first document describes substantially similar subject matter to that described by a second document according to the determination.

Filtering invalid tokens from a document using high IDF token filtering

Systems and methods for determining whether a first document is a potential duplicate of a second document such that the two documents describe the same or substantially the same subject matter, wherein the first and second documents include attribute data in attribute fields. A set of rules is obtained for determining whether the first document is a potential duplicate of the second document. Moreover, for each rule in the set of rules, a determination is made as to whether data in a first set of attributes of the first document is contained in a second set of attributes of the second document. According to the results of the evaluated rules in the rules set, determining whether the first document is a potential duplicate of the second document. If, according to the evaluated rules in the rules set, the first document is determined to be a potential duplicate of the second document, storing a reference to the first document in a set of potential duplicates of the second document.

Comparison engine for identifying documents describing similar subject matter

According to aspects of the disclosed subject matter, a method for identifying a set of documents from a document corpus that are potential duplicates of a source document is provided. A source document is obtained. A list of queries corresponding to a source document is identified. Each query in the identified list of queries is executed on the document corpus, wherein the execution of each query yields a corresponding results set identifying an ordered set of documents in the document corpus. For each document identified in each results set, a document score is generated for the identified document based on the identified document's ordinal position in its results set. A subset of the identified documents of the results set is selected according to the generated document scores that satisfy predetermined selection criteria. The selected subset of identified documents are stored or displayed.

Growing community of inventors

Bellevue, WA, United States of America

Vijai Mohan