Amazon Technologies, Inc.

Gaurav Chanda

Brent Russell Smith

Ronald M Whitman

Sanjeev Jain

Jeffrey T Brownell

Soo-Min Pantel

Homer G Morgan, Iii

Roshan Harish Makhijani

Xuan Guo

Srinivas Vasu Chappidi

Saurabh Wadwekar

Techniques are provided for determining that an item should be provided for presentation to a user based on an item concept relatedness score. For example, a computer system may partition a plurality of items into a plurality of item concepts, whereby each item is assigned to one and only one item concept. The computer system may receive a list of co-selected items, whereby each item of the list is assigned to an item concept. Based at least in part on this list, the system may determine an item concept relatedness score between a first item concept and a second item concept of the plurality of item concepts. The computer system may then receive a selection of a first item assigned to a first item concept. Based at least in part on the item concept relatedness score, the system may provide the second item for presentation to a user.

Item recommendations through conceptual relatedness

A recommendation service that utilizes a machine learning algorithm trained with association vectors is provided. To allow for filtering of unrelated items, a recommendation service generates vectors that represent a quantization of items within a browse node that are considered complementary or a substitute of a selected item. Using an association vector, a machine learning algorithm can be trained to determine whether a particular item recommendation is considered noise, complementary or a substitute. Thereafter, the recommendation service can utilize the trained machine learning algorithm to filter a set of recommendations to remove items considered to be noise or to prioritize items identified as complementary or a substitute.

Recommendations utilizing noise detection and filtering

The present disclosure is directed to filtering co-occurrence data. In one embodiment, a machine learning model can be trained. An output of an intermediate structure of the machine learning model (e.g., an output of an internal layer of a neural network) can be used as a representation of an event. Similarities between representations of events can be determined and used to generate, augment, or modify co-occurrence data.

Prediction filtering using intermediate model representations

Features are provided for the analysis of collections of data and automatic grouping of data having certain similarities. A collection of data regarding user interactions with item-specific content can be analyzed. The analysis can be used to identify groups of items that are of interest to groups of similar users and/or to identify groups of users with demonstrated interests in groups of similar items. Data may be analyzed in a 'bottom-up' manner in which correlations within the data are discovered in an iterative manner, or in a 'top-down' manner in which desired top-level groups are specified at the beginning of the process. A bottom-up process may also be distributed among multiple devices or processors to more efficiently discover groups when using large collections of data.

Grouping of item data using seed expansion

Techniques for determining multiple item comparisons may be provided. For example, a system may monitor user interaction of a plurality of users that includes viewing and ordering items. The system may determine one or more items that compete, such that ordering a first item in the competing category of items lowers a probability that the user will also order a second item. The system may determine a subset of the competing items and providing information about the comparison and/or items for presentation to a user.

Data mining for multiple item comparisons

Disclosed are various embodiments for classifying search queries. A computing device identifies a user account associated with a submission of a search query to an electronic commerce application. The computing device then identifies a network page provided by the electronic commerce application, wherein the network page is requested with the user account. Subsequently, the computing device classifies the search query based at least in part on the requested network page.

Search query classification

Technologies are described herein for relating individual items to categories of items purchased concurrently by a customer. A representative category may be assigned to each of the items offered for sale by an online merchant. Purchase history data containing information regarding each item purchased by a customer on a specific date is augmented with information regarding each representative category of items from which an item was purchased by the customer on the same date. From the purchase history data, an attach rate is calculated between each particular item and a number of representative categories indicating the frequency of concurrency of purchase of the particular item with items assigned to the representative category. A user interface may then be generated containing the attach rates between the particular item and the representative categories and sent to a customer computer for display to the customer.

Relating individual items to similar purchased categories

A system is provided that extrapolates item characteristics from items considered to possess a characteristic to items not known to possess the characteristic. The system may include an item data repository that stores data representing physical items. These items can include first items having a characteristic and second items not known to have the characteristic. A characteristic extrapolation module can extrapolate the characteristic from at least some of the first items to at least some of the second items based at least in part on the strength of associations between the plurality of items. A recommendations module may provide item recommendations based at least partly on the characteristic of the items.

System for extrapolating item characteristics

A process is disclosed for using item price data to assess whether particular item-to-item associations are useful for generating item recommendations. In one embodiment, the process involves generating a compatibility score that represents a degree to which the two items of an item-to-item mapping are compatible in price. This score may then be used to determine whether to use the item-to-item association as a basis for generating item recommendations.

Identifying low utility item-to-item association mappings

Automated processes are disclosed for identifying bundles of three or more items to suggest to users. In one embodiment, item action data, such as item purchase data reflective of the item purchases of a user population, is used to detect pairwise (item to item) associations between particular items represented in an electronic catalog. These pairwise associations are then used (optionally in combination with other information) to select bundles of three or more items to recommend to users. The bundles may, for example, be suggested on item detail pages of the electronic catalog.

Automated selection of three of more items to recommend as a bundle

A computer system for detecting associations between items may include a probabilistic analysis component that accesses user data having information about item selections by a plurality of users. The probabilistic analysis component may programmatically generate associations between certain items by determining a first number of users who selected the items and by estimating a probability that a second number of users would have selected the items due to random chance. The probabilistic analysis component may estimate this probability based at least partly on a number of item selections made by one or more of the users.

System for detecting probabilistic associations between items

A computer-implemented process for reducing category overlap obtains a set of source items and an initial set of item categories associated with the source items. An assessment of the overlap between different item categories in the initial set of item categories is used to selectively eliminate one or more item categories prior to presentation of the source items and associated categories to the user, thereby resulting in reduced overlap between the item categories presented. The resulting set of categories improves the ability of the user to efficiently and effectively access source items of interest, such as recommended items generated by a recommendation engine.

Growing community of inventors

Seattle, WA, United States of America

Gaurav Chanda