Abstract: Embodiments of the present invention provide improved techniques for determining long term relevance and user behavior using query chains. The query chains may first be detected and then annotated into different types of chains based at least in part on various decision rules, machine-learned classifiers, and inter-query relationships. The query chains may then be subsequently used to train models for predicting user behavior and providing more relevant results to a user's queries. A content provider system according to various embodiments may aggregate historical data associated with previous search and/or transaction data, which may be analyzed to detect query chains, for example, whether queries are chained to each other. Determining whether queries are chained to each other may involve incorporating decision rules and reformulation models, analyzing temporal windows between queries, and/or analyzing inter-query relationships.

Abstract: The arrangement and selection of digital content to present to a user can be based at least in part upon probabilities of the user selecting to view more information and/or entering into a transaction with respect to instances of the content. For example, user behavior with respect to various items provided through a content provider can be determined in order to calculate a probability that a user was searching for a particular type of item for a given search. The user behavior can include historical action data, such as information that indicates items that were selected in response to a particular search and/or items that were purchased in response to a particular search. The historical action data can be analyzed to generate an index that indicates a likelihood that the search was intended for a particular type of product. Additionally, the historical data can be used to train language models that can be used to determine a probability of interest for a particular type of product for a given search.

Abstract: Approaches are described for ranking multiple products or other items, such as products obtained in response to a search request submitted to a server. The ranking system determines a ranking score for the products based on both data available online and item data that must be computed offline due to longer computation time or unavailability of data. The ranking score can be used to rank the products and determine which products are the most relevant to the user. A hybrid boosting method is used to first train an online ranking function to produce an online ranking score for the item. In the second phase, an offline ranking function is trained to produce a second ranking score for the item. The online rank score is combined with the offline rank score at query time to produce a combined rank for the items in the search results.

Abstract: Various approaches provide for determining the selection and/or ranking of items to display based at least in part upon the probabilities of a user being interested in those items. The probabilities can be based on profile information of a user. The profile information can include a three-dimensional virtual model of the user. When a search for content is received, the information for the three-dimensional virtual model can be used to determine a set of items. For each item, a matching score quantifying a visual appreciation can be determined based on how well an item virtually “fits” the three-dimensional virtual model. Based on the matching score, the selection and/or ranking of items to display can be determined. In some situations, the items can be shown to appear to be worn by the three-dimensional virtual model. The user can purchase items, cause item information to be presented to other users or sent to other devices, and/or perform other such actions.

Abstract: The relevance or ranking of various dynamically-determined instances of content can be adjusted or otherwise improved based on aspects such as the freshness or seasonality of the content. In many conventional relevance or ranking algorithms, factors such as popularity, performance, and profit are used to determine which content to present to a user, as well as how to display the selected content. Such algorithms do not necessarily reflect domain-specific knowledge very well, and thus fail to accurately select content that is of differing levels appeal at various ages of the content and/or during various seasons of the year. Freshness and/or seasonality adjustment scores can be used to generate improved relevance, selection, or ranking determinations for various categories of content.

Abstract: Video content can be analyzed to identify particular sections of the video content. Speech to text or similar techniques can be used to obtain a transcription of the video content. The transcription can be indexed (e.g., timestamped) to the video content. Information describing how users are interacting with or consuming the video content (e.g., social media information, viewing history data, etc.) can be collected and used to identify the particular sections. Once the particular sections have been identified, other services can be provided. For example, custom trailers and summaries of the video content can be generated based on the identified sections. Additionally, the video content can be augmented to include additional information relevant to the particular sections, such as production information, actor information, or other information. The additional information can be added so as not to interfere with the important sections.

Abstract: Embodiments described herein provide images representing a set of search results based on diversity between results of the search query. Images associated with a set of visually diverse items can be provided to provide a sample of items matching the search query across multiple types of categories. For example, search results can be grouped into types of categories and images from each of the types of categories can be grouped into subsets of visually related images (across one or more different visual attributes). A set of diverse representative images can be selected by taking at least one image from each of the groups of visually related images. The set of representative and diverse images can be displayed to provide an interesting, visually diverse, and aesthetically pleasing set of images to a user.

Abstract: Video content can be analyzed to identify particular sections of the video content. Speech to text or similar techniques can be used to obtain a transcription of the video content. The transcription can be indexed (e.g., timestamped) to the video content. Information describing how users are interacting with or consuming the video content (e.g., social media information, viewing history data, etc.) can be collected and used to identify the particular sections. Once the particular sections have been identified, other services can be provided. For example, custom trailers and summaries of the video content can be generated based on the identified sections. Additionally, the video content can be augmented to include additional information relevant to the particular sections, such as production information, actor information, or other information. The additional information can be added so as not to interfere with the important sections.

Abstract: Video content can be analyzed to identify particular sections of the video content. Speech to text or similar techniques can be used to obtain a transcription of the video content. The transcription can be indexed (e.g., timestamped) to the video content. Information describing how users are interacting with or consuming the video content (e.g., social media information, viewing history data, etc.) can be collected and used to identify the particular sections. Once the particular sections have been identified, other services can be provided. For example, custom trailers and summaries of the video content can be generated based on the identified sections. Additionally, the video content can be augmented to include additional information relevant to the particular sections, such as production information, actor information, or other information. The additional information can be added so as not to interfere with the important sections.

Abstract: Approaches are described for ranking multiple products or other items, such as products obtained in response to a search request submitted to a server. The ranking system determines a ranking score for the products based on both data available online and item data that must be computed offline due to longer computation time or unavailability of data. The ranking score can be used to rank the products and determine which products are the most relevant to the user. A hybrid boosting method is used to first train an online ranking function to produce an online ranking score for the item. In the second phase, an offline ranking function is trained to produce a second ranking score for the item. The online rank score is combined with the offline rank score at query time to produce a combined rank for the items in the search results.

Abstract: Approaches are described for ranking multiple products or other items, such as products obtained in response to a search request submitted to a server. The ranking system determines a ranking score for the products based on both data available online and item data that must be computed offline due to longer computation time or unavailability of data. The ranking score can be used to rank the products and determine which products are the most relevant to the user. A hybrid boosting method is used to first train an online ranking function to produce an online ranking score for the item. In the second phase, an offline ranking function is trained to produce a second ranking score for the item. The online rank score is combined with the offline rank score at query time to produce a combined rank for the items in the search results.