Abstract: Devices and techniques are generally described for search using a micro-partitioned catalog. In some examples, a search system can determine a plurality of items in an online catalog. The search system can generate a similarity graph based on vector representations of the plurality of items and determine an item as a representative item for a cluster of items. The search system can receive a query and determine the cluster of items as a candidate cluster based on a comparison between the query and the representative item. The search system can determine a candidate item of the cluster of items based on a comparison between the query and the cluster of items.

Abstract: Methods and apparatus for identifying features that may have a high potential impact on key application metrics. These methods rely on observational data to estimate the importance of application features, and use causal inference tools such as Double Machine Learning (double ML) or Recurrent Neural Networks (RNN) to estimate the impacts of treatment features on key metrics. These methods may allow developers to estimate the effectiveness of features without running online experiments. These methods may, for example, be used to effectively plan and prioritize online experiments. Results of the online experiments may be used to optimize key metrics of mobile applications, web applications, websites, and other web-based programs.

Abstract: Systems and methods are provided for optimizing campaigns (such as marketing campaigns) based on both short term and long term behaviors of users. A computing system learns an incremental outcome prediction model using training data comprising a marketing campaign log entry of a subject user and another marketing campaign log entry of a corresponding user, which represents a counterfactual outcome for the subject user. A marketing campaign can be selected for another user using the learned model.

Abstract: User-generated tags from viewing web-based content are collected over a predetermined period of time. A subset of distinct or unique tags is identified from among the collected tags. A z-score is calculated for each identified distinct tag, where the z-score is a measure of the statistical significance of the tag. The subset of distinct tags is then sorted based on their corresponding z-score. All distinct tags having a corresponding z-score lower than a predetermined threshold are rejected and the remaining distinct tags, having a corresponding z-score higher than the threshold are used to infer a user's interest. The ability to infer a user's interests from the remaining distinct tags may thus benefit web-based applications by achieving a high degree of accuracy in predicting the interests of users by leveraging on the use of the user generated content tags and keywords.

Abstract: A server determines a plurality of immediate candidate items for a first web page to recommend to a user. For each particular immediate candidate item of the plurality of immediate candidate items, the server determines a separate sequence of two or more subsequent possible candidate items for subsequent web pages to recommend to the user in the event that the user selects the particular immediate candidate item. Further, the server selects a particular immediate candidate item from the plurality of immediate candidate items for the first web page to recommend to the user. The first web page that recommends the plurality of immediate candidate items is generated and sent over the Internet to the user.

Abstract: In one embodiment, access one or more query chains, wherein each one of the query chains comprises two or more search queries, {q1, . . . , qn}, which are recency-sensitive, are related to the same subject matter, and are issued to a search engine sequentially, and actual click-through information associated with each one of the query chains; and smooth each one of the query chains using the actual click-through information associated with the query chain. To smooth one of the query chains comprises, for each one of search queries, qj, in the query chain, where 2?j?n, if one of the network resources identified for qj has actually been clicked in connection with qj by the corresponding one network user, then presume that the one network resource has been clicked in connection with one or more search queries, qk, in the query chain, where 1?k<j.

Abstract: In one embodiment, access one or more query-resource pairs, wherein for each one of the query-resource pairs comprising one of one or more search queries and one of one or more network resources, the one search query is recency-sensitive with respect to a particular time period, and the one network resource is identified for the one search query, and a resource-view count and a resource-click count associated with each one of the query-resource pairs; and construct one or more first click features using the resource-view counts and the resource-click counts associated with the query-resource pairs. To construct one of the first click features in connection with one of the query-resource pairs comprises determine a only-resource-click count associated with the one query-resource pair; and calculate a ratio between the only-resource-click count and the resource-view count associated with the one query-resource pair as the one first click feature.

Abstract: A server determines a plurality of immediate candidate items for a first web page to recommend to a user. For each particular immediate candidate item of the plurality of immediate candidate items, the server determines a separate sequence of two or more subsequent possible candidate items for subsequent web pages to recommend to the user in the event that the user selects the particular immediate candidate item. Further, the server selects a particular immediate candidate item from the plurality of immediate candidate items for the first web page to recommend to the user. The first web page that recommends the plurality of immediate candidate items is generated and sent over the Internet to the user.

Abstract: In one embodiment, access one or more query chains, wherein each one of the query chains comprises two or more search queries, {q1, . . . , qn}, which are recency-sensitive, are related to the same subject matter, and are issued to a search engine sequentially, and actual click-through information associated with each one of the query chains; and smooth each one of the query chains using the actual click-through information associated with the query chain. To smooth one of the query chains comprises, for each one of search queries, qj, in the query chain, where 2?j?n, if one of the network resources identified for qj has actually been clicked in connection with qj by the corresponding one network user, then presume that the one network resource has been clicked in connection with one or more search queries, qk, in the query chain, where 1?k<j.

Abstract: In one embodiment, access one or more query-resource pairs, wherein for each one of the query-resource pairs comprising one of one or more search queries and one of one or more network resources, the one search query is recency-sensitive with respect to a particular time period, and the one network resource is identified for the one search query, and a resource-view count and a resource-click count associated with each one of the query-resource pairs; and construct one or more first click features using the resource-view counts and the resource-click counts associated with the query-resource pairs. To construct one of the first click features in connection with one of the query-resource pairs comprises determine a only-resource-click count associated with the one query-resource pair; and calculate a ratio between the only-resource-click count and the resource-view count associated with the one query-resource pair as the one first click feature.

Abstract: The present invention provides for the detection of abnormal user behavior for a query session of an electronic search engine. A query session is initiated upon receipt of a user search request that includes one or more search terms. The search engine, in accordance with known search technology, generates a search results page that includes various hyperlinks, including for example web content hyperlinks, page navigation hyperlinks and advertising hyperlinks. Tracking user activities generates the clickstream associated with the search results page. The present invention determines a probability score for the clickstream and then this score is normalized. A comparison of the normalized probability score with other normalized probability scores for similar query sessions determines of the normalcy of the query session.

Abstract: The present invention provides for the detection of abnormal user behavior for a query session of an electronic search engine. A query session is initiated upon receipt of a user search request that includes one or more search terms. The search engine, in accordance with known search technology, generates a search results page that includes various hyperlinks, including for example web content hyperlinks, page navigation hyperlinks and advertising hyperlinks. Tracking user activities generates the clickstream associated with the search results page. The present invention determines a probability score for the clickstream and then this score is normalized. A comparison of the normalized probability score with other normalized probability scores for similar query sessions determines of the normalcy of the query session.

Abstract: There are provided methods and systems for inferring a user's interests from user-generated tags of web-based content. In accordance with the invention, user-generated tags from viewing web-based content are collected over a predetermined period of time. A subset of distinct or unique tags is identified from among the collected tags. A z-score is calculated for each identified distinct tag, where the z-score is a measure of the statistical significance of the tag. The subset of distinct tags is then sorted based on their corresponding z-score. All distinct tags having a corresponding z-score lower than a predetermined threshold are rejected and the remaining distinct tags, having a corresponding z-score higher than the threshold are used to infer a user's interest.