Abstract: Described herein is a technique to facilitate filtering during candidate retrieval stage performed by an information retrieval system that utilizes embedding models. An aNN indexing structure is created for each end-user, and in some instances, each activity type. This allows a single request for candidate content items to invoke a single process to obtain content items that satisfy the filtering criteria (e.g., in this case, the in-network requirement) from the separate per-end-user indexes.

Abstract: Methods and systems for improving memory utilization and access speed of database indexes are disclosed. A representation of attributes of items listed on a network-based publication system is identified. The representation includes a plurality of lists of the attributes as a sequence of serialized attributes, each of which stores an attribute identifier (ID) and data corresponding to the attribute identifier. A compact or compressed structure is used to represent the attribute IDs. This representation allows a reader application to quickly determine whether an attribute ID is present in the lists or not. Directly-addressable variable-length codes (DACs) are used, in an unconventional manner, to represent the data. By representing the data in this way, the index takes up less memory for certain types of data and the reader application is able to gain faster access to the index.

Abstract: The disclosed technologies include receiving a selection of an item for which an attribute is to be determined. The attribute is dependent on one or more variable conditions. Data is input that represents a set of rules for determining the attribute. Based on the rules, a parse tree is generated that is usable to determine the attribute based on the one or more variable conditions. The data is parsed and nodes of the parse tree are pre-calculated as values become available. The attribute is determined based on the parsing and pre-calculating. The determined attribute is returned to a requesting process.

Abstract: Methods and systems for improving memory utilization and access speed of database indexes are disclosed. A representation of attributes of items listed on a network-based publication system is identified. The representation includes a plurality of lists of the attributes as a sequence of serialized attributes, each of which stores an attribute identifier (ID) and data corresponding to the attribute identifier. A compact or compressed structure is used to represent the attribute IDs. This representation allows a reader application to quickly determine whether an attribute ID is present in the lists or not. Directly-addressable variable-length codes (DACs) are used, in an unconventional manner, to represent the data. By representing the data in this way, the index takes up less memory for certain types of data and the reader application is able to gain faster access to the index.

Abstract: A set S is initialized. Initially, S is empty; but, as the disclosed process is performed, items are added to it. It may contain one or more samples (e.g., items) from each class. One or more labeled samples for one or more classes may be obtained. A series of operations may be performed, iteratively, until a stopping criterion is reach to obtain the reduced set. For each class of the one or more classes, a point may be generated based on at least one sample in the class having a nearest neighbor in a set S with a different class label than the sample. The point may be added to the set S. The process may be repeated unless a stopping criterion is reached. A nearest neighbor for a submitted point in the set S may be identified and a candidate nearest neighbor may be output for the submitted point.

Abstract: A set S is initialized. Initially, S is empty; but, as the disclosed process is performed, items are added to it. It may contain one or more samples (e.g., items) from each class. One or more labeled samples for one or more classes may be obtained. A series of operations may be performed, iteratively, until a stopping criterion is reach to obtain the reduced set. For each class of the one or more classes, a point may be generated based on at least one sample in the class having a nearest neighbor in a set S with a different class label than the sample. The point may be added to the set S. The process may be repeated unless a stopping criterion is reached. A nearest neighbor for a submitted point in the set S may be identified and a candidate nearest neighbor may be output for the submitted point.