Abstract: Search queries are received and search results are provided. Interaction tracking is used to determine with which search results users interact. The search results having received interactions can be represented as item vectors, which can include a vector representation of a portion of the search result, such as a title, description, or image. For each search query, the item vectors are aggregated, such as by averaging the item vectors. The search queries are stored in an item dataset as collected search queries respectively associated with the aggregate item vectors. When a new search query is received, a search query vector can be compared to the aggregate item description vectors to identify collected search queries that are related. The related collected search queries can be provided as search query recommendations or search results associated with the collected search queries can be provided in response to receiving the new search query.

Abstract: The technology disclosed herein relates to identifying an aspect from a search query based on using a multipartite graph generated using past user behavior and a node embedding algorithm for determining vector representations of nodes of the multipartite graph. For example, nodes of the multipartite graph can include nodes for prior search queries, items or item listings associated with the prior search queries, and one or more of an aspect or category of the items or item listings. In embodiments, the multipartite graph has dynamic edges between the nodes for the prior search queries and the items or item listings. In embodiments, a query expansion is performed based on identifying the aspect using the multipartite graph and node embedding algorithm. In embodiments, search results are provided based on identifying the aspect and performing the query expansion. For example, one or more identified aspects can be provided as selectable options.

Abstract: A search system performs item retrieval using search query categorization that matches query intent. Category embeddings are generated for categories based on hierarchical data and search information. For instance, the category embeddings can be generated using information regarding hierarchical relationships between the categories, co-occurring relationships between categories identified from search information, and initial embeddings that encode query-related information for each category. Category clusters can be formed using the category embeddings. When a search query is received, one or more categories are identified from a category cluster and used for selecting search results for the search query.

Abstract: Search queries are received and search results are provided. Interaction tracking is used to determine with which search results users interact. The search results having received interactions can be represented as item vectors, which can include a vector representation of a portion of the search result, such as a title, description, or image. For each search query, the item vectors are aggregated, such as by averaging the item vectors. The search queries are stored in an item dataset as collected search queries respectively associated with the aggregate item vectors. When a new search query is received, a search query vector can be compared to the aggregate item description vectors to identify collected search queries that are related. The related collected search queries can be provided as search query recommendations or search results associated with the collected search queries can be provided in response to receiving the new search query.