Search result identification using vector aggregation

1. A computerized method for returning search results performed by one or more processors, the method comprising:

mapping collected search queries to a subset of search results determined based on interaction tracking of the collected search queries, wherein search results comprise item descriptions;

determining item vectors for the item descriptions of the subset of search results;

aggregating the item vectors of the item descriptions of the subset of search results to determine an aggregate item description vector for each collected search query, wherein the aggregating averages the item vectors using a weighted averaged determined from the interaction tracking;

matching a search query to a collected search query associated with an aggregate item vector; and

executing a search for the search query using the aggregate item vector of the matching collected search query to identify a set of search results for the search query.

2. The method of claim 1 , further comprising:

populating an item dataset with the collected search queries and the aggregate item description vector for each of the collected search queries.

3. The method of claim 1 , further comprising:

identifying a plurality of related collected search queries, the plurality of related collected search queries identified based on a cosine similarity between aggregate item description vectors for the plurality of related collected search queries and the aggregate item description vector for the matching collected search query; and

providing the plurality of related collected search queries as search query recommendations in response to receiving the search query.

4. The method of claim 1 , further comprising:

training a neural network on training data comprising the collected search queries and the aggregate item description vector for each of the collected search queries.

5. The method of claim 1 , wherein the aggregate item description vectors are determined from a selected number of top ranked search results mapped to the collected search queries.

6. The method of claim 1 , further comprising determining a proportional number of interactions received by each search result of the subset of search results, wherein the weighted average of the item vectors is determined according to the proportional number of interactions.

7. One or more computer storage media storing computer-readable instructions that when executed by a processor, cause the processor to perform operations for returning search results, the operations comprising:

receiving a search query;

determining a search query vector;

identifying a related collected search query from collected search queries within an item dataset, the item dataset comprising the collected search queries associated with aggregate item description vectors for items listed on an item database, the aggregate item description vectors determined using a weighted average of item vectors for search results of the collected search queries, the weighted average based on interaction tracking of the search results, wherein the related collected search query is identified based on the search query vector and an aggregate item description vector for the related collected search query; and

executing a search using the related collected search query to identify search results for the search query.

8. The media of claim 7 , further comprising performing a clustering analysis on the item dataset to determine aggregate item description vector clusters, wherein the related collected search query is identified based on the search query vector and the aggregate item description vector being included within a same aggregate item description vector cluster.

9. The media of claim 7 , wherein the item dataset is a nearest neighbor database, and the related collected search query is identified based on the aggregate item description vector for the related collected search query being a nearest neighbor relative to the search query vector.

10. The media of claim 7 , wherein the item dataset is a nearest neighbor database, and the related collected search query is identified based on the aggregate item description vector for the related collected search query being within a distance threshold value, the distance threshold value determined based on a variance for an aggregate item description vector of a collected search query that matches the search query.

11. The media of claim 7 , further comprising employing a trained classifier to predict a classification for the search query, wherein the related collected search query is further identified based on the predicted classification.

12. A system returning search results, the system comprising:

at least one processor; and

one or more computer storage media storing computer-readable instructions thereon that when executed by the at least one processor, cause the at least one processor to:

generate an item dataset comprising collected search queries, classifications of the collected search queries, and aggregate item description vectors, the aggregate item description vectors being determined by aggregating item vectors of item descriptions associated with the collected search queries, the item vectors aggregated using a weighted average, wherein weights associated with the weighted average are determined from interaction tracking of the search results for the collected search queries corresponding to the item vectors being aggregated;

train a neural network using the item dataset as training data, wherein based on the training, the neural network is configured to receive a search query and predicted classification for the search query and to predict a vector representation of the search query; and

store the trained neural network for use in predicting vector representations for search queries.

13. The system of claim 12 , wherein the aggregate item description vectors are determined by averaging the item vectors for items provided as search results for the collected search queries.

14. The system of claim 12 , further comprising employing a trained classifier to predict the classifications for each of the collected search queries.

15. The system of claim 12 , further comprising:

employing the neural network to predict a search query vector for a received search query;

identify a related collected search query from the collected search queries of the item dataset based on the search query vector and an aggregate item description vector for the related collected search query; and

return search results associated with the related collected search query in response to receiving the search query.