Item retrieval using fitment data-based part clustering

1 . One or more computer storage media storing computer-useable instructions that, when used by one or more computing devices, cause the one or more computing devices to perform operations, the operations comprising:

accessing item data for each of a plurality of part item listings, the item data for each part item listing comprising fitment data;

for each part item listing, embedding the item data including the fitment data to generate an item-level embedding to provide a plurality of item-level embeddings for the plurality of part item listings, wherein each item-level embedding represents a different part item listing from the plurality of part item listings based on at least the fitment data for each part item listing;

clustering the plurality of item-level embeddings for the plurality of part item listings to generate a plurality of clusters, each cluster comprising one or more a subset of the plurality of item-level embeddings for a corresponding subset of different part item listings from the plurality of part item listings; and

storing cluster data for the plurality of clusters, the cluster data for each cluster identifying interchangeable part items by associating a cluster identifier and an item listing identifier for each part item listing represented by each item- level embedding in the cluster.

2 . The one or more computer storage media of claim 1 , wherein clustering the item-level embeddings to generate the plurality of clusters comprises:

generating similarity scores between pairs of item-level embeddings; and

forming the clusters using the similarity scores.

3 . The one or more computer storage media of claim 2 , wherein forming the clusters using the similarity scores comprises:

generating a similarity matrix that identifies similarity between pairs of item-level embeddings based on the similarity scores;

generating a plurality of graphs, each graph having nodes representing part item listings and edges between nodes based on similarity between the item-level embeddings corresponding to the nodes; and

identifying each graph as a cluster.

4 . The one or more computer storage media of claim 3 , wherein the similarity matrix is constrained by one or more constraint rules.

5 . The one or more computer storage media of claim 3 , wherein each graph from the plurality of graphs is fully connected.

6 . The one or more computer storage media of claim 3 , wherein forming the clusters using the similarity scores further comprises:

identifying a first part item listing in a first graph as an outlier based on price; and

removing the first part item listing from the first graph.

7 . The one or more computer storage media of claim 1 , wherein the operations further comprise:

receiving a search query;

identifying, from the plurality of clusters, a first cluster based on the search query; and

providing one or more search results based on the cluster data for the first cluster.

8 . The one or more computer storage media of claim 7 , wherein the search query comprises a part number, and wherein the first cluster is identified based on the first cluster having an item listing identifier for a part item listing having the part number.

9 . The one or more computer storage media of claim 7 , wherein identifying, from the plurality of clusters, the first cluster based on the search query comprises:

identifying one or more part item listings relevant to the search query; and

identifying the first cluster based on the one or more part item listings.

10 . The one or more computer storage media of claim 1 , wherein the operations further comprise:

receiving a seed item;

identifying, from the plurality of clusters, a first cluster based on the seed item; and

providing one or more part item recommendations based on the cluster data for the first cluster.

11 . A computer-implemented method comprising:

accessing item data for each of a plurality of part item listings, the item data for each part item listing comprising fitment data;

for each part item listing, embedding the item data including the fitment data to generate an item-level embedding to provide a plurality of item-level embeddings for the plurality of part item listings, wherein each item-level embedding represents a different part item listing from the plurality of part item listings based on at least the fitment data for each part item listing;

clustering the plurality of item-level embeddings for the plurality of part item listings to generate a plurality of clusters, each cluster comprising a subset of the plurality of item-level embeddings for a corresponding subset of different part item listings from the plurality of part item listings;

storing, in a data store, cluster data for the plurality of clusters, the cluster data for each cluster identifying interchangeable part items by associating a cluster identifier and an item listing identifier for each part item listing represented by each item-level embedding in the cluster;

receiving a search input;

identifying, from the plurality of clusters, a first cluster based on the search input; and

providing, as a response to the search input, an indication of one or more part item listings based on the cluster data for the first cluster.

12 . The computer-implemented method of claim 11 , wherein clustering the item-level embeddings to generate the plurality of clusters comprises:

generating similarity scores between pairs of item-level embeddings; and

forming the clusters using the similarity scores.

13 . The computer-implemented method of claim 12 , wherein forming the clusters using the similarity scores comprises:

generating a similarity matrix that identifies similarity between pairs of item-level embeddings based on the similarity scores;

generating a plurality of graphs, each graph having nodes representing part item listings and edges between nodes based on similarity between the item-level embeddings corresponding to the nodes; and

identifying each graph as a cluster.

14 . The computer-implemented method of claim 13 , wherein the similarity matrix is constrained by one or more constraint rules.

15 . The computer-implemented method of claim 13 , forming the clusters using the similarity scores further comprises:

identifying a first part item listing in a first graph as an outlier based on price; and

removing the first part item listing from the first graph.

16 . The computer-implemented method of claim 11 , wherein the search input comprises a search query identifying a part number, and wherein the first cluster is identified based on the first cluster having an item listing identifier for a part item listing having the part number.

17 . The computer-implemented method of claim 11 , wherein the search input comprises a search query, and wherein identifying, from the plurality of clusters, the first cluster based on the search query comprises:

identifying an initial set of one or more part item listings relevant to the search query; and

identifying the first cluster based on the initial set of one or more part item listings.

18 . The computer-implemented method of claim 11 , wherein the search input comprises a seed item.

19 . A computer system comprising:

one or more processors; and

one or more computer storage media storing computer-useable instructions that, when used by the one or more processors, cause the computer system to perform operations, the operations comprising:

accessing item data for each of a plurality of part item listings, the item data for each part item listing comprising fitment data;

for each part item listing, embedding the item data including the fitment data to generate an item-level embedding to provide a plurality of item-level embeddings for the plurality of part item listings, wherein each item-level embedding represents a different part item listing from the plurality of part item listings based on at least the fitment data for each part item listing;

generating a constrained similarity matrix based on similarity scores between pairs of item-level embeddings and one or more constraint rules;

generating a plurality of graphs using the constrained similarity matrix, each graph having nodes representing part item listings and edges between nodes based on the constrained similarity matrix, wherein each graph is identified as a cluster to provide a plurality of clusters, each cluster comprising a subset of the plurality of item-level embeddings for a corresponding subset of different part item listings from the plurality of part item listings; and

storing cluster data for the plurality of clusters, the cluster data for each cluster identifying interchangeable part items by associating a cluster identifier and an item listing identifier for each item embedding in the cluster.

20 . The computer system of claim 19 , wherein the operations further comprise:

receiving an input;

identifying, from the plurality of clusters, a first cluster based on the input; and

providing an indication of one or more part item listings based on the cluster data for the first cluster.