Solving sparse data problems in a recommendation system with freezing start

Patent metadata

Field	Value
Publication number	US-2025384477-A1
Application number	US-202519302833-A
Country	US
Kind code	A1
Filing date	Aug 18, 2025
Priority date	Dec 16, 2022
Publication date	Dec 18, 2025
Grant date	—

Abstract

Recommendation services typically struggle with sparse data scenarios. A freezing generated item start technique can use a matrix of external generated items to find a linking generated item. Embeddings can be used to determine distance between items. The technologies are useful for providing recommendations even in scenarios involving little or no transaction data.

First claim

What is claimed is:

1 . A system comprising:
a generated-item-by-generated-item matrix stored in one or more non-transitory computer-readable media and representing observed co-occurrences within transactions of external generated items represented in the matrix; an internal representation of a high-dimensional warm embedded vector space in which a plurality of warm generated items out of the external generated items are represented, and a high-dimensional freezing embedded vector space in which a plurality of freezing generated items are represented; and a recommendation extension service configured to, by at least one hardware processor:
for a freezing generated item from the plurality of freezing generated items, identify a warm approximate nearest neighbor of the freezing generated item in the high-dimensional warm embedded vector space;
search candidate linking items found in matrix entries of the warm approximate nearest neighbor for linked freezing items within a threshold distance of the candidate linking items in the high-dimensional freezing embedded vector space, and
responsive to finding a linked freezing generated item, associate the freezing generated item with the linked freezing generated item using the generated-item-by-generated-item matrix as a bridge to enable the association.

2 . The system of claim 1 further comprising:
a recommendation service configured to receive a request for a generated item recommendation for an input generated item, wherein the recommendation service provides a recommendation comprising the linked freezing generated item based on the association between the freezing generated item and the linked freezing generated item.

3 . The system of claim 1 , wherein the plurality of warm generated items are represented in the high-dimensional warm embedded vector space based on respective textual descriptions of the warm generated items.

4 . The system of claim 1 further comprising:
an internal representation of a hierarchical category tree;
wherein the linked freezing generated item is filtered with fallback category filtering.

5 . The system of claim 1 further comprising a pruner configured to remove items having less than a threshold number of co-occurrences from the generated-item-by-generated-item matrix to generate a pruned matrix, wherein the pruned matrix represents the plurality of warm generated items.

6 . The system of claim 1 , wherein the recommendation extension service is configured to normalize the generated-item-by-generated-item matrix.

7 . The system of claim 6 , wherein the normalization is based on diagonal values of the generated-item-by-generated-item matrix, wherein the diagonal values represent numbers of observed co-occurrence transactions for respective generated items.

8 . The system of claim 6 , wherein the normalization comprises dividing entries in the generated-item-by-generated-item matrix by a sum of values for entries in a corresponding row or column of the generated-item-by-generated-item matrix.

9 . The system of claim 1 , wherein the high-dimensional freezing embedded vector space and the high-dimensional warm embedded vector space are a same vector space represented by different indexes.

10 . The system of claim 9 , wherein identifying the warm approximate nearest neighbor in the high-dimensional warm embedded vector space comprises searching a first index, and the searching for the linked freezing items within the threshold distance comprises searching a second index different from the first index.

11 . The system of claim 1 , wherein the searching for linked freezing items within the threshold distance comprises searching for an approximate nearest neighbor of a candidate linking item in the high-dimensional freezing embedded vector space.

12 . The system of claim 1 , wherein the recommendation extension service is configured to iterate over the plurality of freezing generated items to perform the identifying, searching, and associating for each of the plurality of freezing generated items.

13 . The system of claim 1 , wherein the recommendation extension service is configured to identify the candidate linking items based on non-zero entries in a row or column of the generated-item-by-generated-item matrix corresponding to the warm approximate nearest neighbor.

14 . The system of claim 1 , wherein the recommendation extension service is configured to identify a plurality of linked freezing generated items for the freezing generated item and to output a ranked list of the linked freezing generated items.

15 . The system of claim 14 , wherein the ranked list of the linked freezing generated items comprises N linked freezing generated items having N greatest affinity values, where N is greater than or equal to 1.

16 . The system of claim 15 , wherein the recommendation extension service is configured to determine the affinity values based on a relative number of co-occurrences, as indicated in the generated-item-by-generated-item matrix, between the warm approximate nearest neighbor and respective candidate linking items.

17 . The system of claim 15 , wherein the recommendation extension service is configured to determine the affinity values based on distances, in the high-dimensional freezing embedded vector space, between respective candidate linking items and linked freezing generated items, wherein smaller distances indicate greater affinity.

18 . The system of claim 1 , wherein the recommendation extension service is configured to incorporate category information of the generated items into their representation in the embedded vector space by adding a vector representing the category information to a title-based representation of the item or by concatenating a one-hot encoding of the category information to the title-based representation.

19 . A computing system, comprising:
memory; at least one processor in communication with the memory; and a non-transitory computer-readable medium storing instructions that, when executed by the at least one processor, cause the computing system to perform operations comprising: building a generated-item-by-generated-item matrix representing correlations between external generated items based on observed transactions for the external generated items from an external database; pruning, by a pruner, the generated-item-by-generated-item matrix; embedding the external generated items into a high-dimensional warm embedded vector space; for a plurality of freezing generated items, embedding the plurality of freezing generated items into a high-dimensional freezing embedded vector space; for a freezing generated item out of the plurality of freezing generated items, finding a nearest neighbor external generated item r of the freezing generated item within a complementary threshold distance of the freezing generated item in the high-dimensional warm embedded vector space; for a given linking external generated item out of a plurality of possible linking external generated items appearing in the matrix for r, determining whether a linked freezing nearest neighbor generated item exists for the given linking external generated item in the high-dimensional freezing embedded vector space; and responsive to determining that a linked freezing nearest neighbor generated item exists, associating the freezing generated item with the linked freezing nearest neighbor generated item using the generated-item-by-generated-item matrix as a bridge to enable the association, and outputting a recomm

Assignees

• Sap Se

Inventors

• Pendar Nick
• Faghihi Seyed Farhad

Classifications

• G06Q30/0202 Primary

  Market predictions or forecasting for commercial activities · CPC title

• G06Q30/0631 Primary

  Recommending goods or services · CPC title