# US 12,112,520 B2 · Scalable pipeline for machine learning-based base-variant grouping > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,112,520 B2 - Document ID: US12112520B2 - Status: Granted - Kind: B2 - Application number: 17/589,768 - Filed: 2022-01-31 - Published: 2022-07-14 - Granted: 2024-10-08 - Earliest claimed: 2020-01-31 - Assignees: Walmart Apollo, LLC - Applicants: Walmart Apollo, LLC - Inventors: Yanxin Pan; Swagata Chakraborty; Ekaterina Pirogova - Web page: https://patentgrep.com/patent/US12112520B2 ## CPC - G06F 16/55: Clustering; Classification - G06V 10/7625: {Hierarchical techniques, i.e. dividing or merging patterns to obtain a tree-like representation; Dendograms} - G06F 16/5846: {using extracted text} - G06N 3/044: Recurrent networks, e.g. Hopfield networks - G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] - G06N 3/045: Combinations of networks - G06N 3/0464: Convolutional networks [CNN, ConvNet] - G06N 3/09: Supervised learning - G06N 20/00: Machine learning - G06Q 30/0603: {Catalogue creation or management} - G06Q 30/0641: {utilising user interfaces specially adapted for shopping} - G06V 10/7747: {Organisation of the process, e.g. bagging or boosting} - G06V 10/82: using neural networks - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) - G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS - G06Q: INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR ## Abstract A system including one or more processors and one or more non-transitory computer-readable media storing computing instructions configured to run on the one or more processors and perform: creating an adjacency list for candidate items using a distance threshold; generating graphs of the candidate items in the adjacency list, wherein nodes of the graphs represent the candidate items, and wherein edges of the graphs represent respective predicted variant neighbor links between pairs of the candidate items; determining, using breakdown logic, first graphs of the graphs that exceed a predetermined size; performing divisive hierarchical clustering on each of the first graphs; and identifying recommended variant groups of the candidate item in the nested subclusters of the hierarchy dendrogram below the respective cut-off value. Other embodiments are described. ## Claims ### Claim 1 A system comprising: one or more processors; and one or more non-transitory computer-readable media storing computing instructions that, when executed on the one or more processors, perform: creating an adjacency list for candidate items using a distance threshold, wherein the adjacency list for the candidate items are derived from N nearest neighbors, wherein an Nth neighbor of the N nearest neighbors is a last neighbor with a maximum allowable neighbor distance equal to a Siamese model cut-off; loading data for the candidate items in the adjacency list by at least one of: (i) loading Siamese embeddings into a similarity search algorithm; (ii) building an index; or (iii) generating connected components using the Nth neighbor of the N nearest neighbors; generating graphs of the candidate items in the adjacency list, wherein nodes of the graphs represent the candidate items, and wherein edges of the graphs represent respective predicted variant neighbor links between pairs of the candidate items; determining, using breakdown logic, first graphs of the graphs that exceed a predetermined size; when the connected components exceed a predetermined threshold, performing divisive hierarchical clustering on each of the first graphs comprising: building a respective hierarchy dendrogram of nested subclusters of each of the first graphs using respective linkage distances between item embeddings of the candidate items of the pairs of the candidate items of the first graphs; and determining a respective cut-off value based on a respective p-th percentile of density of the respective linkage distances for each of the first graphs; and identifying recommended variant groups of the candidate item in the nested subclusters of the respective hierarchy dendrogram below the respective cut-off value. ### Claim 2 (depends on 1) The system of claim 1, wherein the computing instructions that, when executed on the one or more processors can further perform, before creating the adjacency list: generating text embeddings of the candidate items, using a first pre-trained Siamese neural network and a first contrastive loss function, wherein the item embeddings of the candidate items comprise the text embeddings, wherein the text embeddings are converted to first N-dimensional feature embeddings, and wherein the first N-dimensional feature embeddings comprise encoded vector representations of the candidate items that are used to compute similarity between the candidate items for variant grouping. ### Claim 3 (depends on 2) The system of claim 2, wherein the computing instructions that, when executed on the one or more processors can further perform, before creating the adjacency list: generating image embeddings of the candidate items, using a second pre-trained Siamese neural network, wherein the item embeddings of the candidate items further comprise the image embeddings, wherein an output from the second pre-trained Siamese neural network is flattened to second N-dimensional feature embeddings, wherein the second N-dimensional feature embeddings comprise numeric vector representations of the candidate items to predict whether the candidate items are variants of one another. ### Claim 4 (depends on 3) The system of claim 3, wherein the computing instructions that, when executed on the one or more processors can further perform: indexing, using the similarity search algorithm, each of the item embeddings in a feature space for retrieval of the N nearest neighbors. ### Claim 5 (depends on 4) The system of claim 4, wherein the similarity search algorithm is Facebook Artificial Intelligence Similarity search (FAISS). ### Claim 6 (depends on 1) The system of claim 1, wherein creating the adjacency list further comprises: for each of the candidate items, determining respective nearest neighbors of each of the candidate items based on a respective pre-trained Siamese neural network model threshold generated for each of the candidate items. ### Claim 7 (depends on 1) The system of claim 1, wherein the distance threshold is determined using a union logic for the pairs of the candidate items. ### Claim 8 (depends on 1) The system of claim 1, wherein the predetermined size is approximately 5000. ### Claim 9 (depends on 1) The system of claim 1, wherein the respective cut-off value is inversely proportional to a respective size of each of the first graphs. ### Claim 10 (depends on 1) The system of claim 1, wherein building the respective hierarchy dendrogram of the nested subclusters further comprises sorting the respective linkage distances for each of the first graphs. ### Claim 11 A method being implemented via execution of computing instructions configured to run on one or more processors and stored at one or more non-transitory computer-readable media, the method comprising: creating an adjacency list for candidate items using a distance threshold, wherein the adjacency list for the candidate items are derived from N nearest neighbors, wherein an Nth neighbor of the N nearest neighbors is a last neighbor with a maximum allowable neighbor distance equal to a Siamese model cut-off; loading data for the candidate items in the adjacency list by at least one of: (i) loading Siamese embeddings into a similarity search algorithm; (ii) building an index; or (iii) generating connected components using the Nth neighbor of the N nearest neighbors; generating graphs of the candidate items in the adjacency list, wherein nodes of the graphs represent the candidate items, and wherein edges of the graphs represent respective predicted variant neighbor links between pairs of the candidate items; determining, using breakdown logic, first graphs of the graphs that exceed a predetermined size; when the connected components exceed a predetermined threshold, performing divisive hierarchical clustering on each of the first graphs, comprising: building a respective hierarchy dendrogram of nested subclusters of each of the first graphs using respective linkage distances between item embeddings of the candidate items of the pairs of the candidate items of the first graphs; and determining a respective cut-off value based on a respective p-th percentile of density of the respective linkage distances for each of the first graphs; and identifying recommended variant groups of the candidate item in the nested subclusters of the respective hierarchy dendrogram below the respective cut-off value. ### Claim 12 (depends on 11) The method of claim 11 further comprising, before creating the adjacency list: generating text embeddings of the candidate items, using a first pre-trained Siamese neural network and a first contrastive loss function, wherein the item embeddings of the candidate items comprise the text embeddings, wherein the text embeddings are converted to first N-dimensional feature embeddings, and wherein the first N-dimensional feature embeddings comprise encoded vector representations of the candidate items that are used to compute similarity between the candidate items for variant grouping. ### Claim 13 (depends on 12) The method of claim 12 further comprising, before creating the adjacency list: generating image embeddings of the candidate items, using a second pre-trained Siamese neural network, wherein the item embeddings of the candidate items further comprise the image embeddings, wherein an output from the second pre-trained Siamese neural network is flattened to second N-dimensional feature embeddings, wherein the second N-dimensional feature embeddings comprise numeric vector representations of the candidate items to predict whether the candidate items are variants of one another. ### Claim 14 (depends on 13) The method of claim 13 further comprising: indexing, using the similarity search algorithm, each of the item embeddings in a feature space for retrieval of the N nearest neighbors. ### Claim 15 (depends on 14) The method of claim 14, wherein the similarity search algorithm is Facebook Artificial Intelligence Similarity search (FAISS). ### Claim 16 (depends on 11) The method of claim 11, wherein creating the adjacency list further comprises: for each of the candidate items, determining respective nearest neighbors of each of the candidate items based on a respective pre-trained Siamese neural network model threshold generated for each of the candidate items. ### Claim 17 (depends on 11) The method of claim 11, wherein the distance threshold is determined using a union logic for the pairs of the candidate items. ### Claim 18 (depends on 11) The method of claim 11, wherein the predetermined size is approximately 5000. ### Claim 19 (depends on 11) The method of claim 11, wherein the respective cut-off value is inversely proportional to a respective size of each of the first graphs. ### Claim 20 (depends on 11) The method of claim 11, wherein building the respective hierarchy dendrogram of the nested subclusters further comprises sorting the respective linkage distances for each of the first graphs.