# US 2023/0267481 A1 · Hierarchical neural network based implementation for predicting out of stock products

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2023/0267481 A1
- Document ID: US20230267481A1
- Status: Published application
- Kind: A1
- Application number: 17/675,817
- Filed: 2022-02-18
- Published: 2023-08-24
- Earliest claimed: 2022-02-18
- Assignees: salesforce.com, inc.
- Applicants: salesforce.com, inc.
- Inventors: Akash Singh; Rajdeep Dua
- Web page: https://patentgrep.com/patent/US20230267481A1

## CPC

- G06Q 30/0202: Market predictions or forecasting for commercial activities
- 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 3/08: Learning methods
- G06Q 10/087: Inventory or stock management, e.g. order filling, procurement or balancing against orders
- G06Q 10/08726: {for replenishment processing, procedures, or recommendations using forecasting or optimisation}
- G06Q 30/0201: Market modelling; Market analysis; Collecting market data
- G06N 5/01: Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
- G06N 20/20: Ensemble learning
- G06Q 10/0874: {Inventory fulfillment (shipping G06Q10/083)}
- 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
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A hierarchical neural network for predicting out of stock products comprises an input layer that receives data from data sources that store disparate datasets having different levels of attribute detail pertaining to products for sale in stores of a retailer. A first level of neural networks processes the data from the data sources into respective learned intermediate vector representations. A second level comprises a concatenate layer that concatenates the learned intermediate vector representations from the second level into a combined vector representation. A third level comprises a feed forward network that receives the combined vector representation and outputs to the retailer an out of stock probability indicating which store and product combinations are likely to have out of stock products over a predetermined timeframe.

## Claims

### Claim 1

A hierarchical neural network (HNN) for predicting out of stock products, the hierarchical neural network executed by one or more processors, the HNN comprising:
one or more data sources that store disparate datasets having different levels of attribute detail pertaining to products for sale in one or more stores of a retailer, the disparate datasets comprising: low-level store data including daily store visits and past occurrences of out of stock products, categorical data including store identifiers and associated city identifiers, store-level product data, and time-series product sales data including monthly or weekly sales per product;
a first level that processes the data from one or more data sources into respective learned intermediate vector representations, the first level comprising:
a convolutional neural network (CNN) layer that processes the low-level store data;
an embedding layer that processes the categorical data;
a long short-term memory (LSTM) layer that processes the time-series product sales data;
a second level comprising a concatenate layer that concatenates the learned intermediate vector representations from the second level into a combined vector representation; and
a third level comprising a feed forward network that receives the combined vector representation and outputs to the retailer an out of stock probability indicating which store and product combinations are likely to have out of stock products over a predetermined timeframe.

### Claim 2 (depends on 1)

The HNN of claim 1, wherein the first level first translates the data into initial vector representations.

### Claim 3 (depends on 2)

The HNN of claim 2, wherein a set of initial vector representations representing the store-level product data are not processed by the first level 1 but passed on to the second level.

### Claim 4 (depends on 3)

The HNN of claim 3, wherein the concatenate layer concatenates the set of initial vector representations of the store-level product data with the respective learned intermediate vector representations.

### Claim 5 (depends on 1)

The FINN of claim 1, wherein CNN layer is based on a shared-weight architecture of convolution kernels that slide along the visit data and the visit data is convolved with one the convolutions kernels to provide feature maps, where ones of the feature maps learn different relevant features.

### Claim 6 (depends on 5)

The FINN of claim 5, wherein CNN layer includes a first attention layer such that the CNN layer processes low-level granular and sequential features and the first attention layer examines data from a predefined time period relevant to generating the OOS probability.

### Claim 7 (depends on 1)

The HNN of claim 1, wherein embedding layer processes the store identifiers and the city identifiers to capture information including interaction of the stores and cities with other variables.

### Claim 8 (depends on 1)

The FINN of claim 1, wherein the LSTM layer include stacked LSTM layers to learn historical patterns that are represented by LSTM state vectors, wherein the LSTM state vectors are fed into a second attention layer to enable the LSTM layer to examine data from a predefined time period relevant to generating the OOS probability.

### Claim 9 (depends on 1)

The HNN of claim 1, wherein the one or more data sources are stored in: a data center of a service provider, ii) by a cloud service provider, iii) local to server comprising the one or more processors, or iv) by a computer system of the retailer or a consumer goods company.

### Claim 10

A method for implementing a hierarchical neural network (HNN) for predicting out of stock products, the HNN executed by a computer server, the HNN comprising:
receiving one or more data sources that store disparate datasets having different levels of attribute detail pertaining to products for sale in one or more stores of a retailer, the disparate datasets comprising: low-level store data including daily store visits and past occurrences of out of stock products, categorical data including store identifiers and associated cities, store-level product data, and time-series product sales data including monthly or weekly sales per product;
processing, by a first level, the data from the one or more data sources into respective learned intermediate vector representations, the first level comprising: a convolutional neural network to process the low-level store data, an embedding layer to process the categorical data; and a long short-term memory (LSTM) layer to process the time-series product sales data;
concatenating, by a second level comprising a concatenate layer, the respective learned intermediate vector representations into a combined vector representation; and
receiving, by a third level comprising a feed forward network, the combined feature vector representation and outputting to the retailer an out of stock probability indicating which store and product combinations are likely to have out of stock products over a predetermined timeframe.

### Claim 11 (depends on 10)

The method of claim 10, further comprising translating the data into initial vector representations.

### Claim 12 (depends on 10)

The method of claim 10, further comprising generating a set of initial vector representations representing the store-level product data, and passing the set of initial vector representations on to the second level instead of the first level.

### Claim 13 (depends on 12)

The method of claim 12, further comprising concatenating, by the concatenate layer, the set of initial vector representations of the store-level product data with the respective learned intermediate vector representations.

### Claim 14 (depends on 10)

The method of claim 10, further comprising basing the CNN layer on a shared-weight architecture of convolution kernels that slide along the visit data and the visit data is convolved with one the convolutions kernels to provide feature maps, where ones of the feature maps learn different relevant features.

### Claim 15 (depends on 14)

The method of claim 14, further comprising providing the CNN layer with a first attention layer such that the CNN layer processes low-level granular and sequential features and the first attention layer examines data from a predefined time period relevant to generating the OOS probability.

### Claim 16 (depends on 10)

The method of claim 10, further comprising processing, by the embedding layer, the store identifiers and the city identifiers to capture information including interaction of the stores and cities with other variables.

### Claim 17 (depends on 10)

The method of claim 10, further comprising providing the LSTM layer with stacked LSTM layers to learn historical patterns that are represented by LSTM state vectors, and feeding the LSTM state vectors into a second attention layer to enable the LSTM layer to examine data from a predefined time period relevant to generating the OOS probability.

### Claim 18 (depends on 10)

The method of claim 10, further comprising storing the one or more data sources in: a data center of a service provider, ii) by a cloud service provider, iii) local to server comprising the one or more processors, or iv) by a computer system of the retailer or a consumer goods company.

### Claim 19

A non-transitory machine-readable storage medium that provides instructions that, if executed by a processor are configurable to cause said processor to perform operations comprising:
receiving data from one or more data sources that store disparate datasets having different levels of attribute detail pertaining to products for sale in one or more stores of a retailer, the disparate datasets comprising: low-level store data including daily store visits and past occurrences of out of stock products, categorical data including store identifiers and associated cities, store-level product data, and time-series product sales data including monthly or weekly sales per product;
processing, by a first level, the data from the one or more data sources into respective learned intermediate vector representations, the first level comprising: a convolutional neural network to process the low-level store data, an embedding layer to process the categorical data; and a long short-term memory (LSTM) layer to process the time-series product sales data;
concatenating, by a second level comprising a concatenate layer, the respective learned intermediate vector representations into a combined vector representation; and
receiving, by a third level comprising a feed forward network, the combined feature vector representation and outputting to the retailer an out of stock probability indicating which store and product combinations are likely to have out of stock products over a predetermined timeframe.
