Hierarchical Demand Forecasting Model for Sparse Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models for demand forecasting face challenges with large, imbalanced, and heterogeneous data sets, increasing complexity and resource requirements, and making it difficult to extrapolate demand for items with little historical data or new locations.
Innovation Solution
A hierarchical modeling system that generates batches of training data, trains a base model to learn an uncensored demand structure, and uses embeddings to train downstream models for efficient forecasting, reducing the need for extensive data and computing resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large data sets are used to train machine learning models, then model accuracy is improved, but training complexity and computing resource requirements increase
Solution Approach 1:
The patent segments the training process into two distinct stages: a first stage that trains a base model on a subset of training data to learn uncensored demand structure, and a second stage that trains multiple downstream models using embeddings from the base model. This segmentation allows the system to achieve high model accuracy without requiring all training data to be processed simultaneously, thereby reducing training complexity and computing resource requirements.
Solution Approach 2:
The patent extracts key information from the training data in the form of embeddings generated by the base model. These embeddings capture the essential demand structure patterns and are used to train downstream models, effectively separating the heavy computational task of learning demand structure from the task of generating specific forecasts. This extraction approach reduces the data processing burden while maintaining model accuracy.
2Measurement precision
If large data sets are used to train machine learning models, then model accuracy is improved, but the time required for training increases
Solution Approach 1:
The training process is divided into temporal segments: first stage training the base model on a subset of data, then second stage training multiple downstream models using the base model's embeddings. This temporal segmentation allows the system to leverage pre-computed embeddings from the base model, significantly reducing the total training time required compared to training all models from scratch on the full dataset.
Solution Approach 2:
The base model is trained in advance on a subset of training data to learn the uncensored demand structure and generate embeddings. This preliminary action creates a foundation that can be reused by multiple downstream models, eliminating the need to re-process the entire training data for each downstream model and thereby reducing overall training time.
3Measurement precision
If large data sets are used to train machine learning models, then model accuracy is improved, but computing resources required increase
Solution Approach 1:
The computing workload is segmented into two stages: first stage computes embeddings from the base model on a subset of data, and second stage trains multiple downstream models using these pre-computed embeddings. This segmentation significantly reduces the total computing resources required compared to training all models simultaneously on the full dataset, as the expensive embedding computation is performed only once and reused across all downstream models.
Solution Approach 2:
The patent extracts demand structure information into embeddings that can be efficiently stored and reused. By taking out this essential information in a compressed form, the system avoids the need to process the entire training data multiple times for different downstream models, thereby reducing computing resource consumption while maintaining high model accuracy.
4Adaptability or versatility
If models are trained on imbalanced and heterogenous data, then coverage is expanded, but learning the data structure becomes difficult
Solution Approach 1:
The patent segments the data processing into two stages: first stage trains the base model on a subset of training data to learn uncensored demand structure, and second stage uses the base model's embeddings to train downstream models for specific forecasting tasks. This segmentation allows the system to handle imbalanced and heterogenous data more effectively, as the base model learns the general structure from representative data while the downstream models specialize in specific areas, reducing the overall complexity of learning the data structure.
Data Source
AI summary
Methods and systems are described for providing hierarchical demand forecasting for state space reduction. Using a hierarchical architecture, a base model may be trained to capture a range of shared structure in a first data set that can be used to draw inferences on using smaller sets of data representative of the “whole picture.” For example, training data may be sampled and prepared and used to train a base model in a first stage. In a next stage, one or more downstream models may be trained on the structure and samples of uncensored demand generated by the base model to produce forecasts for items and locations, including items and locations for which there may be little or no historical data. The downstream models that would otherwise require a large amount of data for training can be generated on demand using less training data, training time, computing processing, and memory.