Cyclic Boosting for Explainable Demand Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Supervised machine learning methods, particularly in complex ensemble or deep learning models, fail to provide transparent predictions and struggle to learn rare observations, making it difficult to identify the underlying causes of predictions and accurately forecast demand in supply chain management.
Innovation Solution
The cyclic boosting system, which includes a machine learning method that trains a prediction model using historical sales data, generates predictions, and identifies the contribution of each feature variable to the prediction, providing transparent insights into the prediction process and effectively handling rare observations by using a combination of data processing, training, and prediction modules within a supply chain network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complex ensemble or deep learning models are used to improve prediction accuracy, then prediction accuracy is improved, but model interpretability deteriorates
Solution Approach 1:
The patent segments the prediction process into multiple components by introducing feature importance scores and contribution metrics that break down the complex model output into interpretable parts. Each feature's contribution to the prediction is explicitly calculated and reported, allowing users to understand which features drive the prediction and why.
Solution Approach 2:
The patent introduces an intermediary layer between the complex model and the user by computing and reporting feature importance scores and contribution metrics. This intermediary provides a bridge that translates the black-box model outputs into explainable insights, showing how individual features influence the prediction without requiring users to understand the complex underlying model architecture.
2Productivity
If standard machine learning techniques are used to handle common observations, then they work well for frequent data, but they fail to learn rare observations
Solution Approach 1:
The patent applies local quality by treating rare and common observations differently through weighted sampling and importance weighting. Rare observations are given higher weights during training to ensure they are adequately learned, while common observations maintain their natural frequency. This allows the model to pay special attention to rare events without ignoring them entirely.
Solution Approach 2:
The patent performs preliminary actions by pre-processing the data to identify and highlight rare observations before training. Techniques such as rare instance detection and weighted sampling are applied in advance to ensure that rare observations are properly represented in the training set, preparing the model to learn from them effectively during the training process.
3Loss of information
If feature importance quantification is added to explain predictions, then interpretability is improved, but computational complexity increases
Solution Approach 1:
The patent applies partial action by computing feature importance metrics only for the top-contributing features rather than all features. This selective approach provides sufficient interpretability for the most important decisions while reducing computational overhead. The model identifies and reports only the features that make the largest contribution to the prediction, avoiding unnecessary computations for less relevant features.
Data Source
AI summary
A system and method are disclosed including a computer and a processor and memory. The computer receives historical sales data comprising aggregated sales data for one or more items from one or more store for at least one past time period. The computer further trains a cyclic boosting model to learn model parameters by iteratively calculating for each feature and each bin factors for at least one full feature cycle. The computer further predicts one or more demand quantities during a prediction period by applying a prediction model to historical supply chain data, wherein a training period is earlier than the prediction period, and each of the one or more demand quantities is associated with at least one item of the one or more items and at least one stocking location of the one or more stocking locations during the prediction period and rendering a demand prediction feature explanation visualization.