Abstract:Demand forecasts are the crucial basis for numerous business decisions, ranging from inventory management to strategic facility planning. While machine learning (ML) approaches offer accuracy gains, their interpretability and acceptance are notoriously lacking. Addressing this dilemma, we introduce Hierarchical Neural Additive Models for time series (HNAM). HNAM expands upon Neural Additive Models (NAM) by introducing a time-series specific additive model with a level and interacting covariate components.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: **How to achieve a balance between high accuracy and interpretability in time - series prediction, especially in the field of demand forecasting**.
Specifically, although existing machine - learning methods have significantly improved in prediction accuracy, their black - box nature leads to the non - interpretability of the models, thus affecting their acceptance and application in actual business. While traditional statistical methods have a certain degree of interpretability, they are often inferior to complex machine - learning models in prediction performance. To solve this dilemma, the author introduced a new model - **Hierarchical Neural Additive Models (HNAM)**, aiming to combine the powerful expressive ability of neural networks and the interpretability of additive models.
### Main contributions of the paper:
1. **Proposing the HNAM model**: This is a hierarchical neural additive model specifically used for time - series prediction. It realizes the transparency and interpretability of the prediction process by decomposing the prediction results into the influence of each covariate.
2. **User - specified interaction levels**: HNAM allows users to define the interaction levels between covariates according to business needs. For example, the effect of the day of the week can be estimated independently, while the holiday effect may depend on the day of the week and promotional activities. This design enables the model to better reflect the actual business logic and increase user trust.
3. **Competitive prediction performance**: The experimental results show that HNAM not only provides a reasonable explanation but also is comparable to the state - of - the - art black - box models (such as Temporal Fusion Transformers, TFT) in prediction accuracy, and even performs better in some cases.
4. **Improving human - machine interaction**: By providing an intuitive prediction interface, analysts can observe the specific contribution of each known covariate to the prediction result, so as to better integrate their professional knowledge and enhance their trust in the model.
### Formula representation:
The prediction formula of HNAM can be expressed as:
\[ y = g(S, T, P)+\sum_{i = 0}^{nc - 1}f_i(S, T, P, C[:, :i + 1])\cdot t(C[i, Th:]) \]
where:
- \( g(S, T, P) \) represents the baseline demand level, which is determined by static, time and past covariates.
- \( f_i(S, T, P, C[:, :i + 1]) \) is the coefficient network of the \( i \) - th causal covariate, taking into account all lower - level causal covariates.
- \( t(C[i, Th:]) \) is the transformation function of the causal covariate \( C \), including one - hot encoding and standardization.
In this way, HNAM can provide detailed explanations while maintaining prediction accuracy, thus solving the interpretability and acceptance problems faced by current models in practical applications.