Artificial intelligence models are increasingly used in manufacturing to inform decision making. Responsible decision making requires accurate forecasts and an understanding of the models’ behavior. Furthermore, the insights into the models’ rationale can be enriched with domain knowledge. This research builds explanations considering feature rankings for a particular forecast, enriching them with media news entries, datasets’ metadata, and entries from the Google knowledge graph. We compare two approaches (embeddings-based and semantic-based) on a real-world use case regarding demand forecasting. The embeddings-based approach measures the similarity between relevant concepts and retrieved media news entries and datasets’ metadata based on the word movers’ distance between embeddings. The semantic-based approach recourses to wikification and measures the Jaccard distance instead. The semantic-based approach leads to more diverse entries when displaying media events and more precise and diverse results regarding recommended datasets. We conclude that the explanations provided can be further improved with information regarding the purpose of potential actions that can be taken to influence demand and to provide “what-if” analysis capabilities.

No abstract

No abstract

While increasing empirical evidence suggests that global time series forecasting models can achieve better forecasting performance than local ones, there is a research void regarding when and why the global models fail to provide a good forecast. This paper uses anomaly detection algorithms and explainable artificial intelligence (XAI) to answer when and why a forecast should not be trusted. To address this issue, a dashboard was built to inform the user regarding (i) the relevance of the features for that particular forecast, (ii) which training samples most likely influenced the forecast outcome, (iii) why the forecast is considered an outlier, and (iv) provide a range of counterfactual examples to understand value changes, in the feature vector or the predicted value, can lead to a different outcome. Moreover, a modular architecture and a methodology were developed to iteratively remove noisy data instances from the train set, to enhance the overall global time series forecasting model performance. Finally, to test the effectiveness of the proposed approach, it was validated on two publicly available real-world datasets.