Cloud-based data warehouse BigQuery officially provides users with Explainable AI (XAI) functions, enabling users to better understand how machine learning models make decisions. The official mentions that BigQuery provides various XAI methods for different model types, and comprehensively supports explainable artificial intelligence, allowing users to use a single SQL query and get millions of explanations in seconds.
Many companies have introduced machine learning technology internally, but to make AI decisions more credible, AI must be interpretable. Google mentioned that when training machine learning models, there are two types of interpretability related to features. are global interpretability and regional interpretability, respectively.
The global interpretable describes the overall impact of features on the model, allowing users to understand whether specific features have a greater impact on the prediction of the model than other features. When the model has hundreds or thousands of features, users want to know which features are the main contributors to the model, global interpretability can be particularly useful, allowing users to prune less important features to improve the generality of the model.
Regional interpretability, on the other hand, describes the contribution of each feature to a specific prediction. Taking house prices as an example, different characteristics of houses may contribute to house prices. For example, a house with 3 bedrooms may contribute additionally to house prices. $50,000, $100,000 nearer downtown, built-in 2010, maybe $30,000, etc. Google says that the purpose of regional interpretability is to understand the exact contribution of each feature of the model to predictions.
BigQuery Explainable AI enables users to gain insight into the results generated by machine learning models in classification and regression tasks by defining the contribution of each feature in the data column to the predicted results. Google mentioned that this is called feature attribution, which This information can be used to verify that the model is behaving as expected, to identify biases in the model, and to obtain ways to improve the model and training data.
BigQuery Explainable AI is suitable for a variety of models, including supervised learning models with independent and identically distributed data (IID Data), and time series models. For each model, BigQuery Explainable AI has different explainability methods, and each explainable Sexual methods have their own way of operation.