Feature attribution methods – Explainable AI
Each approach for attributing features is based on Shapley values, which is an algorithm derived from cooperative game theory that gives credit for a given result to each participant in a game. When this concept is applied to models for machine learning, it indicates that each model feature is dealt with as if it were a “player” in the game. The Vertex Explainable AI provides a certain amount of credit to each individual characteristic based on its weight in the overall forecast:
- Sampled Shapley method: The sampled Shapley technique offers an estimate of the actual Shapley values via the use of sampling. Tabular models created using AutoML make use of the sampled Shapley approach to determine the relevance of features. For these models, which are meta-ensembles of tree and neural network structures, the Sampled Shapley method performs quite well.
- Integrated gradients method: Along an integral route, the integrated gradients approach calculates the gradient of the prediction output with respect to the characteristics of the input. This is done to get an accurate result. Calculations of the gradients are performed at various time intervals along a scaling parameter. Utilizing the Gaussian quadrature rule allows for the calculation of the size of each interval. (When dealing with picture data, think of this scaling option as a “slider” that sets all the image’s pixels to a black value.) The integration of the gradients is done as follows:
- An approximation of the integral may be found by using a weighted average.
- Calculations are performed to get the element-wise product of the original input and the averaged gradients.
- XRAI method: To discover which parts of a picture that contribute the most to a certain prediction of class, the XRAI approach uses a combination of the integrated gradients method and some extra phases.
- Pixel-level attribution: XRAI can do pixel-level attribution for the picture that is sent into it. In this stage of the process, XRAI makes use of the integrated gradients approach, applying it to both, a black and a white baseline.
- Over segmentation: XRAI generates a tiny patch over the picture by over segmentation , which is done independently of pixel-level attribution. In order to construct the picture segments, XRAI takes advantage of Felzenswalb’s graph-based technique.
- Region selection: XRAI compiles the pixel-level attribution included inside each segment to calculate the attribution density of that segment. XRAI assigns a ranking to each segment based on these values, and then it arranges the segments from most positive to least positive. This identifies which parts of the picture contribute the most strongly to a certain class prediction, as well as which parts of the image are most prominent.
All types of models are supported for feature-based explanations. Classification models are supported for AutoML images and classification, and regression models are supported for AutoML tabular models.