Modeling feature interactions in tabular data remains a key challenge in predictive modeling, for example, as used for insurance pricing. This paper proposes the Tree-like Pairwise Interaction Network (PIN), a novel neural network architecture that explicitly captures pairwise feature interactions through a shared feed-forward neural network architecture that mimics the structure of decision trees. PIN enables intrinsic interpretability by design, allowing for direct inspection of interaction effects. Moreover, it allows for efficient SHapley's Additive exPlanation (SHAP) computations because it only involves pairwise interactions. We highlight connections between PIN and established models such as GA2Ms, gradient boosting machines, and graph neural networks. Empirical results on the popular French motor insurance dataset show that PIN outperforms both traditional and modern neural networks benchmarks in predictive accuracy, while also providing insight into how features interact with each another and how they contribute to the predictions.

This paper presents a method to explain the internal representation structure of a neural network for image generation. Specifically, our method disentangles primitive feature components from the intermediate-layer feature of the neural network, which ensures that each feature component is exclusively used to generate a specific set of image regions. In this way, the generation of the entire image can be considered as the superposition of different pre-encoded primitive regional patterns, each being generated by a feature component. We find that the feature component can be represented as an OR relationship between the demands for generating different image regions, which is encoded by the neural network. Therefore, we extend the Harsanyi interaction to represent such an OR interaction to disentangle the feature component. Experiments show a clear correspondence between each feature component and the generation of specific image regions.

A very popular model-agnostic technique for explaining predictive models is the SHapley Additive exPlanation (SHAP). The two most popular versions of SHAP are a conditional expectation version and an unconditional expectation version (the latter is also known as interventional SHAP). Except for tree-based methods, usually the unconditional version is used (for computational reasons). We provide a (surrogate) neural network approach which allows us to efficiently calculate the conditional version for both neural networks and other regression models, and which properly considers the dependence structure in the feature components. This proposal is also useful to provide drop1 and anova analyses in complex regression models which are similar to their generalized linear model (GLM) counterparts, and we provide a partial dependence plot (PDP) counterpart that considers the right dependence structure in the feature components.

Deep learning models have gained great popularity in statistical modeling because they lead to very competitive regression models, often outperforming classical statistical models such as generalized linear models. The disadvantage of deep learning models is that their solutions are difficult to interpret and explain, and variable selection is not easily possible because deep learning models solve feature engineering and variable selection internally in a nontransparent way. Inspired by the appealing structure of generalized linear models, we propose a new network architecture that shares similar features as generalized linear models, but provides superior predictive power benefiting from the art of representation learning. This new architecture allows for variable selection of tabular data and for interpretation of the calibrated deep learning model, in fact, our approach provides an additive decomposition in the spirit of Shapley values and integrated gradients.

A vastly growing literature on explaining deep learning models has emerged. This paper contributes to that literature by introducing a global gradient-based model-agnostic method, which we call Marginal Attribution by Conditioning on Quantiles (MACQ). Our approach is based on analyzing the marginal attribution of predictions (outputs) to individual features (inputs). Specificalllly, we consider variable importance by mixing (global) output levels and, thus, explain how features marginally contribute across different regions of the prediction space. Hence, MACQ can be seen as a marginal attribution counterpart to approaches such as accumulated local effects (ALE), which study the sensitivities of outputs by perturbing inputs. Furthermore, MACQ allows us to separate marginal attribution of individual features from interaction effect, and visually illustrate the 3-way relationship between marginal attribution, output level, and feature value.

This paper aims to define, quantify, and analyze the feature complexity that is learned by a DNN. We propose a generic definition for the feature complexity. Given the feature of a certain layer in the DNN, our method disentangles feature components of different complexity orders from the feature. We further design a set of metrics to evaluate the reliability, the effectiveness, and the significance of over-fitting of these feature components. Furthermore, we successfully discover a close relationship between the feature complexity and the performance of DNNs. As a generic mathematical tool, the feature complexity and the proposed metrics can also be used to analyze the success of network compression and knowledge distillation.

Electrical network frequency (ENF) is the signature of a power distribution grid which represents the nominal frequency (50 or 60 Hz) of a power system network. Due to load variations in a power grid, ENF sequences experience fluctuations. These ENF variations are inherently located in a multimedia signal which is recorded close to the grid or directly from the mains power line. Therefore, a multimedia recording can be localized by analyzing the ENF sequences of that signal in absence of the concurrent power signal. In this paper, a novel approach to analyze location forensics using ENF sequences extracted from a number of power and audio recordings is proposed. The digital recordings are collected from different grid locations around the world. Potential feature components are determined from the ENF sequences. Then, a multi-class support vector machine (SVM) classification model is developed to validate the location authenticity of the recordings. The performance assessments affirm the efficacy of the presented work.

This paper proposes a set of criteria to evaluate the objectiveness of explanation methods of neural networks, which is crucial for the development of explainable AI, but it also presents significant challenges. The core challenge is that people usually cannot obtain ground-truth explanations of the neural network. To this end, we design four metrics to evaluate explanation results without ground-truth explanations. Our metrics can be broadly applied to nine benchmark methods of interpreting neural networks, which provides new insights of explanation methods. Nowadays, many methods are proposed to explain the feature representations of a deep neural network (DNN) in a post-hoc manner. However, some methods usually pursue attribution maps which look reasonable from the perspective of human users, instead of objectively reflecting the information processing in the DNN. A trustworthy evaluation of the objectiveness of attribution maps is crucial for the development of deep learning and proposes significant challenges to state-of-the-art algorithms. Existing metrics (Y ang & Kim, 2019; Arras et al., 2019) of evaluating explanation methods have certain shortcomings. Issue 1, evaluation of the accuracy of a DNN null evaluation of the objectiveness of attribution maps: Some methods only evaluate whether the visualized attribution map looks reasonable to human users, instead of examining whether an attribution map objectively reflects the truth of a DNN. For example, they added an irrelevant object into the image. Pixels from the irrelevant object are expected to be assigned with zero attributions.

This paper aims to analyze knowledge isomorphism between pre-trained deep neural networks. We propose a generic definition for knowledge isomorphism between neural networks at different fuzziness levels, and design a task-agnostic and model-agnostic method to disentangle and quantify isomorphic features from intermediate layers of a neural network. As a generic tool, our method can be broadly used for different applications. In preliminary experiments, we have used knowledge isomorphism as a tool to diagnose feature representations of neural networks. Knowledge isomorphism provides new insights to explain the success of existing deep-learning techniques, such as knowledge distillation and network compression. More crucially, it has been shown that knowledge isomorphism can also be used to refine pre-trained networks and boost performance.