1. Introduction
Explainable Artificial Intelligence (XAI) is about opening the “black box” decision making of Machine Learning (ML) algorithms so that decisions are transparent and understandable. This ability to explain decision models is important to data scientists, end-users, company personnel, regulatory authorities, or indeed any stakeholder who has a valid remit to ask questions about the decision making of such systems. As a research area, XAI incorporates a suite of ML techniques that enables human users to understand, appropriately trust, and effectively manage the emerging generation of artificially intelligent partners [
1]. Interest in XAI research has been growing along with the capabilities and applications of modern AI systems. As AI makes its way to our daily lives, it becomes increasingly crucial for us to know how underlying opaque AI algorithms work. XAI has the potential to make AI models more trustworthy, compliant, performant, robust, and easier to develop. That can in turn widen the adoption of AI solutions and deliver greater business value.
A key development in the complexity of AI systems was the introduction of AlexNet deep model [
2], a Convolutional Neural Network (CNN) that utilises two Graphical Processing Units (GPUs) for the first time, enabling the training of a model on a very large training dataset whilst achieving state-of-the-art results. With 10 hidden layers in the network, AlexNet was a major leap in Deep Learning (DL), a branch of ML that produces complex multi-layer models that present particular challenges for explainability. Since AlexNet’s unveiling in 2012, other factors have boosted the rapid development of DL: Availability of big data, cloud computing growth, powerful embedded chips, reduction in the cost of systems with high computational power and memory, and the achievement of a higher performance of DL models over traditional approaches. In some application areas, these models achieve as good as human level performance, such as in object recognition [
3,
4], object detection [
5,
6], object tracking [
7]) or games (e.g., beating AlphaGo champion [
8]), stock price predictions [
9], time series forecasting [
10,
11], and health [
12].
In an early work in 2010, researchers in [
13] focused on explaining individual decisions taken by classifiers. From 2012, there has been a year on year improvement in the accuracy of deep learning models, accompanied by greater complexity. Researchers are actively investigating the real-life implications associated with the deployment of these types of models. In addition to ethical concerns, such as privacy or robot autonomy, there are other issues at the heart of ML that are critical to handle. For example, potentially biased decisions due to bias in the training data and the model; a system wrongly predicting/classifying an object with high confidence; lack of understanding of how a decision is taken or what input features were important in this decision; and downstream legal complications, such as the lack of adherence to the “right to explanation” under EU General Data Protection Regulation (GDPR) rule [
14]. For example, a customer whose loan application has been rejected has the right to know why their application was rejected.
Some models are used to make decisions that have life threatening implications, such as the interpretation of potential cancer scans in healthcare. Currently, a doctor is needed as an intermediate user of the system to take the final decision. Other AI scenarios aim to remove the intermediate user. For example, the use of fully autonomous cars would cede full control to the associated AI-based driving system. DL models are at the heart of these types of complex systems. Examples such as these emphasise the critical nature of explaining, understanding, and therefore controlling the decisions of DL models.
Explainability means different things, depending upon the user (audience/stakeholder) of the explanation and the particular concerns they wish to address via an explanation. For example, an end user (customer) may question the individual decision a model has taken about them. A regulatory authority may query whether the model is unbiased with respect to gender, ethnicity, and equality. An intermediate user, such as the doctor with the diagnostic scan decision, will want to know what features of the input have resulted in a particular decision.
Scope: In this paper, we have four main contributions. Firstly, we report recent research work for explaining AI models. We note that there are several comprehensive survey articles on XAI, such as: Tjoa, E., & Guan, C [
15] discuss explainability in health data, Selvaraju et al. [
16] & Margret et al. [
17] cover image data for object detection and recognition, and [
18] discuss financial data/text data. In addition, detailed surveys on XAI as a field are emerging, such as the detailed and comprehensive survey about explainability, interpretability, and understandability covered in [
19]. Secondly, we apply an explainability technique i.e., Layer-wise Relevance Propagation (LRP) for the explanation of a DL model trained over structured/tabular/mixed (in this paper structured, tabular, or mixed is used interchangeably) data as input, in this case a 1-dimensional DL model. Various research works use DL for time series data which is a special case of structured data, where time is a main feature. In our work, we focus on structured data that adheres to a predefined data model in a tabular format but without time features—i.e., non-time series structured data. To the best of our knowledge, this is the first time that LRP has been applied to a model with structured data input. LRP typically uses image as input, providing intuitive visual explanations on the input image.
In our work, we train a one dimensional CNN (1D-CNN) model and apply LRP in order to highlight influential features of the input structure data. This approach enables us to answer questions for our selected use case datasets such as: Which factors are causing customers to churn? Why did this specific customer leave? What aspects of this transaction deem it to be classified as fraudulent? There are several other explainability techniques typically used for image-based models e.g., DeepLIFT [
20], LEMNA [
21], and Grad-CAM [
16]. However, although there are several other perturbation approaches e.g., MAPLE [
22], LORE [
23], and L2X [
24] in this work we compare it with two commonly-used XAI techniques in the field: LIME and SHAP. Finally, we validate the correctness of the LRP explanations (important features) by our approach. This is done by taking the most influential subset of features and using them as input for training classifiers in order to see their performance i.e., to determine whether the new models are achieving equal or better performance on the subset of influence features highlighted in the explanation (testing set) compared to the models trained over the whole set of features.
The paper is organised as follows:
Section 2 gives an overview of related work.
Section 3 explains the proposed approach that includes the datasets used, pre-processing performed, models trained, and finally model explanation details. Then,
Section 4 discusses the results achieved with the proposed approach, highlighting important features, as well as, results with the subset of features. Finally,
Section 5 gives some future directions and main conclusions of our paper.
2. Related Work
Whilst explaining AI systems is not a new research topic, it is still in its early stages. Several survey articles have been published for the domain, including [
17,
19,
25,
26,
27,
28]. Of these, ref. [
19] is the most recent and complete, summarising all others into one. This survey of XAI includes details about the concepts, taxonomies, and research work up to December 2019 along with opportunities and challenges for new or future researchers in the field of XAI. Arrieta et al. [
19] divide the taxonomy of explainable scenarios into two main categories: (1) Transparent ML models that are self-explanatory or answers all or some of the previous questions (e.g., Linear/Logistic regression, Decision trees, K-NN, rule-based learning, and general additive models) and (2) post-hoc models, where a new method is proposed to explain the model for explanation of a decision of a shallow or deep model. The post-hoc category is further divided into model-agnostic, which can be applied to all models to extract specific information about the decision and model-specific, which are specific to the model in use e.g., for SVM, or DL models such as CNN.
In contrasts to Arrieta’s transparent model view, Mittelstadt et al. [
29] give credence to the black box concept by highlighting an alternative point of view about the explanation of AI systems and whether AI scientists can gain more through considering broader concepts. In this work, they focus on ‘what-if questions’, and highlight that decisions of a black box system must be justified and open to discussion and questioning. In [
30], emphasis is put on bringing transparency and trust in AI systems by taking care of issues such as the ‘Clever Hans’ problem [
31] and providing some level of explanation for decisions being made. The authors categorise explanations based on the content (e.g., explaining learned representations, individual predictions, model behaviour, and representative examples) and their methods (e.g., explaining with surrogates, local perturbations, propagation-based approaches, and meta-explanations).
Explainability of DLs for structured data is limited. In the majority of cases, traditional ML techniques such as random forest, XGboost, SVM, logistic regression, etc. are used with explainability techniques LIME [
32], SHAP [
33], or more recently MANE [
34] that is being used with CNN. These methods for explaining predictions from ML algorithms have become well established in the past few years. It is important to highlight that the majority of the XAI methods, which use DL networks such as CNN, show heatmaps [
31] or saliency visualisations [
35] for images input to the network. These techniques are also applied to other types of input data apart from images, including text [
36] and time series data [
10]. However, some of the techniques in these XAI are not general in the sense that they cannot be applied to different ML algorithms and/or input types or both. Hence, here we will discuss briefly the explainability of approaches used for DL models in three main categories of input data i.e., images, text, and time series data. We explain these application of explainability for various DL model inputs to frame our work—but we note the lack of application of such techniques for DL models using tabular (non-time series) data.
XAI in Images: A well-explored area of research in XAI is proposing models (mainly using CNN [
31,
37,
38]) that can interpret and classify an input image. When such models are explained, they benefit from the intuitive visual nature of the input. The portion of the image that influenced the model decision can be highlighted, which is relatively easily understood by different types of recipients e.g., end-user (customer) or data scientists. For example, researchers found Clever Hans [
31] type issues in datasets, which are highly interpretable for this issue [
31].
M. D. Zeiler and R. Fergus [
37,
38,
39] contributed approaches to understanding mid- and high-level features that a network learns as well as visualising the kernels and feature maps by proposing a deconvenet model to reconstruct strong and soft activations to highlight influences on the given prediction. In [
40], a local loss function was utilised with each convolution layer to learn specific features related to object components. These features result in more interpretable feature maps that support explainability. Google’s Model Cards tool [
17] helps to provide insight on trained image models, providing bench-marked evaluation information in a variety of conditions. The tool helps to answer concerns in explainability, such as the avoidance of bias. Such model cards can be used/considered for every model before deployment.
Ramprasaath et al. [
16] proposed a post-hoc method (proposing a new method to explain an existing model for explanation of its decision) that can be applied to several types of CNN models to visualise and explain the decision it provides. The model, termed Grad-CAM, uses a gradient weighted class activation mapping approach in which the gradient targets a class (e.g., cat) and visualises the activations that help in predicting the correct class. A pixel-level visualisation has been proposed in the form of a heatmap that shows where the model is focusing on an output map, and thus influenced the model decision.
Recently, Lapuschkin et al. [
31] explained the decisions of nonlinear ML model systems for Computer Vision (CV) and arcade games. They used LRP [
41] and Spectral Relevance Analysis (SpRAy) technique and compared both with Fisher Vector-based results to detect and highlight the Clever Hans issue in a famous dataset (PASCAL VOC). The proposed SpRAy uses spectral clustering on the heatmaps generated by LRP to identify typical and atypical behaviour in a semi-automated manner. This is done by learning some specific behaviours (anomalies) in the decisions of a system over a large dataset, unlike the LRP approach which manually analyses every output. These models helps in identifying serious issues in what a model learns e.g., a wrong area/patch of an image to correctly classify the category.
XAI in Time Series data: The analysis and forecasting of Time Series (TS) information, like any other area that can benefit from AI, needs to incorporate mechanisms that offer transparency and explainability of its results. However, in DL, the use of these mechanisms for a time series is not an easy task due to the temporal dependence of the data. For instance, surrogate solutions like LIME [
32] or SHAP [
33] do not consider a time ordering of the inputs so their use on TS presents clear limitations.
In [
42], authors propose a visualisation tool that works with CNN and allows different views and abstraction levels for a problem of prediction over Multivariate TS defined as a classification problem. The solution proposes the use of saliency maps to uncover the hidden nature of DL models applied to TS. This visualisation strategy helps to identify what parts of the input are responsible for a particular prediction. The idea is to compute the influence of the inputs on the inter-mediated layers of the neural network in two steps: Input influence and filter influence. The former is the influence of the input in the output of a particular filter and the latter is the influence of the filter on the final output based on the activation patterns. The method considers the use of a clustering stage of filters and optimisation of the input influence, everything with the goal of discovering the main sources of variations and to find similarities between patterns. However, due to clustering to combine the maps, it is time consuming and might not be as fast as other techniques such as LRP, which work on the pre-computed gradients.
ML tools are widely used in financial institutions. Due to regulatory reasons and ease of explainability, interpretability, and transparency many institutions use traditional approaches such as decision trees, random forest, regression, and Generalized Additive Model (GAM), at a cost of lower performance. However, there are examples of DL models that have been applied in financial applications e.g., for forecasting prices, stock, risk assessment, and insurance. Taking specific model examples, GAMs are relatively easy and transparent to understand and are used for risk assessments in financial applications [
43,
44,
45]. The authors in [
46] use traditional XGboost and Logistic Regression (LR), with LR principally used for comparison purposes. After training the model, the Shapley values [
33] from the testing set of the companies are calculated. The testing set contains explanatory variables values. They also use a post-processing phase correlation matrix to interpret the predictive output from a good ML model that provides both accuracy and explainability.
Liu et al. [
9] proposed a DL model to predict stock prices. In the first step of this work, a specific model was used to reduce the noise and make the data clean for LSTM. This system showed good results for predicting stock prices through price rate of change. In [
47], a decision support system from financial disclosures is proposed. It uses a deep LSTM model to predict whether the stock is going up or down. The authors have also focused on answering whether the DL model can give good results on short-term price movement compared to the traditional approach of the bag of words with logistic regression, SVM, etc. The results show that DL based systems, as well as transfer learning and word-embeddings, improve performance compared to naive models. Whilst the performance of these models is not very high, the approach gives a baseline for future research to using DL in financial data.
In [
48], an AI-based stock market prediction model for financial trade called CLEAR-Trade is proposed, which is based on CLEAR (Class Enhanced Attentive Response). It identifies the regions with high importance/activations and their impact on the decision as well as the categories that are highly related to the important activations. The objective is to visualise the decisions for better interpretability. The results on using S&P 500 Stock Index data show that the model can give helpful information about the decision made which can help a company while adopting AI-based systems for addressing requirements from regulatory authorities. Their model uses a CNN architecture with a convolution layer, leaky ReLu, and Global average pooling layer, followed by the SoftMax layer to classify into two categories i.e., the market going up or down. The visualisation shows that in the correct cases, the model weighs the past 4 days of data heavily, whereas in the incorrect cases, it considers data from previous weeks as important. Secondly, in the correct decisions, it considers open, high, and low values for making a decision. Whereas in the incorrect cases, the model considers trade volumes but it is not a strong indicator of correctly predicting the model future. Thirdly, it can/may show that in the correct cases, the probability or output values are high compared to when the model incorrectly predicts.
XAI in Text data: DL has shown good performance over text data for application areas such as text classification, text generation, natural language processing for chat-bots, etc. Similar to vision, financial, and time-series data, several works have been done on text data to explain what and how the text is classified or sentence is generated [
36,
49]. A bi-LSTM is used to classify each sentence in five classes of sentiment. LRP is used to visualise the important word in the sentence. The LRP relevance values are being examined qualitatively and quantitatively. It showed better results than a gradient-based approach.
Summary: XAI is a highly active area of research in the machine learning domain, with a variety of general and model/data specific approaches in the field and continuing to emerge. We have discussed the most relevant explainability approaches related to deep learning models processing images, time-series/financial data, and text input. We note the lack of deep learning XAI approaches applied to structured tabular data. Structured, tabular data is very common in organisations, tending to be an earlier focus for the adoption of machine learning models than unstructured data such as images or text. Explainability of structured data models has been largely limited to those based on traditional machine learning models (with algorithms such as random forest, XGBoost, etc.) using model agnostic techniques such as LIME and SHAP. Organisations want to utilise such data for training a DL network, provided such DL models can be explained.
We focus principally on LRP, an established XAI technique for DL models that is typically used for images but can be utilised with modifications for other forms of inputs, providing intuitive visual explanations in the form of heatmaps. It has a number of distinct advantages: It provides intuitive visual explanation highlighting relevant input, produces results quickly, and has not been tried with 1D CNN over structured data. By visually highlighting high influence parts of the input, it should in theory highlight the important features (input) that contribute most to a model decision e.g., customer churn, credit card fraud detection, and loan or insurance rejection. 1D CNN is never or rarely (not in our knowledge to date) used for structured data but we suggest that it can be, with the sliding kernel approach, learning a combination of features in the table that as a group contribute to model decisions.
Our motivation for using the traditionally image focused approach of 1D-CNN for tabular data was as follows: Firstly, structured data has a large overlap with image input. It is essentially a matrix of numbers, just as an image is a matrix of pixel values numbers. Pixel values have a fixed range, and this can be achieved in structure data using normalisation. In the case of structured data as input to a 1D-CNN, the positions and combinations of numbers has relevance and are in a fixed set of positions (features). Furthermore, although we do not know whether certain features have correlation or dependencies with each other, CNN will learn that patterns/dependencies/uniqueness that drive to particular classifications, individually, or in combination to other features by identifying the occurrence of feature values. Secondly, in traditional machine learning, features are typically selected manually in the feature extraction stage or by using techniques like LBP [
50], SIFT [
51], etc., and than in some cases a features subset selection technique is used to improve the model results. This can be a lengthy iterative process e.g., manual subset selection of features or by using a wrapper feature selection method, with iterative model training to seek out redundant low contribution features. By using 1D-CNN with LRP for explanation, the influential features are highlighted as a by-product of the initial model creation exercise. Our focus is on using and enhancing existing XAI techniques for structured data. In addition to use a 1D-CNN model over structured data with LRP for model explanation, we wish to compare the correctness of LRP against leading explainability methods SHAP and LIME in terms of their similarity in selecting important features and time complexity. In the next section, we will discuss the proposed approach.