Understanding neural networks via rule extraction

Neural Processing Letters, 1998

Three neural-based methods for extraction of logical rules from data are presented. These methods facilitate conversion of graded response neural networks into networks performing logical functions. MLP2LN method tries to convert a standard MLP into a network performing logical operations (LN). C-MLP2LN is a constructive algorithm creating such MLP networks. Logical interpretation is assured by adding constraints to the cost function, forcing the weights to ±1 or 0. Skeletal networks emerge ensuring that a minimal number of logical rules are found. In both methods rules covering many training examples are generated before more specific rules covering exceptions. The third method, FSM2LN, is based on the probability density estimation. Several examples of performance of these methods are presented. I. INTRODUCTION Classification using crisp logical rules is preferable to humans over other methods because it exposes the inherent logical structure of the problem. Although the class of problems with logical structure simple enough to be manageable by humans may be rather limited nevertheless it covers some important applications, such as the decision support systems in financial institutions. One way to obtain logical description of the data is to analyze neural networks trained on these data. Many methods for extraction of logical rules from neural networks exist (for a review and extensive references see [?]). In this paper several new methods of logical rule extraction and feature selection are presented. Although we concentrate on crisp logical rules these methods can easily obtain also fuzzy rules. In contrast with the existing neural rule extraction algorithms based on analysis of small connection weights, analysis of sensitivity to changes in input or analysis Authors are with the Department of Computer Methods, Nicholas

Nonlinear Analysis: Theory, Methods & …, 1997

Proceedings Fourth International Conference on Computational Intelligence and Multimedia Applications. ICCIMA 2001, 2001

Although Artificial Neural Networks (ANNs) have satisfactory emploied in several problems, such as clustering, pattern recognition, dynamic systems control and prediction, they still suffer of a significant limitation. The representations they learn are not usually comprehensible to humans. Several techniques have been suggested to extract meaninful knowledge from trained ANNs. This paper proposes the use of symbolic learning to extract symbolic representations from trained ANNs. These algorithms are commonly used by the Machine Learning community. They learn concepts represented by propositional description. The procedure proposed is similar to that used by the Trepan algorithm (Craven, 1996). This algorithm extract comprehensible, symbolic representations (decision trees) from trained ANNs. Trepan queries a trained network to induce a decision tree that describes the concept represented by the ANN.

Knowledge and Information Systems, 2020

Representing the knowledge learned by neural networks in the form of interpretable rules is a prudent technique to justify the decisions made by neural networks. Heretofore many algorithms exist to extract symbolic rules from neural networks, but among them, a few extract rules from deep neural networks trained using deep learning techniques. So, this paper proposes an algorithm to extract rules from a multi-hidden layer neural network, pre-trained using deep belief network and fine-tuned using back propagation. The algorithm analyzes each node of a layer and extracts knowledge from each layer separately. The process of knowledge extraction from the first hidden layer is different from the other layers. Consecutively, the algorithm combines all the knowledge extracted and refines them to construct a final ruleset consisting of symbolic rules. The algorithm further subdivides the subspace of a rule in the ruleset if it satisfies certain conditions. Results show that the algorithm extracted rules with higher accuracy compared to some existing rule extraction algorithms. Other than accuracy, the efficacy of the extracted rules is also validated with fidelity and various other performance measures.

2000

A common problem in KDD (Knowledge Discovery in Databases) is the presence of noise in the data being mined. Neural networks are robust and have a good tolerance to noise, which makes them suitable for mining very noisy data. However, they have the well-known disadvantage of not discovering any high-level rule that can be used as a support for human decision making. In this work we present a method for extracting accurate, comprehensible rules from neural networks. The proposed method uses a genetic algorithm to find a good neural network topology. This topology is then passed to a rule extraction algorithm, and the quality of the extracted rules is then fed back to the genetic algorithm. The proposed system is evaluated on three public-domain data sets and the results show that the approach is valid

2000

Abstract For neural networks to gain a greater acceptance among machine learning practitioners dealing with real world data, some form of explanation of how a network reaches its decision must be reachable. Recent neural network research has shown that making use of ensembles of neural networks, more accurate predictions can be achieved. Although any advances that reduce this error are welcome, ensembles have added an extra layer of complexity that must be dealt with when extracting rules from the networks.

International Conference on Pattern Recognition, Informatics and Medical Engineering (PRIME-2012), 2012

Though neural networks have achieved highest classification accuracy for many classification problems, the obtained results may not be interpretable as they are often considered as black box. To overcome this drawback researchers have developed many rule extraction algorithms. This paper has discussed on various rule extraction algorithms based on three different rule extraction approaches namely decompositional, pedagogical and eclectic. Also it evaluates the performance of those approaches by comparing different algorithms with these three approaches on three real datasets namely Wisconsin breast cancer, Pima Indian diabetes and Iris plants.

IEEE Transactions on Knowledge and Data Engineering, 1999

Hybrid Intelligent Systems that combine knowledge-based and artificial neural network systems typically have four phases involving domain knowledge representation, mapping of this knowledge into an initial connectionist architecture, network training, and rule extraction, respectively. The final phase is important because it can provide a trained connectionist architecture with explanation power and validate its output decisions. Moreover, it can be used to refine and maintain the initial knowledge acquired from domain experts. In this paper, we present three rule-extraction techniques. The first technique extracts a set of binary rules from any type of neural network. The other two techniques are specific to feedforward networks, with a single hidden layer of sigmoidal units. Technique 2 extracts partial rules that represent the most important embedded knowledge with an adjustable level of detail, while the third technique provides a more comprehensive and universal approach. A rule-evaluation technique, which orders extracted rules based on three performance measures, is then proposed. The three techniques area applied to the iris and breast cancer data sets. The extracted rules are evaluated qualitatively and quantitatively, and are compared with those obtained by other approaches.

IJCNN'99. International Joint Conference on Neural Networks. Proceedings (Cat. No.99CH36339)

Rule extraction from classifiers treated as black boxes is an important topic in explainable artificial intelligence (XAI). It is concerned with finding rules that describe classifiers and that are understandable to humans, having the form of (I f ...T hen...Else). Neural network classifiers are one type of classifier where it is difficult to know how the inputs map to the decision. This paper presents a technique to extract rules from a neural network where the feature space is Boolean, without looking at the inner structure of the network. For such a network with a small feature space, a Boolean function describing it can be directly calculated, whilst for a network with a larger feature space, a sampling method is described to produce rule-based approximations to the behaviour of the network with varying granularity, leading to XAI. The technique is experimentally assessed on a dataset of cross-site scripting (XSS) attacks, and proves to give very high accuracy and precision, comparable to that given by the neural network being approximated.

Algorithms, 2021

In machine learning, ensembles of models based on Multi-Layer Perceptrons (MLPs) or decision trees are considered successful models. However, explaining their responses is a complex problem that requires the creation of new methods of interpretation. A natural way to explain the classifications of the models is to transform them into propositional rules. In this work, we focus on random forests and gradient-boosted trees. Specifically, these models are converted into an ensemble of interpretable MLPs from which propositional rules are produced. The rule extraction method presented here allows one to precisely locate the discriminating hyperplanes that constitute the antecedents of the rules. In experiments based on eight classification problems, we compared our rule extraction technique to “Skope-Rules” and other state-of-the-art techniques. Experiments were performed with ten-fold cross-validation trials, with propositional rules that were also generated from ensembles of interpret...

Encyclopedia of Knowledge Management

Neural networks (NN) as classifier systems have shown great promise in many problem domains in empirical studies over the past two decades. Using case classification accuracy as the criteria, neural networks have typically outperformed traditional parametric techniques (e.g., discriminant analysis, logistic regression) as well as other non-parametric approaches (e.g., various inductive learning systems such as ID3, C4.5, CART, etc.).

International journal of neural systems, 2001

The problem of rule extraction from neural networks is NP-hard. This work presents a new technique to extract "if-then-else" rules from ensembles of DIMLP neural networks. Rules are extracted in polynomial time with respect to the dimensionality of the problem, the number of examples, and the size of the resulting network. Further, the degree of matching between extracted rules and neural network responses is 100%. Ensembles of DIMLP networks were trained on four data sets in the public domain. Extracted rules were on average significantly more accurate than those extracted from C4.5 decision trees.

The 1st Online Workshop on …, 1996

Simple method for extraction of logical rules from neural networks trained with backpropagation algorithm is presented. Logical interpretation is assured by adding an additional term to the cost function, forcing the weight values to be ±1 or zero. Auxiliary constraint ensures that the training process strives to a network with maximal number of zero weights, which augmented by weight pruning yields a minimal number of logical rules extracted by means of weights analysis. Rules are generated consecutively, from most general, covering many training examples, to most specific, covering a few or even single cases. If there are any exceptions to these rules, they are being detected by additional neurons.

Pattern Recognition, 1999

In this paper we present a methodology for extracting decision trees from input data generated from trained neural networks instead of doing it directly from the data. A genetic algorithm is used to query the trained network and extract prototypes. A prototype selection mechanism is then used to select a subset of the prototypes. Finally, a standard induction method like ID3 or C5.0 is used to extract the decision tree. The extracted decision trees can be used to understand the working of the neural network besides performing classi"cation. This method is able to extract di!erent decision trees of high accuracy and comprehensibility from the trained neural network.

2010

This paper describes an efficient rule generation algorithm, called rule generation from artificial neural networks (RGANN) to generate symbolic rules from ANNs. Classification rules are sought in many areas from automatic knowledge acquisition to data mining and ANN rule extraction. This is because classification rules possess some attractive features. They are explicit, understandable and verifiable by domain experts, and can be modified, extended and passed on as modular knowledge. A standard three-layer feedforward ANN is the basis of the algorithm. A four-phase training algorithm is proposed for backpropagation learning. Comparing them to the symbolic rules generated by other methods supports explicitness of the generated rules. Generated rules are comparable with other methods in terms of number of rules, average number of conditions for a rule, and predictive accuracy. Extensive experimental studies on several benchmarks classification problems, including breast cancer, wine, season, golf-playing, and lenses classification demonstrate the effectiveness of the proposed approach with good generalization ability.