Implications of deep learning for the automation of design patterns organization

Highlights

• There is a need to bridge the gap between the semantic relationship between patterns.

• We propose an approach by leveraging a powerful deep learning algorithm named Deep Belief Network (DBN).

• The DBN learns on the semantic representation of documents formulated in the form of feature vectors.

• We performed a case study in the context of a text categorization based automated system.

• The experimental promising results suggest the significance of the proposed approach to construct a more representative feature set.

Abstract

Though like other domains such as email filtering, web page classification, sentiment analysis, and author identification, the researchers have employed the text categorization approach to automate organization and selection of design patterns. However, there is a need to bridge the gap between the semantic relationship between design patterns (i.e. Documents) and the features which are used for the organization of design patterns. In this study, we propose an approach by leveraging a powerful deep learning algorithm named Deep Belief Network (DBN) which learns on the semantic representation of documents formulated in the form of feature vectors. We performed a case study in the context of a text categorization based automated system used for the classification and selection of software design patterns. In the case study, we focused on two main research objectives: 1) to empirically investigate the effect of feature sets constructed through the global filter-based feature selection methods besides the proposed approach, and 2) to evaluate the significant improvement in the classification decision (i.e. Pattern organization) of classifiers using the proposed approach. The adjustment of DBN parameters such as a number of hidden layers, nodes and iteration can aid a developer to construct a more illustrative feature set. The experimental promising results suggest the significance of the proposed approach to construct a more representative feature set and improve the classifier’s performance in terms of organization of design patterns.

Introduction

The text categorization approach has been applied in many domains to introduce an automated system to decrease the cost and computational time across the problems. The aim of automated systems is to classify the texts into appropriate classes with respect to their content, such as spam email filtering [17], web page classification [29], sentiment analysis [23], and author identification [43].

The steady increase in the number of design patterns in the literature and online repositories causes certain issues [2] such as (1) deep knowledge to understand the classification scheme used to organize the patterns [3], (2) the lack of classification scheme [11], (3) formal specification of patterns [19], [20], (4) the heterogeneous description of design patterns [38], and (5) the selections of right design patterns for inexperienced developer in the existence of complex [3] and spoiled patterns [4]. In order to address these issues, the text categorization approach has been employed to automate the organization and selection of design patterns [15], [16]. In the text categorization based automated system, it has been reported that the use of an extreme number of features may have undesirable effects on the classification accuracy and computation time of the automated systems [8], [14], [37]. The wrapper, embedded, and filter are three common schemes which are used to formulate the feature selection methods. The wrapper and embedded based feature selection methods required classifier interaction, while the filter-based methods do not need such interaction to construct a feature set. Besides, the running time of wrappers and embedded methods may be increased due to the classifiers interaction [36], [42]. The wrappers and embedded methods are specific to the learning model as compared to filter-based methods. Consequently, the use of filter-based methods are recommended for the automatic systems implemented via text categorization approach. Though in the context of text categorization based automated systems for the organization and selection of design pattern, the effect of global filter-based feature selection methods have been reported [8], [37]. However, the association of semantic relationship between design patterns (i.e. Document) and the constructed features (i.e. Attributes of feature space constructed through feature selection method) still need to be addressed.

Recently, researchers have adopted the use of deep learning algorithms to improve certain research tasks such as to improve fault localization [21], model the programming languages to suggest the code [40], to retrieve program’s structural information, and produce features from program source code for the prediction of defects [39], [41]. Subsequently, the applicability of unsupervised feature learning and deep learning in certain areas have been reported [1]. In this paper, we look the capacity of deep learning [13] and propose an approach by leveraging a powerful representation learning algorithm Deep Belief Network [12], to learn features from feature vectors extracted from the problem domain of design pattern collection [6], [9], [12]. The aim of the proposed approach is to fill the gap between the semantic relationship between documents(or Patterns) and the features which are used for the organization of design patterns. The Deep Belief Network (DBN) is a generative graphical model to learn a representation which can aid to reconstruct the training data with a high probability. Before applying DBN to learn features from patterns catalog, we perform some pre-processing tasks such as remove stopwords and words stemming to preserve the features (with their frequencies) in the feature vector form [14]. Subsequently, these feature vectors are used as input to the DBN algorithm. We performed certain experiments to evaluate the effectiveness of the proposed approach. Our main contributions in this paper are.

• We leverage the Deep Belief Network (DBN) algorithm and extract the features which are learned from the semantic relationship between design patterns of a collection in order to improve the performance of classifiers employed for the automation to organize the design patterns.

• We propose an evaluation model to evaluate the performance of classifiers in the context of organization of design patterns.

• We evaluate the capacity of the proposed approach in the construction of more illustrative feature sets as compared to existing filter-based feature selection methods, such as Document Frequency (DF) [44], Improved Gini Index [31], Correlation, Gain Ratio, Information Gain [22].

• We used three well-known classifiers [37] and three widely used design pattern collections [6], [9], [12] which have been used as a benchmark in different domains and evaluate the effectiveness of the proposed approach.

The rest of the paper is organized into eight sections. In Section 2, we summarized the related work regarding the classification of design pattern, feature selection methods, and the application of deep learning. In Section 3, we present the brief introduction of text categorization approach. In Section 4, we describe the working procedure of the proposed approach. In Section 5, we describe the experimental procedure and present a case study in the domain of classification of design patterns. In Section 6, we discussed the results of the presented case study. Subsequently, in Section 7, we summarized the evaluation results. Finally, in Sections 8 Threats to validity, 9 Conclusion, we discuss some threats to the validity and the conclusion of our work respectively.