Scribd
Upload
109 views5 pages

A Novel Approach For Code Smells Detection Based On Deep Learning

This document proposes a novel approach for detecting code smells using deep learning. Specifically, it uses convolutional neural networks to identify various types of code smells based on code semantics. The experiments show the approach achieves high F2 scores, particularly for detecting uncontrolled side effects and contrived complexity. Key aspects of the approach include transforming source code to XML, using CNNs to propose and classify code segments, and achieving precision, recall, and F2 scores over 0.75 on average for several common code smells.

Uploaded by

4048 Sivashalini.G
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
109 views5 pages

A Novel Approach For Code Smells Detection Based On Deep Learning

This document proposes a novel approach for detecting code smells using deep learning. Specifically, it uses convolutional neural networks to identify various types of code smells based on code semantics. The experiments show the approach achieves high F2 scores, particularly for detecting uncontrolled side effects and contrived complexity. Key aspects of the approach include transforming source code to XML, using CNNs to propose and classify code segments, and achieving precision, recall, and F2 scores over 0.75 on average for several common code smells.

Uploaded by

4048 Sivashalini.G
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

A Novel Approach for Code Smells Detection Based on

Deep Leaning

Tao Lin1, Xue Fu2, Fu Chen3, Luqun Li4


1 Amazon, Seattle WA 98109, USA, [email protected]
2 Shanghai Normal University, China
3 Central University of Finance and Economics, China, [email protected]
4 Shanghai Normal University, China, [email protected]

Abstract. Compared to software bugs, code smells are more significant in


software engineering research. It is not easy to detect code smells through
traditional methods. In this work, we propose a novel code smells detection
approach based on deep learning. The experiments show that our work achieves
high scores in terms of F2 score.

Keywords: Code Smells, Deep Learning, Convolutional Neural Network.

1 Introduction

Code smells are increasingly generated by modern agile software development. This
is because code changes are much more frequent and occur on a daily basis for large
software companies and dominant open-source communities.
Although there are many more test approaches to detect code smells, these methods
have some defects. Due to the frequent changes, it is increasing probable to generate
code smells overheads. Code smells, like software bugs, are a serious problem in
modern software.
Nowadays, the code smells are being researched by many practitioners. Software
developers are not aware of what is the code smell, although they are aware of software
bugs, thanks integrated environment development kits that can provide many instant
suggestions and notifications when there are bugs.
The question here is how to detect code smells effectively? And what is the
motivation for detecting code smells? Although there are many more test approaches
to detect code smells, these methods have some defects. There are mainly two
categories of deep learning networks. One is Recurrent Neural Networks, and another
is Convolutional networks.

Corresponding Authors: Tao Lin (Amazon, USA, [email protected]), Luqun Li (Shanghai Normal
University, China, [email protected]), Fu Chen (Central University of Finance and
Economics, China, [email protected]), Xue Fu (Shanghai Normal University, China)
2

Convolutional networks have already demonstrated its usage by leveraging


hierarchy features. In this paper, we use fully convolutional networks for code smells
detection based on semantic features. We will use fully convolutional networks for this
work.
We will define the type of neural network we use, and explain how it is used to detect
code smells. An advantage of using convolutional network is its ability to identify and
use local correspondences.
In recognition and machine learning, convolutional networks are increasingly
significant. Convolutional network presents the improvement on image recognition. An
example is using convolutional network on local correspondence. In software
engineering, we definitely can use these kinds of information for code smells detection.
To our knowledge, this is the first work to train a convolutional network for code
smells recognition. The inference is much more improved through convolutional
network.
This work is an extension of the authors previous work [8].
Unlike previous works that needs additional information for code smells detection,
this work does not use any existing information for code smells detection. One of the
major challenges in code smells detection is to find the relationship between code
semantics and code location. There is a tradeoff between identifying the correct
semantics compared to identifying the correct location of the smells.
Although there are several success stories from image recognition by using deep
networks [1].It is hard to transfer these approaches to software engineering, which is
more deterministic. Fully convolutional network has been used for one-layered
computation, and has a potential to be deployed to multi-layered environments.

2 Code Smells Detection Based on Convolutional Networks

We can define a multi-dimensional array to represent the convolutional network, h


* w * d, where h and w are space dimensions, and d is the channel. The first layer is
our source code inputs.
The second layer is the networks for sequence modeling. For example, the inputs are
x0, x1, x2, x3, x4,… xn, and the outputs are y0, y1, y2, y3, y4,… yn. The second layer will
be y’0, y’1, y’2, y’3, y’4,… y’n.
The outputs will be reshaped to a one-dimensional array, where size will be D *1024.
This output array will be dilation blocks. For the encoder task, we should process noise.
Each layer in the encoder is processed by normalization and liner analysis.

3 High level design

With recent development in software engineering, it is easy to find software bugs using
several methods from compiling to running. Our work is based on the state-of-the-art
deep learning methods for detection and recognition task.
3

Firstly, we transform the software source code to XML file, in order to be processed
by deep learning models [2]. Then there are two steps: code segment proposal and
classification. The code segment proposal leverages the heuristic search to generate
following inputs. These segments are processed by CNN classifier. We try to avoid to
use R-CNN, otherwise. One of the main reasons is that R-CNN uses selective search
algorithm, which is time consuming.
We use the following equation for segments:
𝑙
Seg = (max𝑟𝑝2 ⁄𝐷 + 𝑚𝑖𝑛𝑟𝑞) ∗ (𝐼/𝐷)

Seg is the segments, r is the rule of limits, and p is process variable, q is the next
graph inputs, I is interception, and D is next destination.

3.1 Experiments results

We use an open-source database published by the authors’ previous work[11].


The experiments results are shown as following table:

Table 1 Experiments results

Precision Recall F-Score Kappa


Long 0.528 0.674 0.754 0.635
Method
Lazy Class 0.624 0.678 0.613 0.632

Speculative 0.712 0.734 0.689 0.643


Generality
Refused 0.698 0.701 0.711 0.677
Bequest
Duplicated 0.543 0.568 0.594 0.585
code
Contrived 0.783 0.792 0.810 0.802
complexity
Shotgun 0.597 0.596 0.501 0.601
surgery
Uncontrolled 0.801 0.799 0.805 0.810
side effects

From Table 1, this work achieves high performance in terms of F2 score, especially
for the category of uncontrolled side effects and contrived complexity.
4

4 Conclusion

In this work, we conducted a research for code smells detection based on deep learning.
Our solution uses convolutional neural network for training a model to detect several
common code smells problems in software engineering. The solution achieves satisfied
F2 score with the average above 0.75.

Acknowledge

Part of this work is from the author’s PhD study [1], before the author joining
Amazon. Professor Fu Chen from Central University of Finance and Economics
provided many constructive suggestions and perspectives for this work during author’s
PhD study. Professor Fu Chen and this work was supported in part by National Science
Foundation of China under No.61672104.

Reference

[1] T. Lin, “A Data Triage Retrieval System for Cyber Security Operations Center,”
Pennsylvania State Univ. Thesis, 2018.
[2] T. Lin, “A Container - Destructor – Explorer Paradigm to Code Smells
Detection,” J. Chinese Comput. Syst., vol. 37, no. 3, 2016.
[3] T. Lin and X. Fu, “Flame Detection Based on SIFT Algorithm and One Class
Classifier with Undetermined Environment,” Comput. Sci., vol. 42, no. 6, 2015.
[4] T. Lin, C. Zhong, J. Yen, and P. Liu, “Retrieval of Relevant Historical Data
Triage Operations in Security Operation Centers,” in From Database to Cyber
Security, Springer, Cham, 2018, pp. 227–243.
[5] T. Lin, “A Novel Image Matching Algorithm Based on Graph Theory,”
Comput. Appl. Softw., vol. 33, no. 12, 2016.
[6] T. Lin, “Graphic User Interface Testing Based on Petri Net,” Appl. Res.
Comput., vol. 33, no. 3, 2016.
[7] T. Lin, “A Novel Direct Small World Network Model,” J. Shanghai Norm.
Univ., vol. 45, no. 5, 2016.
[8] T. Lin, J. Gao, X. Fu, and Y. Lin, “A Novel Bug Report Extraction Approach,”
in International Conference on Algorithms and Architectures for Parallel
Processing, 2015, pp. 771–780.
[9] C. Zhong, T. Lin, P. Liu, J. Yen, and K. Chen, “A cyber security data triage
operation retrieval system,” Comput. Secur., vol. 76, pp. 12–31, 2018.
[10] T. Lin, “Deep Learning for IoT,” 39th IEEE -- International Performance
Computing and Communications Conference, 2020.
[11] T.Lin, “Security Operations Center Retrieval,” 2021.
https://github.com/ltaocs/SecurityOperationsCenterRetrieval.
5

You might also like