ﺑﻬﺒﻮد ﺗﺸﺨﯿﺺ ﺑﺮونﻫﺸﺘﻪاي دادهﻫﺎ ﺑﺎ ﯾﺎدﮔﯿﺮي ﻋﻤﯿﻖ

Improve Anomaly Detection with Deep Learning

دﻟﯿﻞ اﺻﻠﯽ ﮐﻪ ﺑﺎﻋﺚ ﺷﺪ دادهﮐﺎوي، ﻣﻮرد ﺗﻮﺟﻪ ﺻﻨﻌﺖ اﻃﻼﻋﺎت ﻗﺮار ﺑﮕﯿﺮد، ﻣﺴﺌﻠﻪ در دﺳﺘﺮس ﺑﻮدن ﺣﺠﻢ وﺳﯿﻌﯽ از دادهﻫﺎ و اﺳﺘﺨﺮاج اﻃﻼﻋﺎت و داﻧﺶ ﺳﻮدﻣﻨﺪ از آنﻫﺎ اﺳﺖ. در ﻋﻤﻠﯿﺎت ﭘﺎكﺳﺎزي داده، ﻣﺸﮑﻞ ﮐﯿﻔﯿﺖ دادهﻫﺎ ﺑﺮﻃﺮف ﻣﯽﺷﻮد. ﯾﮑﯽ از ﻣﺸﮑﻼﺗﯽ ﮐﻪ ﺑﺮ ﮐﯿﻔﯿﺖ دادهﻫﺎ ﺗﺄﺛﯿﺮ ﻣﯽﮔﺬارد، دادهﻫﺎي ﺑﺮونﻫﺸﺘﻪ ﻫﺴﺘﻨﺪ. اﯾﻦ ﻧﻤﻮﻧﻪﻫﺎ رﮐﻮردﻫﺎﯾﯽ ﻫﺴﺘﻨﺪ ﮐﻪ ﻣﻘﺎدﯾﺮ ﻣﺸﺨﺼﻪ آنﻫﺎ ﺑﺎ رﮐﻮردﻫﺎي دﯾﮕﺮ ﺑﺴﯿﺎر ﺗﻔﺎوت دارد. در اﯾﻦ ﺗﺤﻘﯿﻖ از ﯾﮏ روش ﻣﺒﺘﻨﯽ ﺑﺮ ﯾﺎدﮔﯿﺮي ﻋﻤﯿﻖ و ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻋﻤﯿﻖ 14 ﻻﯾﻪاي ﺑﺮ روي ﭘﮑﯿﺞ ﺗﻨﺴﻮرﻓﻠﻮ و ﮐﺮاس ﺑﺮاي ﺗﺸﺨﯿﺺ ﺑﺮونﻫﺸﺘﻪاي و ﺑﻬﺒﻮد ﻋﻤﻠﮑﺮد آن اﺳﺘﻔﺎده ﺷﺪه اﺳﺖ. ﻣﺠﻤﻮﻋﻪ داده ﻣﻮرد اﺳﺘﻔﺎده در اﯾﻦ ﺗﺤﻘﯿﻖ ﻣﺠﻤﻮﻋﻪاي ﺑﺎ 2 درﺻﺪ ﺑﺮونﻫﺸﺘﻪاي اﺳﺖ. ﻣﯿﺰان ﺻﺤﺖ روش ﭘﯿﺸﻨﻬﺎدي ﻣﻘﺪار 97/08 را ﻧﺸﺎن داد و ﻣﻌﯿﺎرﻫﺎي ﺑﺎزﺧﻮاﻧﯽ و دﻗﺖ ﻧﯿﺰ 97 درﺻﺪ ﻣﺤﺎﺳﺒﻪ ﺷﺪه اﺳﺖ. روش ﭘﯿﺸﻨﻬﺎدي ﺑﺎ 5 ﻣﺪل دﯾﮕﺮ ﻣﺒﺘﻨﯽ ﺑﺮ ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﮐﺎﻧﻮﻟﻮﺷﻦ و ﺷﺒﮑﻪ ﺑﺎزﮔﺸﺘﯽ LSTM ﻧﯿﺰ ﻣﻘﺎﯾﺴﻪ ﺷﺪﻧﺪ. ﻣﻘﺪار ﻣﻌﯿﺎرﻫﺎي ارزﯾﺎﺑﯽ ﮐﻼسﺑﻨﺪﻫﺎ ﻧﺸﺎن از ﺑﻬﺒﻮد ﺑﺴﯿﺎر ﺧﻮب روش ﭘﯿﺸﻨﻬﺎدي در ﻣﻘﺎﺑﻞ روشﻫﺎي ﺳﻨﺘﯽ و ﺣﺘﯽ روشﻫﺎي ﻣﺒﺘﻨﯽ ﺑﺮ ﯾﺎدﮔﯿﺮي ﻋﻤﯿﻖ را داده اﺳﺖ.

The main reason that data mining has become the focus of attention in the information industry is the availability of large volumes of data and the urgent need to extract useful information and knowledge from this data. In data cleaning operation, the problem of data quality is solved. One of the problems that affects the quality of data is skewed data or abnormal data. These are records whose attribute values are very different from other records. In this research, a method based on deep learning and 14-layer deep neural network on the tensorflow and cross package has been used to diagnose the abnormality and improve its performance. The data set used in this research is a data set with 2% anomalies. The accuracy of the proposed method was 97/08 and the readability and accuracy criteria were 97%. The proposed method was compared with 5 other models based on convolutional neural network and LSTM recursive network. The value of the classification evaluation criteria showed a very good improvement over the proposed method compared to traditional methods and even methods based on deep learning.