Profit agent classification using feature selection eigenvector centrality

Chia sẻ: Nguyễn Thảo | Ngày: | Loại File: PDF | Số trang:11

Thêm vào BST

Báo xấu

14
lượt xem 1
download

Download Vui lòng tải xuống để xem tài liệu đầy đủ

In this paper we applied a feature selection based on graph method, graph method identifies the most important nodes that are interrelated with neighbors nodes.

Chủ đề:

Bình luận(0) Đăng nhập để gửi bình luận!

Lưu

Nội dung Text: Profit agent classification using feature selection eigenvector centrality

International Journal of Mechanical Engineering and Technology (IJMET) Volume 10, Issue 3, March 2019, pp. 603–613, Article ID: IJMET_10_03_062 Available online at http://www.iaeme.com/ijmet/issues.asp?JType=IJMET&VType=10&IType=3 ISSN Print: 0976-6340 and ISSN Online: 0976-6359 © IAEME Publication Scopus Indexed PROFIT AGENT CLASSIFICATION USING FEATURE SELECTION EIGENVECTOR CENTRALITY Zidni Nurrobi Agam Computer Science Department, Bina Nusantara University, Jakarta, Indonesia 11480 Sani M. Isa Computer Science Department, Bina Nusantara University, Jakarta, Indonesia 11480 Abstract Classification is a method that process related categories used to group data according to it are similarities. High dimensional data used in the classification process sometimes makes a classification process not optimize because there are huge amounts of otherwise meaningless data. in this paper, we try to classify profit agent from PT.XYZ and find the best feature that has a major impact to profit agent. Feature selection is one of the methods that can optimize the dataset for the classification process. in this paper we applied a feature selection based on graph method, graph method identifies the most important nodes that are interrelated with neighbors nodes. Eigenvector centrality is a method that estimates the importance of features to its neighbors, using Eigenvector centrality will ranking central nodes as candidate features that used for classification method and find the best feature for classifying Data Agent. Support Vector Machines (SVM) is a method that will be used whether the approach using Feature Selection with Eigenvalue Centrality will further optimize the accuracy of the classification. Keywords: Classification, Support Vector Machines, Feature Selection, Eigenvalue Centrality, Graph-based. Cite this Article: Zidni Nurrobi Agam and Sani M. Isa, Profit Agent Classification Using Feature Selection Eigenvector Centrality, International Journal of Mechanical Engineering and Technology (IJMET)10(3), 2019, pp. 603–613. http://www.iaeme.com/ijmet/issues.asp?JType=IJMET&VType=10&IType=3 http://www.iaeme.com/IJMET/index.asp 603 editor@iaeme.com
Zidni Nurrobi Agam and Sani M. Isa 1. INTRODUCTION In this era data is a very important commodity used in almost all existing technologies, data makes researchers examine more data in order to find hidden patterns that can be used as information. but with the increasing number of data, there are also many data that irrelevant and redundant dataset, making the quality of the data less good. Feature Selection is a method that selects a subset of variables from the input which can efficiently describe the input data while reducing effects from noise or irrelevant variables and still provide good prediction results [1]. Usually, feature selection operation both ranking and subset selection [2][3] to get most relational or most important value from a dataset. n described as total feature the goal of feature selection is to select the optimal feature I, so the optimal feature selection is I < n. With Feature Selection processing data will improve the overall prediction because optimal dataset that improves by feature selection. we applied feature selection for optimizing classification based on graph feature selection, this feature selection ranked feature based on Eigenvector Centrality. in graph theory, ECFS measures a node that has major impact on other nodes in the network. all nodes on the network are assigned relative scores based on the concept that nodes that have high value contribute more to the score of the node in question than equal connections to low-scoring nodes [4]. A high eigenvector score means that a node is connected to many nodes who themselves have high scores. so, relationship between feature (nodes) are measure by weight the connection between nodes. The problem from feature subset selection refers the task of identifying and selecting a useful subset of attributes to be used to represent patterns from a larger set of often mutually redundant, possibly irrelevant, attributes with different associated measurement costs and/or risks [5]. we try to find the most influential feature to predict the profit agent with ECFS. There are many studies that research about Eigenvector Centrality such as Nicholas J. Bryan and Ge Wang [6] , Nicholas J. Bryan and team research about how music with so many features can create pattern network between song and used to help describe patterns of musical influence in sample-based music suitable for musicological analysis. [7] To analyze rank influence feature between genre music with Eigenvector Centrality. and on 2016 Giorgio Roffo & Simione Melzi research about Feature ranking vie Eigenvector Centrality, in Giorgio Roffo & Simione Melzi research important feature by identifying the most important attribute into an arbitrary set of cues then mapping the problem to find where feature are the nodes by assessing the importance of nodes through some indicator of centrality. for building the graph and the weighted distance between nodes Giorgio Roffo & Simione Melzi use Fisher Criteria. The Goal of this paper is to applied Chi-Square and ECFS feature selection and compare both features with different dataset. Both Feature Selection test with HCC and Profit agent dataset, this test validates with K Fold Cross Validation feature selection to test the model’s ability then evaluated with confusion matrix to measure misclassification. Based on ECFS results we try to determine which attribute from profit agent that have a major impact on another attribute on profit agent dataset. 2. RESEARCH METHOD A discussion about Feature selection and ranking based on Graph network [8] method will be discussed in this paragraph. To build the graph first we have to define how to design and build the graph. http://www.iaeme.com/IJMET/index.asp 604 editor@iaeme.com
Profit Agent Classification Using Feature Selection Eigenvector Centrality 2.1. Graph Design Define the graph G = (V,E), V is a vertices corresponding one by one to each variable x, x is a set of features X = {x(1),x(2),.....x(n)}. E define as a (Weight) edges between nodes (features). To represent the graph that have relationship between edge. Define node (feature) into adjency matrix represent on binary matrix: (i) αij = ℓ(x ,x(j)) (1) When the graph not complex, the adjacency matrix are 0 all 1 or (there are no weights on the edges or multiple edges) and the diagonal entries are all [9] , but if the adjacency matrix has weighted on edges the adjacency matrix will not 0 and 1 but fill with weighted as figure 1. ( ) ( ) Figure 1 Matrix A with Weight and Matrix A no Weight A design graph is a part to weight the graph according to how good the relationship between two feature in the dataset. we apply Fisher linear discriminant [10] to find the mean and standard deviation, this methods find a linear combination of features which characterizes or separates two or more classes of objects or events. | | ( ) (2) Where: m = represent a mean. s = represent a variance. (standard deviation) Subscripts = the subscripts denote the two classes. After we measure the relationship between to class and we get the weight from fisher linear Discriminant, then we implement Eigenvector Centrality to rank and filter data from spearman correlation weight that generated from the relationship between 2 nodes (feature). For G: = (V, E) with |V| vertices let A = (av,t) adjacency matrix, the relative centrality score of vertex v can be defined as : ∑ ( ) ∑ (3) where M(v) is a set of the neighbors of and is a constant. With a small rearrangement this can be rewritten in vector notation as the eigenvector equation: (4) However, as we count longer and longer paths, this measure of accessibility converges to an index known as eigenvector centrality measure (EC). Example for node and adjacency matrix [9] described in figure 2 and Table 1. http://www.iaeme.com/IJMET/index.asp 605 editor@iaeme.com
Zidni Nurrobi Agam and Sani M. Isa Figure 2 Node Data Agent Company XYZ Table 1 Example adjacency Matrix Example Age Gender City Balance Total Trans Age 0 8.56 0 0 0 Gender 0 0 3.4 7.4 0 City 0 8.56 0 2.6 0 Balance 0 7.4 2.6 0 0 Total Trans 0 0 8.7 0 0 2.2. SVM Classification Method SVM Classification is classification analysis which fall into the category a supervised learning algorithm. Given a set of training examples, each marked as belonging to one or the other of two categories, Support vectro machine builds a model that assigns new examples to one category or the other, making it a non-probabilistic binary linear classifier. Its basic idea is to map data into a high dimensional space and find a separating hyperplane with the maximal margin. Given a training dataset of n points of the form: (⃗ ) (⃗⃗⃗⃗⃗ ) (5) where the are 1 or −1 [11], each point indicating the class to which the point ⃗⃗⃗⃗ belongs. Each ⃗⃗⃗⃗ is p-dimensional real vector. We want to find the "maximum-margin hyperplane" that divides the group of points ⃗⃗⃗⃗ for which = 1 from the group of points for which = - 1 which is defined so that the distance between the hyperplane and the nearest point from either group is maximized. http://www.iaeme.com/IJMET/index.asp 606 editor@iaeme.com
Profit Agent Classification Using Feature Selection Eigenvector Centrality Figure 3 SVM try to find best Hayperlane to split class -1 and + 1 2.3. Confusion Matrix To measure how to optimize SVM Classification used Eigenvector Centrality, we used a confusion matrix, confusion matrix is a table that is often used to describe the performance of a classification model (or "classifier") on a set of test data for which the true values are known. The table of confusion matrix will describe on figure 4 there is a TP (True Positives), TN (True Negatives), FP (False Positives) and FN (False Negatives). Figure 4 Confusion Matrix True Positive: We predicted yes (they have failure), and they do have the failure. False Positives: We predicted yes, but they don't actually have the failure. True Negatives: We predicted no, and they don't have the failure. False Negatives: We predicted no, but they actually do have the failure. 3. RESULT & ANALYSIS In this step we discuss result from compare between Eigenvector Centrality FS and Chi- Square FS, then we used Support Vector Machines Classification to calculate and compare accuracy between that feature selection. 3.1. Dataset Dataset PT. XYZ is chosen analysis feature selection Eigenvector Centrality with the scenario. first, we describe the dataset used for feature selection as how many features used for prediction, data agent is a categorical data collected from transaction data and have 6 attributes that used for analysis: http://www.iaeme.com/IJMET/index.asp 607 editor@iaeme.com
Zidni Nurrobi Agam and Sani M. Isa 1. Type Application Description: a device used by the agent to the transaction. Categorical: 1. EDC 2. Android 2. Age Description: Age agent PT. XYZ Categorical: 1. 23 Years and 29 Years 3. City Description: City based on Agent stay. Categorical: Convert every city with a number. 4. Balance Agent Description: Wallet Agent on PT.Xyz Categorical: 1. Rp.500.000 and Rp.2.000.000 5. Transaction Description: Transaction Agent every day. Categorical: every transaction agent /day 6. Joined Description: Duration Agent join with PT.Xyz. Categorical: Count per day Agent from join until now. 7. Gender Description: Agent Gender. Categorical: 1= L 0 = P 8. PulsaPrabayar, PLNPrabayar, TVBerbayar, PDAM, PLNPasca, Telpon, Speedy, BPJS, Cashin, Asuransi, Gopay, & TiketKereta. Description: Any transaction from Pt.Xyz application. Categorical: Count perday detail transaction agent. Number of data profit: 558 Number of profit: 441 3.2. FS Comparison Approach Is this step, we compare analysis about Eigenvector Centrality Feature Selection, we applied chi-square feature selection, chi-square is a numerical test that measures deviation considering the feature event is independent of the class value [12]. In this section, we applied to compare between eigenvector centrality and chi-square feature selection used SVM Classification to dataset Agent and HCC (hepatocellular carcinoma) survival public dataset have 49 attributes and 2 class (Survive, Not Survive). HCC is the most common type of primary liver cancer. Hepatocellular carcinoma occurs most often in people with chronic liver diseases, such as cirrhosis caused by hepatitis B or hepatitis C infection. The entire dataset is taken for our analysis containing 165 records and have numerical category characteristics. Result from the comparison between ECFS with data agent, data HCC and Chi-Square with data agent, data HCC will be analyzed whether there are significant differences between the two methods. The correlation between the attributes can influence the classification result[13] and Eliminating crucial features accidentally can reduce the classification result. Details from dataset describe on table 2 includes further details of considered datasets such as http://www.iaeme.com/IJMET/index.asp 608 editor@iaeme.com
Profit Agent Classification Using Feature Selection Eigenvector Centrality the number of samples and variables, number of classes and accuracy result from classification [14]. Table 2 Dataset used in the comparison of feature selection. for each dataset, the following detailed are reported. No Dataset Samples Variables Classes 1 Agent Profit 1000 19 2 2 HCC 165 49 2 3.3. Test Models & Performance Analysis This step try to test accuracy both of feature selection we used SVM classification and measure accuracy with confusion matrix. For Eigenvector Centrality dataset will split into data and . describe the feature that used for classification model and describe the label form dataset, profit label will give -1 and non profit willl give 1. The data will split into training and test then data then split into two class, profit class will assume -1 and nonprofit 1. Mean data will used for find the mutual information, linear descriminant used for measure mutual information between two clasess and find best mutual for process building the graph on Eigenvector Centrality. the Eigenvector Centrality will rank result by how strongly each node is conneceted to the other nodes. selected 10 best strongly central attribute (node) to predict the model with SVM, SVM model used for this classification is FITCSVM. This models will validate with 10 k fold validation and used random seed (1) to control random number generation for every feature selection. Table 3 shows the result obtained by comparing the accuracy from different datasets used for classification. HCC dataset starts from attribute 9 to attribute 45, after we try 1 iteration to 8 iteration the accuracy of chi-square decreases to 60 % and Agent Profit dataset start from dataset 10 to 19 attribute, after we try 1 iteration to 9 iteration the accuracy of chi-square decreases to 60 % so we don’t display the accuracy from 1-9 for HCC and 1-10 for Profit Agent. Best attribute found based on manual reduction to get the best accuracy from iteration. Table 3 Performance Analysis dataset HCC with Chi-Square and ECFS Chi Chi Chi Itteration ECFS Itteration ECFS Itteration ECFS Square Square Square 45 70,4412 71,4706 32 71,6176 67,9779 19 67,9044 71,6176 44 69,8529 69,8529 31 69,8897 69,1912 18 67,9412 68,6029 43 70,4779 69,8529 30 69,8529 69,1544 17 68,5294 69,1912 42 69,2279 69,8162 29 68,0515 67,9779 16 67,3529 66,7647 41 69,2279 68,6397 28 66,2132 67,2794 15 67,9779 66,2132 40 67,3897 66,8015 27 67,4632 69,7794 14 66,8382 66,8382 39 65,5515 64,9265 26 67,4265 69,7426 13 66,8015 67,3897 38 66,8015 66,7279 25 66,1765 68,5662 12 65,625 69,7794 37 65 64,9265 24 67,8676 69,7426 11 65,5147 71,0294 36 65,5515 64,8897 23 65,4412 69,8162 10 63,1618 68,6397 35 66,8015 63,0882 22 67,2426 69,1912 9 59,4853 67,3529 34 68,0515 67,9779 21 64,2279 69,8162 33 71,6176 67,9779 20 66,0294 71,0294 http://www.iaeme.com/IJMET/index.asp 609 editor@iaeme.com
Zidni Nurrobi Agam and Sani M. Isa Figure 5 Performance Analysis dataset HCC with Chi-Square and ECFS In the chi-square, the highest iteration was generated in 32,33 iteration and produce the highest accuracy of 71,6176%. while the highest ECFS iteration was generated in 19 iterations and produced the highest accuracy 71.6176%. Chi-Square succeeded in producing faster accuracy at 33 iterations while ECFS achieved maximum accuracy in the 19th iteration. both of these FS were tested on the number of datasets that had 45 attributes, then the test will be carried out on dataset agent profit that has the number of attributes 19 and showing on table 4 and figures 6. Table 4 Performance Analysis dataset Agent Profit with Chi-Square and ECFS Itteration ECFS Chi square 19 84,5838 84,7838 18 84,9838 84,8828 17 85,0848 85,7838 16 85,3859 86,3889 15 85,0848 86,7889 14 85,3859 86,8889 13 87,7889 87,5899 12 90,3838 87,5909 11 90,4848 87,8899 10 83,7869 88,6899 9 67,6869 89,1899 http://www.iaeme.com/IJMET/index.asp 610 editor@iaeme.com
Profit Agent Classification Using Feature Selection Eigenvector Centrality Figure 6 Performance Analysis dataset Agent Profit with Chi-Square and ECFS The Accuracy produced by ECFS on the Agent profit dataset is 90.48% better than chi- square which produces a maximum accuracy 89,18% and iteration for maximum accuracy is obtained by ECFS in 11 iterations while chi-square is in iteration 9. the overall performance for both feature selection indicates that performance from ECFS is more robust than chi- square because on the results obtained from test HCC dataset and Agent Profit Dataset, but Chi-Square better when attribute is more than 20 attributes when chi-square can reach maximum accuracy on 32 and 33 iteration. but when chi-square and ECFS reach less than 20 attributes from HCC Dataset, ECFS succeeded reach maximum accuracy on iteration 19. When the iteration from attributes reach smaller iteration such as 9 attributes, Chi-Square shows that accuracy decreases significantly as in dataset HCC and agent profit. ECFS is more robust when the attribute reduced even the accuracy decreases ECFS not significantly decreases. Value of ECFS has the largest increase in performance when the attribute reduces [15] on dataset agent profit as many attributes have strong relationships with others. Every attribute on ECFS rank according to how well they descriminant between two class and reduces the attribute that doesn’t have a major impact on other attribute make the result from reduces attribute better. 3.4. Evaluating & Analysis Attribute From this comparison both feature selection has each best accuracy and evaluated with Confusion Matrix that shown in table 5, Confusion matrix only show 1 best accuracy from ECFS and Chi-Square. For ECFS best accuracy obtained from 10fold Dataset, HCC is 71,61% and Dataset Profit Agent is 90,48%. For Chi-Square best accuracy obtained from 10fold Dataset, HCC is 71,61% and Dataset Profit Agent is 89,18%. http://www.iaeme.com/IJMET/index.asp 611 editor@iaeme.com
Zidni Nurrobi Agam and Sani M. Isa Table 5 Confussion Matrix Best Accuracy ECFS (Dataset Agent Profit) Profit NonProfit Profit 40 4 NonProfit 5 51 Chi-Square (Dataset Agent Profit) Profit NonProfit Profit 34 10 NonProfit 1 55 ECFS (Dataset HCC) Profit NonProfit Profit 2 4 NonProfit 1 9 Chi-Square (Dataset HCC) Profit NonProfit Profit 2 4 NonProfit 1 9 There are 12 attributes from dataset Agent that have a major impact on other attributes after tested by ECFS and may have an impact for analysis profit agent as shown in figure 7. Figure 7 Attributes Analysis Dataset Profit Agent with ECFS From figure 7 we analysis attributes that have a major impact for other attribute is City and many profit agents are determined by where is agent doing the transaction. That is possible analysis from ECFS because the city is one attribute that has a major impact in real conditions because people in big cities and small cities have different habits such as knowledge about technology and culture. http://www.iaeme.com/IJMET/index.asp 612 editor@iaeme.com
Profit Agent Classification Using Feature Selection Eigenvector Centrality 4. CONCLUSION & FUTURE WORK In this paper we try to build a model from Eigenvector Centrality, this feature selection used eigenvector to weight a between the node and find the best node then rank feature based on the most important feature. The result from this paper tries to compare between ECFS and Chi-Square and produce analysis that ECFS more robust than Chi-Square if the attribute has been optimized with reduced attribute and Chi-Square, can reach the maximum accuracy when having more attribute than ECFS. If Chi-Square test with the reduced attribute such as fewer than 9 attributes, Chi-Square significantly decreases than ECFS. With this ECFS we can know each attribute that have any major impact on Agent Profit. Future work from this paper is trying to boost the performance of feature selection itself when the dataset has many attributes and try the other dataset with more attribute to test ECFS and make ECFS can develop on other tools or framework so the resource can be developed and easy access to many researchers. REFERENCES [1] I. Guyon and A. Elisseeff, “An Introduction to Variable and Feature Selection,” J. Mach. Learn. Res., Volume 3, Number 3, pp. 1157–1182, 2003. [2] R. Wan, L. Vegas, M. Carlo, and P. Ii, Computational Methods of Feature Selection.” [3] P. S. Bradley, Feature Selection via Concave Minimization and Support Vector Machines, Number 6. [4] L. Solá, M. Romance, R. Criado, J. Flores, A. García del Amo, and S. Boccaletti, Eigenvector centrality of nodes in multiplex networks, Chaos, Volume 23, Number 3, pp. 1–11, 2013. [5] J. Yang and V. Honavar, Feature Subset Selection Using a Genetic Algorithm, IEEE Intell. Syst., vol. 13, pp. 44–49, 1998. [6] N. Bryan and G. Wang, Musical Influence Network Analysis and Rank of Sample-Based Music, Proc. 12th Int. Soc. Music Inf. Retr. Conf., no. ISMIR, pp. 329–334, 2011. [7] G. Roffo and S. Melzi, “Ranking to learn: Feature ranking and selection via eigenvector centrality,” Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), vol. 10312 LNCS, pp. 19–35, 2017. [8] F. R. Pitts, “A graph theoretic approach t o historical geography,” pp. 15–20. [9] F. Harary, S. Review, and N. Jul, “No Title,” vol. 4, no. 3, pp. 202–210, 2008. [10] B. Scholkopft and K. Mullert, “Fisher Discriminant Analysis With,” pp. 41–48, 1999. [11] Y. Bazi, S. Member, F. Melgani, and S. Member, “Toward an Optimal SVM Classification System for Hyperspectral Remote Sensing Images,” no. December 2013, 2006. [12] S. Thaseen and C. A. Kumar, “Intrusion Detection Model Using fusion of Chi-square feature selection and multi class,” J. KING SAUD Univ. - Comput. Inf. Sci., no. 2016, 2015. [13] I. Sumaiya Thaseen and C. Aswani Kumar, “Intrusion detection model using fusion of chi-square feature selection and multi class SVM,” J. King Saud Univ. - Comput. Inf. Sci., vol. 29, no. 4, pp. 462–472, 2017. [14] D. Ballabio, F. Grisoni, and R. Todeschini, “Multivariate comparison of classification performance measures,” Chemom. Intell. Lab. Syst., vol. 174, no. March, pp. 33–44, 2018. [15] E. M. Hand and R. Chellappa, “Attributes for Improved Attributes: A Multi-Task Network for Attribute Classification,” pp. 4068–4074, 2016. http://www.iaeme.com/IJMET/index.asp 613 editor@iaeme.com