JISIB-vol-12_Nr-3(2022).pdf Journal of Intelligence Studies in Business Vol. 12 No. 3 (2022) Open Access: Freely available at: https://ojs.hh.se/ pp. 18–26 Intelligence on the hiring process Mohamed Beraich Imane Amouri Cheklekbire Malainine ABSTRACT: In this paper, we proposed a decision support tool for recruiters to improve their the process opted by the methods and techniques related to Data Mining. As a result, after completing the modelling process, we were able to obtain a model capable of obtained a model with an accuracy of 99% as well as with a very low error rate. support tool for recruiters while minimising the cost and time of processing applications and maximising the accuracy of the decisions made. KEYWORDS: (AI), Digital Enterprise, Recruitment 19 INTRODUCTION In the current era of globalisation and the emergence of new technologies, as well as the competition of the global business market, companies cannot afford to continue to adopt traditional methods in the various business the smooth running of the company, especially the hiring process follows important steps such as the selection and appointment of suitable candidates for such a vacancy post. Every enterprise invests a lot of money and and wastes resources in searching for poten- tial candidates. The total investment becomes a loss if the selected candidates do not meet the company’s requirements after completing the whole hiring process. Therefore, the objec- tive of this empirical study is to propose a deci- sion support tool to improve the hiring process Neural Network (ANN) method to build a pre- dictive model of the decision in the hiring pro- cess. Our methodology consists of adopting the process applied by Data Mining techniques, starting with a pre-processing and exploratory analysis of the data, then building our model - uating the model using the proportion of test data, using the various validation metrics ema- nating from the confusion matrix. 1. LITERATURE REVIEW the importance of exploiting new methods from - ciplines. to build a model capable of analysing the per- formance of students’ academic records as well grade in different classes (Excellent, Average, Poor), however the results obtained revealed the robustness of these methods. In the employment market, and more Intelligence techniques in the hiring process, several studies have been carried out in this sense in order to improve the said process. Therefore, the study conducted by (D. Alao et Al., 2013), the authors constructed a set of rules using the decision tree method in order to build a model capable of predicting new employee attrition, however, the results obtained yielded a model with an accuracy of 74%. a decision support model for ranking candidates in the employee hiring process using a vari- achieved a maximum accuracy of 88.24% using the decision tree method. 2. HIRING PROCESS sound tactics into the hiring process. Owever, the hiring process can be internal or external, therefore, it can take many forms that differ from one company to another, but remains faithful to the single purpose of choos- interview is a crucial step and represents more than 50% of the rating assigned to the pre-se- lected candidates. 3. ARTIFICIAL NEURAL NETWORK (ANN) research that deals with learning and reason- - niques, unlike parametric techniques, ANNs Neural networks, as systems capable of learning, implement the principle of induction, i.e., learning from experience. By confronta- - grated decision system whose generic charac- ter depends on the number of learning cases encountered and their complexity in relation to the complexity of the problem to be solved. - ally composed of a succession of layers, each of which takes its inputs from the outputs of the previous layer. Each layer i is composed of Ni neurons, taking their inputs from the N 1 20 called the input layer and the last layer, com- posed of a single neuron, is called the output layer. The intermediate layers are called hid- den layers. Networks (ANN) - posed of a succession of layers, each of which takes its inputs from the outputs of the pre- vious layer. Each layer i is composed of Ni Neurons (nr) taking their inputs from the neu- called the input layer and the last layer, con- sisting of a single neuron, is called the output layer. The intermediate layers are called hid- 3.2 Structure and operation of an that receives input from other neurons and weights it with real values called synaptic coef- Consider the neuron of a layer i. Let us note x , x , ..., x the N 1 inputs from the layer i–1 to the neuron of the layer i. We also con- sider the N 1 weights denoted , , ..., . The neuron calculates the sum of its inputs - cients, to which it adds a constant term called the bias b . This gives the formula: = + The bias is an external parameter of the neu- ron . It can be integrated into the weighted sum, as the signal which takes the value 1, weighted by the weight whose value is equal to the bias : The sum can thus be written as: = + 1 2 0 1 2 1 2 3 The hidden layer or The i-layer The Output Layer (s) 0 0 1 2 21 To this sum the neuron applies an acti- vation or transfer function to obtain an out- put The output (output) of the neuron neu- ron in the i layer is sent to other neurons or to the outside. 3.3 Matrix writing We consider the layer i layer composed of M1 neurons. with 1 < < Mi we put: So: = = . = . We pose: So: We put: So: The outputs of the Mi neurons of the layer are then written: + + The weighted sum The applica on of the ac va on func on The predic on of the neuron ( ) 22 So: The transfer function = . The summation function Architecture and functioning of ANN (Source : Author). The list of activation functions (Source: Author). The function title The function The Graphic Representation Sigmoid ReLu 23 3.4 Activation function The transfer function or activation function or thresholding function, also called the activa- tion function, is the function used to propagate information from layer to layer. The most com- mon functions cited in the literature are listed in the following table (Table 1): 3.5 Error functions To calculate the correct weights (parameters), the error between the expected output and the output produced by the network must be calculated. Methods for calculating the error include: • : With: • : m The number of individuals or objects to be predicted or the number of observations. : network The vast majority of neural networks have a “training” algorithm which consists of modifying the synaptic weights according to a set of data presented as input to the net- work. The purpose of this training is to allow the neural network to learn from the examples. If the training is carried out correctly, the net- work is able to provide output responses very close to the original values of the training data- set. But the interest of neural networks lies in their ability to generalise from the test set. It is therefore possible to use a neural network to - ory. Supervised learning occurs when the net- state as it is presented with a pattern. In contrast, in unsupervised learning, state when presented with a pattern. ANN learning can be achieved, among other things, by: i) Changing weights, (creation or deletion of neurons or con- nections, or layers), iii) The use of appropriate attractors or other appropriate steady state points, iv) The choice of activation functions. Since backpropagation training is a gradi- ent descent process, it can get stuck in local minima in this weight space. It is because of this possibility that neural network models are characterised by high variance and instability. Back-propagation Backpropagation consists of backpropa- gating the error committed by a neuron to its synapses and the neurons connected to - gation of the error gradient is usually used, which consists of correcting errors according to the importance of the elements that have actually participated in the making of these errors: the synaptic weights that contribute to - ated a marginal error. How to choose the number of layers and neurons The number of neurons and layers directly of prediction quality. Indeed, to determine the number of hidden layers, we can follow a process that consists in starting with a single hidden layer and adapting it to reach the ideal architecture. So if one layer does not produce satisfac- tory results, then we automatically have to think about adding another until we get satis- factory results. The same goes for the number 24 of neurons, we try to modify it until we get the desired results. The number of neurons in each layer must not exceed the number of input variables. So, you have to think about doing several tests to arrive at a relevant and pow- erful ANN in terms of accuracy in predicting the output variables. On the other hand, the more layers you increase the capacity of the network, the more you risk overlearning if you exaggerate in terms of the number of layers or neurons, and the same thing if you decrease the number of layers, you risk underlearning. To avoid the problem and we try to divide the data into 4 parts and try to alternate the combinations between these parts. By applying this tech- nique, we will have a perfect test of the data since all parts will be used for the test. 4 METHODOLOGY, METRICS AND DATA 4.1 Methodology The aim of this empirical study is to build a model that can be implemented as a decision support tool for recruiters to effectively hire - Intelligence, for which we adopted the pro- cess of data mining techniques. This process is initially based on the prepa- ration of the data, followed by the splitting intended to train the prediction model, while the second serves as a test proportion for the accuracy of the resulting model. 4.2 Metric To evaluate the model obtained from the mod- metrics to assess the performance of the model - sion matrix using the test data set. Given that the test data set represents 25% of the overall data and the training set represents 75% of the overall data. allows us to indicate the number of correct predictions for each class and the number of incorrect predictions for each class organised according to the predicted class. Each row of the table corresponds to a predicted class, and each column corresponds to an actual class. Confusion matrix. True Positives (TP) True Negatives (TN) With: - Applicant database (Accepted or rejected) The Entry Layer Training the neural network 1 2 3 Predict whether a candidate will be accepted or rejected 25 ratios can be calculated: 4.3 Data In the data preparation stage, we used a data- base that includes 1000 rows of applicants from a recruitment agency. In addition, this data- base has 8 explanatory variables and only one dichotomous variable to be explained which takes 2 binary values (Accept / Reject), so we coded all categorical variables according to the table below (See Table 3): 5. RESULTS After preparing the data for the modelling, The function function is used to train our model over 50 iterations, allowing us to Coding of explanatory variable values. Code 1 2 3 4 5 Speciality Computer Science Secretariat Management Right Current Status Unemployment Assets French level A1 A2 B1 B2 C1 English level A1 A2 B1 B2 C1 Computer level Beginner Medium Advanced Excellent Decision Reject (0) Accept (1) The architecture of our neural network. Figure 6. The evolution of the error of our model. Figure 7. The evolution of the accuracy of our model. 26 choose the right values for the weight matrix W. The calculations are performed using the gradi- ent descent method. The training data used are stored on (starting values) and the evolution of the accuracy and the error (loss) of the model in the training phase. decreases and the accuracy increase with iter- ations, as the training algorithm continuously updates the weights and biases in the neural network according to the training data. We curves (blue and orange) are very close for both Test and Train data sets, which means that the model has been well trained. We also notice the Test and Train data sets decrease towards 0, which means that the model performs well. Thus, we calculated the metric for the training and test data and obtained an accuracy equal to 99,33% using the test data Indeed, according to the value of the met- ric obtained, we can conclude that our model has a fairly high level of predictability, which will help us to make accurate predictions of the recruitment decision. 6. CONCLUSION Selecting and hiring the right candidate is a daunting task for the company. Therefore, companies are looking for tools that can collect, sort and analyse a large amount of informa- tion about candidates to assess their person- Intelligence provides to improve this hiring process. It is in this context that our paper is writ- ten, we have tried to detect the importance of using these techniques in the construction of a model capable of predicting the recruitment decision of new candidates for a company. so we have exploited a database that includes a range of explanatory variables that describe the level of competence of candidates. After following the process adopted by data mining techniques, we were able to achieve - racy obtained at the end of the modelling pro- cess, which exceeds 99%, reveals the robust- ness of the model obtained, which will improve the hiring process for companies. REFERENCES [1] C. E. A. Pah and D. N. Utama, “Decision support model for employee recruitment - tional Journal, vol. 8, no. 5, 2020. [2] D. Alao and A. Adeyemo, “Employee at- trition analysis using decision tree algo- rithms”,Computing, Information Systems, Development Informatics and Allied Re- search Journal, vol. 4, no. 1, pp. 17–28, 2013. [3] Dana Pessach, Gonen Singer, Dan Avra- - li, Irad Ben-Gal, “Employees recruitment: A prescriptive analytics approach via machine learning and mathematical pro- gramming”, Decision Support Systems 134 (2020). - ligence to Recruit Employees”, Citations - ume 18, 2021. enterprise human resource management computer hardware”, Microprocessors and Microsystems 82 (2021). - gineering students’ performance for re- techniques”, International Journal of Com- puter Science, Engineering and Technolo- gy, vol. 3, no. 2, p. 31, 2013. - mance Analysis of Engineering Students Mining Techniques”, Samrat Singh et. al 31–37.