تشخيص فرار مالياتي با استفاده از سيستم هوشمند تركيبي

Tax Evasion Detection by Using Combinatory Intelligent System

با توجه به اجرايي شدن سامانه عمليات الكترونيكي موديان مالياتي و ايجاد پايگاه داده هاي مالياتي، امكان پايش اطلاعات موجود با مدل هاي مختلف فراهم شده است. در اين پژوهش، از الگوريتم بهينه سازي جستجوي هارموني به منظور بهينه سازي همزمان پارامترهاي شبكه عصبي پرسپترون چند لايه و تركيب مناسب ورودي ها استفاده شده است. علاوه بر آن نتايج با رگرسيون لجستيك به عنوان هسته سيستم مورد مقايسه قرار گرفته است. متغيرهاي ورودي به سيستم 21 مورد بوده كه با بررسي پژوهش هاي مشابه انجام شده طي 30 سال اخير، اعمال ويژگي هاي مالياتي ايران و نظرخواهي از خبرگان انتخاب شده است. مقايسه نتايج حاصل از شبكه عصبي و رگرسيون لجستيك در دو صنعت مواد غذايي و نساجي نشان مي دهد، استفاده از شبكه عصبي داراي دقت هاي بالاتري بوده و اين تفاوت از لحاظ آماري معنادار مي باشد. در شبكه عصبي به ترتيب در صنعت مواد غذايي و نساجي دقت كلي %78/83 و %85/84، دقت تشخيص شركت هاي فراري %31/80 و %34/84 و دقت تشخيص شركت هاي سالم %20/87 و %36/85 مي باشد. با اعمال مجموعه مدل هاي نهايي سيستم بر روي اطلاعات عملكرد سال 91 اشخاص حقوقي و مقايسه آن با نتايج حسابرسي مالياتي در دو صنعت مواد غذايي و نساجي، به ترتيب دقت كلي %22/92 و %35/82، دقت تشخيص شركت هاي فراري %87/83 و %05/84 و دقت تشخيص شركت هاي سالم %71/92 و %22/82 حاصل شده است. نتايج در داده هاي آزمون بر مبناي اعتبارسنجي ضربدري 10 بخشي با تكرار و ميانگين گيري بر روي 8 حلقه موازي ارايه شده است.

With the electronic taxpayers system being operationalized and the digital storage of tax data developed in Iran, it is now possible to design different models to analyze the available data. There are two main areas that have not been the focus of the fairly limited current studies in this field; one being the parallel optimization of parametric AI models and the other area is the selection of input variable combination. For this reason, in this study, we have used the harmony search (HS) optimization algorithm to do parallel optimization of multilayer perceptron (MLP) neural network parameters and also to find a suitable combination of input variables. In addition to that, the results have been compared with logistic regression results as the core of the system. In the present research, 21 initial input variables are selected for the system based on the survey done on similar studies in the last thirty years and it took into account the specifications of the tax system in Iran and the opinions of the experts in the field were asked. After running the system on the data from the food and textile sectors and comparing the results from the neural network and logistic regression, we have concluded that neural network can produce more accurate results and this difference is statistically meaningful. In the neural network method the overall accuracy is 83.78% and 84.85%, the accuracy of detecting fraudulent companies is 80.31% and 84.34% and the accuracy of detecting non-fraudulent companies is 87.20% and 85.36% in the food and textile sectors respectively. By applying all the final models of the system on data from the 1391 tax year and comparing the results with data from the auditing of companies in the food and textile sectors ,92.22% and 82.35% overall accuracy, 83.87% and 84.05% accuracy for detecting fraudulent companies and 92.71% and 82.22% for detecting non-fraudulent companies have been obtained. The results for the test data has been presented based on 10-fold cross-validation and averaging on 8 parallel loops.