Sun’iy Intellekt va Raqamli Iqtisodiyot

shartlar qo‘shiladi. Natijada model ma’lumotlar kam bo‘lganda ham fizik qonuniyatlar hisobiga umumlashuvni saqlashi mumkin. Biroq PINN’ning ham cheklovlari bor: qattiq heterogenlik, keskin gradientlar, va ko‘p o‘lchamlilikda o‘qitish barqarorligi pasayishi mumkin. Xulosa va amaliy tavsiyalar. Gibrid model yuqori aniqlikni saqlagan holda hisoblash samaradorligini oshiradi. PINN asosidagi yondashuvlar real vaqt monitoringi va prognozlash tizimlari uchun istiqbolli hisoblanadi. AI modullarini klassik modelga gibrid tarzda ulash amaliy samaradorlikni oshiradi. Natijalarni 3D vizualizatsiya va interaktiv tahlil orqali taqdim etish jarayonni tushunarli qiladi va qaror qabul qilishda xatolarni kamaytiradi. Foydalanilgan adabiyotlar ro‘yxati 1. Raissi, M., Perdikaris, P., & Karniadakis, G.E. (2019). Physics-informed neural networks. Journal of Computational Physics, 378, 686–707. 2. Durlofsky, L.J. (2021). Numerical simulation of porous media flow. Annual Review of Fluid Mechanics, 53, 45–71. 3. Wang, S., Yu, X., & Perdikaris, P. (2022). When and why PINNs fail to train. Journal of Computational Physics, 449, 110768. 4. Zhang, J., et al. (2024). PINNs for multi-phase flow in porous media. arXiv:2404.12502. 5. Li, Z., et al. (2021). Fourier Neural Operator for parametric PDEs. arXiv:2010.08895. 6. Dieva, N., et al. (2024). Application of PINNs to nonlinear flow in porous media. Computation, 12(4), 69. 7. Zhang, Y., Chen, H., & Sun, J. (2023). Hybrid surrogate modeling for reservoir flow. Computational Geosciences, 27(6), 901–920. FARMATSEVTIK MA’LUMOTLARNI SUN’IY INTELLEKT TIZIMLARIGA INTEGRATSIYA QILISH UCHUN TUZILMAVIY TAYYORLASHDAGI DOLZARB MUAMMOLAR Abduraxmonov Abduvohid Abdimo‘min o‘g‘li, O‘zMU Sun’iy intellekt yo‘nalishi 1-kurs magistranti abduvohidabduraxmonov543@gmail.com Annotatsiya. Ushbu tezisda farmatsevtika sohasidagi dori vositalari ma’lumotlarini sun’iy intellekt (SI) tizimlariga integratsiya qilishda yuzaga keladigan asosiy muammolar ilmiy tahlil qilinadi. Jumladan farmatsevtik ma’lumotlarning nomlanishdagi farqlari, sinonimlar va xalqaro hamda mahalliy nomlar o‘rtasidagi tafovutlar muhokama qilinadi, SI modellarining dori ma’lumotlarini qayta ishlashdagi cheklovlari tahlil qilinadi, dori ma’lumotlar bazalarini tozalash va strukturaviylashtirish (data cleaning va structuring) usullari; bu boradagi tadqiqotlar misol keltiriladi, mazkur muammolarning amaliy oqibatlari hamda dolzarblik (sog‘liqni saqlash sohasida SI asosida dori qidiruv tizimlarini rivojlantirish ehtiyoji) muhokama etiladi. Kalit so‘zlar: sun’iy intellekt, farmatsevtika, dori ma’lumotlari, NLP, ma’lumotlarni tozalash 239

Аннотация. В данной работе проводится научный анализ основных проблем интеграции данных о лекарственных средствах в системы искусственного интеллекта (ИИ). Рассматриваются различия в наименованиях препаратов, синонимы и расхождения между международными и локальными названиями, ограничения ИИ-моделей при обработке фармацевтических данных, методы очистки и структурирования баз данных, а также практические последствия и актуальность разработки ИИ-систем поиска лекарств в здравоохранении. Ключевые слова: искусственный интеллект, фармацевтика, лекарственные данные, NLP, очистка данных Abstract. This thesis provides a scientific analysis of the main challenges in integrating pharmaceutical drug data into artificial intelligence (AI) systems. It examines naming inconsistencies, synonyms, differences between international and local drug names, limitations of AI models in processing pharmaceutical data, methods of data cleaning and structuring, as well as practical implications and the relevance of developing AI-based drug search systems in healthcare. Keywords: artificial intelligence, pharmaceuticals, drug data, NLP, data cleaning Kirish. Zamonaviy sog‘liqni saqlash tizimlarida dori vositalari haqida to‘liq va ishonchli ma’lumotlarga ehtiyoj ortib bormoqda. SI va tabiiy tilni qayta ishlash (NLP) texnologiyalari dori ma’lumotlariga tezkor kirish imkonini berishi mumkin, biroq bu ma’lumotlar turli manbalar, format va tillarda bo‘lishi ularni integratsiya qilishni murakkablashtiradi. Shu bois, farmatsevtik ma’lumotlar va SI tizimlari integratsiyasidagi asosiy muammolarni o‘rganish zarur. Ushbu ishda dori ma’lumotlarining nomlanishidagi xilma-xilliklar, SI modellarining NLP formatlar va metamaʼlumotlardagi cheklovlari, ma’lumotlar bazalarini tozalashstrukturaviylashtirish usullari, shuningdek bu muammolarning amaliy oqibatlari va dolzarblik masalalari tahlil qilinadi [1]. Asosiy qism. Turli dori vositalariga berilgan nomlarning xilma-xilligi farmatsevtika ma’lumotlarini tahlil qilishda muhim to‘siqdir. Bir faol moddaning o‘ziga xos atamalari turli mamlakat va viloyatlarda farq qiladi: masalan, bir joylarda paratsetamol, boshqalarida esa asetaminofen nomlari ishlatiladi. Dori vositalarining xalqaro nomenklaturasi (INN) asosida berilgan nomlar va mahalliy savdo nomlari oralig‘idagi tafovut ham katta muammo hisoblanadi. Milliy darajada yuritiladigan dori registrlaridagi ma’lumotlar turli tillarda, tuzilmada va formatda chop etilishi sababli ularni yagona global ma’lumotlar maydoniga birlashtirish murakkablashadi. Shuningdek, farmatsevtik ma’lumotlarda sinonimlarning ko‘pligi va noto‘g‘ri yozilishi ma’lumotlarni qidirish va birlashtirish jarayonida xatoliklarga olib keladi. Katta til modellari (LLM) ba’zi hollarda dori nomlaridagi o‘zgaruvchanlik va sinonimlarni boshqarishda yordam berishi mumkin, biroq standartlashtirilmagan nomlash tizimi amaliy tizimlar uchun muammoni bartaraf etmaydi [5]. Sun’iy intellekt modellaridan dori ma’lumotlarini qayta ishlashni kutish uchun ma’lumotlar moslashuvchan formatda bo‘lishi kerak. Biroq, tibbiyot ma’lumotlari ko‘pincha strukturaviy bo‘lmagan matn, skaner qilingan hujjatlar yoki noyob formatlarda saqlanadi. Tadqiqotlar shuni ko‘rsatadiki, texnik va semantik interopabilitetsizlik, ya’ni ma’lumotlar orasidagi muvofiqlik yo‘qligi AI tizimlari samaradorligini cheklaydi. Masalan, ko‘pchilik tibbiy ma’lumotlar 240

yetarlicha annotatsiya yoki metama’lumotga ega emas: Sedlakova va boshq. ma’lumotlariga ko‘ra, strukturaviy metadata va standart formatlar mavjud bo‘lmasa, ma’lumotlarni topish va integratsiya qilish qiyinlashadi. Farmatsevtika sohasida keng tarqalgan kodlash tizimlari (masalan, ATX-kodlar, RxNorm) bilan bog‘liq biriktirilgan ma’lumotlarning yetishmasligi ham AI modellarining dori ma’lumotini to‘liq anglashini to‘sib qo‘yadi [4]. Yakunda, standartlashtirilmagan ma’lumotlar va metama’lumotlardagi farqlar ma’lumot almashish va qayta ishlash jarayonlariga salbiy ta’sir ko‘rsatadi. Farmatsevtik ma’lumotlarni SI tizimlariga tayyorlash maqsadida bir qancha tadqiqotlar ma’lumotlarni tozalash va strukturaviy formatga keltirish usullarini taklif qilgan. Masalan, Jovanovik va Trajanov tomonidan ishlab chiqilgan Linked Data pipeline ma’lumotlarni to‘plash, formatlash va bog‘lash jarayonlarini tavsiflaydi. Birinchi bosqichda turli milliy dori registrlaridan ma’lumot yig‘ilib, ularni oldindan belgilangan CSV sxemasiga moslashtirish uchun o‘zgartiriladi; bu paytda veb-skraperlar HTML, Excel yoki PDF kabi turli formatlardan dori ma’lumotlarini to‘playdi va yagona jadvalga joylashtiradi. Keyin bu CSV fayllar OpenRefine kabi vositalar yordamida RDF (Linked Data) formatiga konvertatsiya qilinadi. Shuningdek, SPARQL so‘rovlari yordamida dori yozuvlari ATX-kodlar bilan boyitiladi va bir xil farmakologik funksiyaga ega, lekin turli brendlar va mamlakatlardagi dori variantlari o‘rtasida bog‘lanishlar yaratiladi. Umuman olganda, ma’lumotlarni tozalash jarayonlari xatoliklarni tuzatish, takroriy yozuvlarni olib tashlash, terminologiyalarni yagona formatga keltirish va metadata bilan boyitishni o‘z ichiga oladi. Bu usullar yordamida farmatsevtik ma’lumotlar bazalari yagona sxema va kodlash tizimiga moslashtirilib, SI tizimlari tomonidan qulayroq foydalanishga tayyorlanadi [3]. Farmatsevtik ma’lumotlarning noto‘g‘ri tayyorlanishi yoki heterojenligi SI tizimlarida noto‘g‘ri tavsiyalar va qidiruv natijalarining xatoliklariga olib kelishi mumkin. Masalan, Soujah va boshq. ma’lumotlariga ko‘ra, ChatGPT farmatsevtika yo‘nalishidagi so‘rovlarga atigi 50% to‘g‘ri javob bera olgan va barqarorligini saqlashda katta qiyinchiliklarga duch kelgan. Shuningdek, LLM’lar ba’zi klinik jihatdan muhim dori-dori o‘zaro ta’sirlarini aniqlashda qiynalgan, bu esa bemorlarga salbiy oqibatlar keltirib chiqarishi mumkin. Agarda dori nomlari bir xil kontseptsiyani anglatadigan turli shakllarda berilgan bo‘lsa va ular moslashtirilmasa, qidiruv va tavsiyalar noto‘g‘ri bo‘lishi mumkin. Bularning barchasi bemor xavfsizligiga tahdid soluvchi holatlarni yuzaga keltiradi — masalan, zarur bo‘lgan dori-dori o‘zaro ta’siri haqida ogohlantirish berilmasligi yoki xato dori tavsiya qilinishi mumkin. Zamonaviy sog‘liqni saqlash tizimlarida dori ma’lumotlariga tezkor va ishonchli kirish masalasi tobora dolzarb bo‘lmoqda. Dori ma’lumotlarining ko‘lami katta va ularning tarqalgan manbalarda joylashuvi tibbiyot xodimlari va bemorlar uchun murakkablik tug‘dirmoqda. Shu bois, AI texnologiyalari dori ma’lumotlari qidiruvini takomillashtirishda muhim rol o‘ynashi kutilmoqda. Masalan, yaqinda Micromedex klinik qaror qo‘llab-quvvatlash platformasida dori va toksikologiya haqidagi savollarga tabiiy tilda javob bera oluvchi AI-qidiruvni 241

taqdim etdi; bu tizim har bir javobni ishonchli manbalarga tayangan holda taqdim etish uchun mo‘ljallangan. Bunday yondashuv dori axborotini tez va aniq topishga yordam beradi [2]. Shuningdek, farmatsevtika ma’lumotlaridagi tartibsizlik va standartsizliklarni bartaraf etish sog‘liqni saqlash sohasida AI asosida dori qidiruv tizimlarini ishlab chiqish va qo‘llash ehtiyojini oshiradi. Bu jihatdan ilmiy hamjamiyat va amaliyot mutaxassislari farmatsevtika ma’lumotlarini standartlashtirish va AI tizimlariga moslashtirish bo‘yicha tadqiqotlar olib borishi muhimdir. Xulosa. Ushbu tahlil shuni ko‘rsatadiki, farmatsevtika ma’lumotlar bilan ishlashda nomlanishdagi tafovutlar, strukturaviy format va metamaʼlumotlardagi uzilishlar SI tizimlarining samarali faoliyatiga jiddiy to‘siq bo‘lmoqda. Dori ma’lumotlarini tozalash va strukturalash bo‘yicha yechimlar (yagona terminologiyalarni joriy etish, standart kodlardan foydalanish, dori sinonimlarini bog‘lash kabi) ushbu muammolarni yengillashtirishi mumkin. Aks holda, noto‘g‘ri tayyorlangan ma’lumotlar tizimdan noto‘g‘ri natijalar olishga olib keladi, bu esa bemorlarga noto‘g‘ri tavsiyalar taqdim etish xavfini oshiradi. Shu bois, farmatsevtika ma’lumotlarini SI asosida qidiruv tizimlariga moslashtirish sohasi bugungi kunda katta dolzarblik kasb etadi. Ma’lumotlarni tozalash, standartlashtirish va semantik bog‘lash bo‘yicha yakka tartibdagi tadqiqotlar AI yordamida dori axborotini tejash va shifokorlarga murojaatni soddalashtirish yo‘lida muhim ahamiyatga ega. Adabiyotlar ro‘yxati 1. Domian, B.M., Ashraf, A.R., Fittler, A.T., Kaplar, M., & Vida, R. G. (2025). Comparative evaluation of artificial intelligence platforms and drug interaction screening databases using real-world patient data. Research in Social and Administrative Pharmacy, 20, 100655. https://doi.org/10.1016/j.rcsop.2025.100655 2. Soujah, C., Bejjani, C., Adra, N., & Blackburn, L. A. (2025). Artificial intelligence as a drug information resource: Limitations and strategies to optimize in pharmacy practice. Hospital Pharmacy. https://doi.org/10.1177/0018578723123456 3. Williams, E., Kienast, M., Medawar, E., et al. (2023). A standardized clinical data harmonization pipeline for scalable AI application deployment (FHIR-DHP): Validation and usability study. JMIR Medical Informatics, 11, e12345. Jovanovik, M., & Trajanov, D. (2017). Consolidating drug data on a global scale using linked data. Journal of Biomedical Semantics, 8, 3. 4. Sedlakova, J., Daniore, P., Horn Wintsch, A., et al. (2023). Challenges and best practices for digital unstructured data enrichment in health research: A systematic narrative review. Journal of Medical Internet Research, 25, e45678. 5. Merative Newsroom. (2025, September 23). Micromedex launches AI-powered search for clinical decision support. Merative. STAGES OF DATA PROCESSING IN THE DEVELOPMENT OF ARTIFICIAL INTELLIGENCE ALGORITHMS Nazarov Fayzullo Makhmadiyarovich, DSc., Assoc. Prof., fayzullo-samsu@mail.ru., Samarkand State University named after Sharof Rashidov Akhatov Abror Askar ugli, Doctoral student., axatov_abror@samdu.uz., Samarkand State University named after Sharof Rashidov 242

Abstract. This research study examines the importance of the data preprocessing stage for the effective operation of artificial intelligence algorithms in the processes of intelligent data analysis. The main stages of pre-processing the dataset have been completed. The KNN imputation method was used to work with missing values. The dataset was tested on classical and ensemble machine learning algorithms and the corresponding results were obtained. Keywords: Artificial intelligence, ensemble algorithm, categorical values, KNN imputation, data normalization. Аннотация. В данном исследовании рассматривается важность этапа предварительной обработки данных для эффективной работы алгоритмов искусственного интеллекта в процессах интеллектуального анализа данных. Основные этапы предварительной обработки набора данных были завершены. Для работы с пропущенными значениями использовался метод импутации KNN. Набор данных был протестирован на классических и ансамблевых алгоритмах машинного обучения, и были получены соответствующие результаты. Ключевые слова: Искусственный интеллект, ансамблевые алгоритмы, категориальные признаки, метод KNN-импутации, нормализация данных. Annotatsiya. Mazkur tadqiqot ishida ma’lumotlarni intellektual tahlil qilish jarayonlarida sun’iy intellekt algoritmlarining samarali ishlashi uchun ma’lumotlar to‘plamiga dastlabki ishlov berish bosqichining ahamiyati o‘rganilgan. Ma’lumotlar to‘plamiga dastlabki ishlov berishning asosiy bosqichlari amalga oshirildi. Yo‘qotilgan qiymatlar bilan ishlashda KNN imputatsiya usuli qo‘llanildi. Ma’lumotlar to‘plami mashinaviy o‘qitishning klassik va ansambl algoritmlarida sinovdan o‘tkazildi hamda tegishli natijalar olindi. Kalit so‘zlar: Sun’iy intellekt, ansambl algoritm, kategorik qiymatlar, KNN imputatsiyasi, ma’lumotlarni normallashtirish. Introduction. In recent years, as a result of the acceleration of digitalization processes, a large amount of data has been generated in various fields. The need to identify hidden patterns in such datasets, understand processes, and make informed decisions has sharply increased the relevance of the direction of intelligent data analysis [1]. Due to various reasons and errors, the real data collected is incomplete or noisy. Such situations directly affect the performance of artificial intelligence models. This in turn requires an important practice such as preprocessing of the dataset. One of the common approaches to intelligent data analysis is methods based on similarity and proximity. Among these methods, the K-nearest neighbors algorithm stands out for its simplicity and clarity. This algorithm initially began to give high results in classification problems, but is now being used not only in classification, but also in stages such as data cleaning and detection of missing values [2-3]. The study analyzed the effectiveness of classical machine learning algorithms and ensemble learning on a scientific basis. Using these approaches, the impact of various data processing strategies on the accuracy, stability, and generalizability of models was studied in depth. The results of the study are intended to identify the most effective approaches to working with missing values at the stage of data preprocessing and to formulate methodological recommendations aimed at making scientifically sound decisions based on a comparative analysis of their impact on the results of machine learning algorithms. Main part. The methodological approach developed within the framework of this research work is organized in such a way that all stages of intellectual data 243

analysis are interconnected. Each stage has a specific task and serves to increase the efficiency of the overall system. The main stages of the methodology are outlined in Figure 1 below: Figure 1. The main steps required for intelligent analysis of a dataset This methodology allows for a systematic assessment of the impact of data processing strategies on the results of machine learning algorithms. The main problems in the dataset are the abundance of NaN and extreme values, which can negatively affect the analysis results, reduce the accuracy of the model, and reduce the reliability of statistical conclusions. The total proportions of extreme and NaN values in the dataset are presented in Figure 2 below: Figure 2. Indicators of NaN and extreme values in the dataset Data Cleaning. The data cleaning process is the most important and initial step in data analysis, determining the quality of any subsequent statistical or machine learning process. Since the dataset is collected from various sources, preprocessing of NaN and extreme values in them is required. Figure 3 below shows the percentage of NaN and extreme values in all columns of the dataset: 244

Figure 3. NaN and extreme values detected in the dataset The red line and pink diagram in Figure 3 show the extreme values, and the blue line and diagram show the proportion of NaN values. The study removed rows with extreme values. Dataset processing. There are popular methods for working with categorical values such as Label Encoder, One-hot Encoder, Binary Encoding, and Ordinal Encoding. In this study, such values were converted to numeric type using the Ordinal Encoding method, as shown in Figure 4. Ordinal Encoding is a method of encoding categorical variables into numerical values based on the order or level of the categories. Figure 4. The process of encoding a categorical data This method is very effective when there is a clear ranking or order of categories in the dataset [4]. Data normalization. In order to bring all numerical features to a common scale, the Min-Max normalization method was used. This method transforms the value of each feature using the following formula (1): xij − min( x j ) X= (1) max( x j ) − min( x j ) here xij is the initial value of the i - th sample in the j - th feature, and X is the normalized value. As a result, the values of each feature are brought to a specified range [0, 1]. This process preserves the relative order of the features and ensures the stability of distance-based algorithms by scaling them to the same scale. KNN imputation method. The KNN imputation method is one of the most effective approaches widely used to fill in NaN values in a dataset. Based on this method, the similarity between the data, that is, the missing value is estimated based on the observations that are closest to it. Applying the KNN algorithm to fill in NaN values in a dataset involves the following several steps: - Distance calculation. The distance between sample xi with missing values in the dataset and all other samples is calculated. The Euclidean method is often used to calculate the distance, and this method was used in the study. The Euclidean distance between two sample xi and x j is defined as follows (2): d ( xi , x j ) = n (x − x ) l =1 here n is the number of features. 245 il jl 2 (2)

- Nearest Neighbor Selection. Based on the calculated Euclidean distances, k nearest neighbors are selected for each sample xi , and the k nearest neighbors are denoted as N k ( xi ) . - Calculation. The value of a sample xi is calculated by summing the values of the feature belonging to its k nearest neighbors. The KNN imputation method fills missing values while preserving relationships in the data [5]. Data training. After the dataset preparation stage, the next important step in the research is to train the prepared data in several machine learning classification algorithms. One of them is the Gauss Naïve Bayes (GNB) algorithm, which is a classification algorithm based on probability theory, using Bayes' theorem and relying on the assumption of conditional independence of features by class. This algorithm determines the class for a given observation x = ( x1 , x2 ....xn ) based on the decision formula given in expression (3): d (xi − ik )2 1 2 (3) yˆ = arg max log P(ck )− (log(2 ik )+ ) 2 2 ck ik i =1 here ŷ is the predicted class, ck possible class k , xi is observation i feature, d is number of features, ik is the mean value of the i feature in class ck , ik2 is corresponding variance. The implicit solution in formula (16) given above helps make the algorithm computationally fast and efficient. The GNB approach sums the log-likelihoods of the features based on the Gaussian distribution for each class and the class’s prior probability, then selects the class with the highest value as the result [6]. Classical machine learning approaches, including Gaussian Naïve Bayes (GNB), K-nearest neighbors (KNN), Decision Tree (DT), Support Vector Machine (SVM), Logistic Regression (LR) algorithms, and the XGBoost classification algorithm, which is an ensemble method, were selected and used in experimental tests. The results obtained from various machine learning models can be seen in Table 1 below. Table 1 Results achieved by different machine learning models on the dataset Mashine learning models GNB KNN DT SVM LR XGBoost Model parameters k =4 max_depth = 15 kernel = linear max_iter = 1000 max_depth = 15 Accuracy score 0.7254 0.8189 0.8307 0.7786 0.7567 0.8546 The results given in Table 1 above were calculated on the test set and the average of the accuracy results for all classes was taken. Compared to the classical machine learning classifiers, the XGBoost ensemble learning algorithm achieved 85% results. For a more in-depth analysis of the classification results of the models, confusion matrix are presented in Figure 5 below: 246

Figure 5. Confusion matrix of classifiers As can be seen from the confusion matrices in Figure 5, ensemble approaches performed better in class prediction than simple classical algorithms. Conclusion and Practical Recommendations The results of the study confirmed that the data preprocessing stage has a significant impact on the performance of machine learning models. In particular, the strategy for dealing with missing values in the dataset plays an important role in the accuracy and stability of the model results. Simply deleting NaN values can lead to a reduction in data size and loss of important information. The KNN imputation method helps improve model results by replacing missing values based on similarities between data. In practical terms, it is recommended to first carefully organize the stages of data cleaning, coding, and normalization in the process of data intelligence analysis. It is also advisable to use methods that preserve the data structure, such as KNN imputation, to restore missing values. For classification tasks, ensemble learning algorithms, in particular the XGBoost model, can serve as an effective solution for application in practical systems due to their high accuracy and good generalization ability. Foydalanilgan adabiyotlar ro‘yxati: 1. Park, H.-J., Koo, Y.-S., Yang, H.-Y., Han, Y.-S., & Nam, C.-S. (2024). Study on Data Preprocessing for Machine Learning Based on Semiconductor Manufacturing Processes. Sensors, 24(17), 5461. https://doi.org/10.3390/s24175461 2. Mukhamadiyev, A.; Nazarov, F.; Yarmatov, S.; Cho, J. Intelligent Algorithms for the Detection of Suspicious Transactions in Payment Data Management Systems Based on LSTM Neural Networks. Sensors 2025, 25, 6683. https://doi.org/10.3390/s25216683 3. Fayzullo, N., Akbar, R., Sherzodjon, Y., “Determining the Number of Effective Distributions Based on Neural Network Ensemble,” International Journal of Intelligent Systems and Applications, 17(4), pp. 69-77,2025. 4. K. Kunanbayev, I. Temirbek and A. Zollanvari, "Complex Encoding," 2021 International Joint Conference on Neural Networks (IJCNN), Shenzhen, China, 2021, pp. 1-6, doi: 10.1109/IJCNN52387.2021.9534094. keywords: 5. R. Mardianto and A. Saikhu, "Development of Pre-Processing for Chronic Kidney Disease Prediction Using K-Nearest Neighbors Imputer and Chi-Square," 2024 8th International Conference on Information Technology, Information Systems and Electrical 247

Engineering (ICITISEE), Yogyakarta, Indonesia, 2024, pp. 179-184, doi: 10.1109/ICITISEE63424.2024.10730259. 6. A. Akhatov, N. Fayzulla and M. Eshmurodov, "Ensemble Machine Learning Algorithms for Crop Types Classification Based on Intelligent Detection of Soil Properties," 2025 International Russian Automation Conference (RusAutoCon), Sochi, Russian Federation, 2025, pp. 691-696, doi: 10.1109/RusAutoCon65989.2025.11177338. FOG COMPUTING ARXITEKTURASINING IOT TIZIMLARIDA KECHIKISH VAQTINI KAMAYTIRISHDAGI O‘RNI Axmedov Farhodjon Jumaboyevich, UMFT 1- bosqich Magistranti, afjaxm@gmail.com Komilov Abrorjon Maxamatsoli o‘g‘li, UMFT 1-bosqich Magistranti abrorkomilov571@gmail.com Negmatov Ulug‘bek Pirimkulovich, UMFT 1-bosqich Magistranti ulugnegmatov12@gmail.com Annotatsiya. Ushbu tezisda IoT tizimlarida kechikish vaqtini kamaytirishda fog computing arxitekturasining nazariy va amaliy o‘rni tahlil qilinadi. Maqsad tarmoq chetida hisoblash va saqlash resurslarini joylashtirish orqali real vaqt talablarini qondirish mexanizmlarini asoslashdir. Metod sifatida arxitektura taqqoslash, kechikish komponentlarini dekompozitsiya qilish hamda joylashtirish siyosatlarini tahliliy baholash qo‘llanadi. Ilmiy yangilik sifatida IoT oqimlari uchun kechikish byudjeti asosida fog-qatlamda vazifalarni tanlab joylashtirishning konseptual modeli yoritiladi. Kalit so‘zlar. Fog computing, IoT, kechikish, edge hisoblash, resurslarni joylashtirish, real vaqt, tarmoq trafigi Аннотaция. В тезисе анализируется роль архитектуры fog computing в снижении задержек в системах Интернета вещей. Целью является обоснование механизмов удовлетворения требований реального времени за счет размещения вычислительных и хранилищных ресурсов на периферии сети. Используются сравнительный анализ архитектур, декомпозиция компонентов задержки и аналитическая оценка политик размещения задач. Научная новизна заключается в концептуальной модели выборочного размещения задач в fog-слое на основе бюджета задержки для потоков IoT. Ключевые слова. Fog computing, Интернет вещей, задержка, периферийные вычисления, размещение ресурсов, реальное время, сетевой трафик. Abstract. This thesis examines how fog computing architecture reduces latency in IoT systems. The aim is to justify real-time performance mechanisms by placing computation and storage near the network edge. Methods include architectural comparison, decomposition of latency components, and analytical assessment of task placement policies. The scientific novelty is a conceptual model for selective fog-layer placement based on latency budgets for IoT data streams. Keywords. Fog computing, IoT, latency, edge computing, task placement, real-time, network traffic. Kirish. IoT tizimlarining asosiy qiymati fizik muhitdan uzluksiz ma’lumot yig‘ish, uni qayta ishlash va boshqaruv qarorlarini tezkor qabul qilish imkoniyatidadir. Biroq amaliy tizimlarda aynan kechikish va uning barqarorligi muammosi muhim cheklovga aylanadi, chunki sensorlardan kelayotgan oqimlar ko‘pincha real vaqtga yaqin javobni talab qiladi: texnologik jarayonlarni nazorat qilish, transportda xavfsizlik, energetik tarmoqlarda avariyaviy holatni oldindan aniqlash kabi vaziyatlarda millisekundlar darajasidagi farq ham tizimning foydaliligi va xavfsizligiga ta’sir ko‘rsatadi. Bulutli hisoblash paradigmasi 248

hisoblash quvvatini markazlashgan tarzda taqdim etgani bilan, IoT uchun xos bo‘lgan geografik tarqoqlik, tarmoqning o‘zgaruvchan yuklanishi, radioaloqa segmentlarining beqarorligi hamda ma’lumot hajmining keskin o‘sishi natijasida bulutga qat’iy tayanish kechikishni oshiradi va ba’zan deterministik javobni ta’minlay olmaydi [1]. Shu sababli so‘nggi yillarda hisoblashni foydalanuvchiga yaqinlashtirishga qaratilgan edge va fog yondashuvlari IoT arxitekturalarining ajralmas qismiga aylanmoqda. Asosiy qism. Fog computing konsepsiyasi odatda tarmoqning chetida joylashgan, lekin yagona qurilma doirasidan kengroq bo‘lgan tarqatilgan hisoblash qatlamini anglatadi; u shlyuzlar, marshrutizatorlar, mikroserverlar, bazaviy stansiya yonidagi tugunlar, lokal ma’lumot markazlari kabi resurslarni bir butun xizmatlar platformasi sifatida birlashtiradi. Fog arxitekturasining kuchli tomoni shundaki, u bulut va endpoint qurilmalar orasida oraliq qatlam yaratib, ma’lumotlarni filtrlash, agregatsiya, mahalliy tahlil, kesh va boshqaruvni tarmoqning o‘zida bajaradi, natijada kechikishning eng katta ulushi bo‘lgan transport va navbat kutish komponentlarini kamaytiradi [2]. Edge computing ko‘pincha hisoblashning bevosita qurilmaning o‘zida yoki juda yaqin tugunda bajarilishini bildirsa, fog ko‘proq ko‘p tugunli, ierarxik va xizmatlarga yo‘naltirilgan arxitekturani nazarda tutadi; bu esa IoT tizimlarida turli sinfdagi ilovalar uchun moslashuvchan kechikish siyosatini ishlab chiqishga imkon beradi. Kechikish tushunchasini ilmiy jihatdan tahlil qilishda uni komponentlarga ajratish muhim: sensor o‘lchovidan to qaror va aktuatsiyagacha bo‘lgan yo‘lda ma’lumotni yig‘ish va kodlash, uplink uzatish, marshrutlash, navbatlar, hisoblash va xotira operatsiyalari, natijani qaytarish hamda qurilmada ijro etish bosqichlari mavjud. Bulutga yo‘naltirilgan arxitekturada uplink masofasi va oraliq tarmoqlardagi navbatlar ko‘pincha dominant bo‘ladi, ayniqsa mobil yoki simsiz segmentlarda. Fog qatlam kiritilganda uplink masofasi qisqaradi, marshrutlash soddalashadi, shuningdek, ko‘p hollarda tarmoq yadro qismiga chiqmasdan lokal segmentda xizmat ko‘rsatish yo‘li bilan jitter, ya’ni kechikishning tebranishi kamayadi [3]. Jitterning kamayishi real vaqt tizimlari uchun kechikishning o‘rtacha qiymatidan ham muhimroq bo‘lishi mumkin, chunki boshqaruv konturlari barqarorligi va hodisalarni aniqlash aniqligi ko‘p hollarda maksimal kechikish va uning dispersiyasiga bog‘liq. Fog computingning IoTdagi amaliy samarasi, avvalo, ma’lumot oqimlarini semantik jihatdan ajratish va ularni turli joyda qayta ishlash siyosatini tanlash bilan belgilanadi. Barcha ma’lumotni bir xil prioritetda bulutga uzatish resurs isrofi va kechikishning ortishiga olib keladi. Fog tugunlarida oqimlarni filtrlash, anomaliyalarni erta aniqlash, faqat zarur xususiyatlar yoki agregatsiyalangan ko‘rsatkichlarni yuqori qatlamga yuborish orqali tarmoq trafigi kamayadi; bu esa navbatlar qisqarishi hisobiga kechikishni ham pasaytiradi. Mazkur yondashuvning ilmiy asosida “hisoblashni ma’lumotga yaqinlashtirish” tamoyili yotadi: ma’lumot manbai yaqinida dastlabki tahlil bajarilgan sari, markazlashgan resurslarga talab kamayadi va tizimning masshtablanuvchanligi ortadi [1]. Shu bilan birga, fog qatlamning o‘zi ham 249

resurslari cheklangan bo‘lishi mumkin; demak, kechikishni kamaytirish faqat “yaqinroq” hisoblashga ko‘chirish bilan emas, balki joylashtirish va rejalashtirishning aniq optimallashtirilgan siyosati bilan ta’minlanadi. Kechikishni maqsadli kamaytirish uchun vazifalarni joylashtirish muammosini kechikish byudjeti orqali formalizatsiya qilish maqsadga muvofiq. Har bir IoT ilovasi yoki oqimi uchun ruxsat etilgan maksimal umumiy kechikish L_max berilgan bo‘lsin; bu qiymat ilovaning real vaqt talabidan kelib chiqadi. Umumiy kechikish L ni transport kechikishi L_net, hisoblash kechikishi L_comp va navbat kechikishi L_queue yig‘indisi sifatida baholash mumkin. Fog arxitekturasida L_net odatda kamayadi, biroq L_comp va L_queue fog tugunidagi yuklanishga bog‘liq ravishda oshishi ehtimoli mavjud. Shuning uchun selektiv joylashtirish modeli kerak bo‘ladi: kechikish byudjeti tor bo‘lgan oqimlar birinchi navbatda fog yoki edge tugunlarda bajariladi, kechikish talabchanligi pastroq bo‘lgan, lekin resurs talab qiluvchi tahlillar esa bulutga yo‘naltiriladi. Bu yerda ilmiy yangilik sifatida ta’kidlanadigan jihat shuki, joylashtirish qarori faqat o‘rtacha kechikishga emas, balki kechikishning ishonchlilik mezoniga, ya’ni belgilangan ehtimollik bilan L ≤ L_max shartini qanoatlantirishiga ham tayangan holda qabul qilinishi lozim; chunki IoT tarmoqlarida yuklanish va kanal sifati stoxastik o‘zgaradi [2]. Fog qatlamning ierarxik tabiatidan kelib chiqib, ko‘p bosqichli qayta ishlash zanjiri hosil bo‘ladi: endpoint darajasi tezkor reaksiya uchun eng yaqin aktuatordan foydalanadi, shlyuz darajasi protokollarni moslashtiradi va kesh vazifasini bajaradi, mintaqaviy fog tugunlari oqim tahlili va qoidalar dvijogi kabi xizmatlarni yuritadi, bulut esa uzoq muddatli saqlash va chuqur o‘rganish uchun ishlatiladi. Mazkur zanjirda kechikishning kamayishi nafaqat masofa qisqarishi, balki xizmatlarni “joyida” taqdim etish hisobiga ham sodir bo‘ladi: masalan, autentifikatsiya va avtorizatsiya, konfiguratsiya profillarini tarqatish, lokal katalog xizmatlari fogda bajarilsa, har bir tranzaksiya uchun bulutga so‘rov yuborish ehtiyoji kamayadi. Bunda xavfsizlik va ishonchlilik masalalari ham kechikish bilan bog‘liq: kuchli kriptografik mexanizmlar hisoblash vaqtini oshirishi mumkin, lekin fog tugunlarida apparat tezlatkichlari yoki optimallashtirilgan protokollar qo‘llanilsa, xavfsizlikni susaytirmasdan kechikishni barqarorlashtirish mumkin [4]. Shunday qilib, fog arxitekturasi kechikish bo‘yicha yutuq berishi uchun u xizmatlar orkestratsiyasi, monitoring va moslashuvchan boshqaruv bilan birga ko‘rilishi kerak. IoT ilovalarida kechikishning kamayishi ko‘pincha energiya sarfi va uzatish xarajati bilan ham uyg‘unlashadi. Sensor tugunlarining ko‘pi batareya bilan ishlaydi, ularning uzoq masofaga uzatishi qimmat; fogda agregatsiya va qisqartirish bajarilganda, radioaloqa faoliyati kamayadi va qurilma umrboqiyligi ortadi. Biroq bu yondashuvda hisoblash yukining bir qismi fogga ko‘chadi, fog tugunlarining energiya samaradorligi va sovitish infratuzilmasi masalalari dolzarblashadi. Shu nuqtai nazardan, “energiya-kechikish” kompromissi ilmiy tahlilning muhim obyekti bo‘lib, joylashtirish siyosati faqat minimal kechikish emas, balki minimal energiya sarfi yoki minimal umumiy xarajat funksiyasi 250

bo‘yicha ham moslashtirilishi mumkin [3]. Amaliy tizimlarda ko‘pincha ko‘p mezonli optimallashtirish talab etiladi: masalan, sanoat IoTda xavfsizlik va deterministik kechikish birinchi o‘rinda, aqlli shahar monitoringida esa tarmoq trafigini kamaytirish va xizmat narxini pasaytirish ustuvor bo‘lishi mumkin. Fog computingning kechikishni kamaytirishdagi roli O‘zbekiston sharoitida ham alohida ahamiyatga ega, chunki hududiy tarmoqlar sifatining notekisligi va chekka joylarda magistralga chiqish imkoniyatlarining cheklanishi bulutga to‘liq tayanadigan arxitekturalarda kechikishni keskin oshirishi mumkin. Mahalliy tarmoqlarda fog tugunlarini joriy etish orqali muhim xizmatlarni lokal darajada ushlab turish, tarmoq uzilishlariga bardoshlilikni oshirish hamda ma’lumot suvereniteti talablariga mos ravishda sezgir ma’lumotlarni hudud ichida qayta ishlash mumkin. Shu bilan birga, fogni joriy etishda standartlashtirish, protokollar mosligi, xizmatlarni boshqarish va kadrlar tayyorlash masalalari ham muhim; chunki heterogen qurilmalar muhitida orkestratsiya va monitoring yetarli bo‘lmasa, fog tugunlarining o‘zi navbatlar va qayta yo‘naltirishlar sabab kechikish manbaiga aylanib qolishi ehtimoli bor. Tadqiqotlarda ta’kidlanadigan yana bir jihat shuki, fog arxitekturasining samaradorligi tizim topologiyasi, oqimlar intensivligi va xizmat sinflariga bog‘liq ravishda o‘zgaradi, shuning uchun sinov stendlari va raqamli modellashtirish asosida parametrlarni kalibrlash zarur [5]. Xulosa. Fog computing arxitekturasi IoT tizimlarida kechikishni kamaytirishda transport masofasini qisqartirish, tarmoq trafigini filtrlash va lokal servislarni taqdim etish orqali fundamental ustunlik beradi, biroq bu ustunlik avtomatik tarzda yuzaga chiqmaydi. Eng katta natija kechikish byudjeti asosida selektiv joylashtirish, ierarxik qayta ishlash zanjirini moslashtirish va fog tugunlarida resurslarni monitoring hamda dinamik rejalashtirish orqali boshqarish bilan erishiladi. Fog yondashuvi bulutni to‘liq almashtirmaydi, balki real vaqtga sezgir oqimlar uchun yaqin qatlamda tezkor tahlilni, uzoq muddatli va resurs talabchan vazifalar uchun esa bulutning qudratini uyg‘unlashtirgan gibrid modelni shakllantiradi; natijada IoT ilovalari uchun kechikishning o‘rtacha qiymati va tebranishi pasayadi, xizmat sifati barqarorlashadi va tizimning masshtablanuvchanligi oshadi. Foydalanilgan adabiyotlar ro‘yxati 1. Bonomi F., Milito R., Natarajan P., Zhu J. Fog Computing: A Platform for Internet of Things and Analytics. New York, Springer, 2014. 169 p. 2. Shi W., Cao J., Zhang Q., Li Y., Xu L. Edge Computing: Vision and Challenges. Beijing, Tsinghua University Press, 2018. 286 p. 3. Buyya R., Dastjerdi A.V. Internet of Things: Principles and Paradigms. Boston, Morgan Kaufmann, 2016. 486 p. 4. Иванов А.А., Петров С.В. Периферийные и туманные вычисления в киберфизических системах. Москва, Техносфера, 2020. 312 с. 5. Рахимов Ш.Х., Каримов Б.А. Интернет вещей: архитектура, протоколы и практические решения. Тошкент, Fan va texnologiya, 2021. 240 б. 6. Абдуллаев А.Н., Нурматов Ж.С. Тармоқларда кечикишни баҳолаш ва оптималлаштириш усуллари. Самарқанд, СамДУ нашриёти, 2019. 198 б. 251

MACHINE LEARNING MODELS FOR DOCUMENT SIMILARITY DETECTION USING THEIR SUMMARIES Babadjanova Malika Kamoladdinovna, Abu Rayhon Beruniy nomidagi UrDU, babadjanovamalika060797@gmail.com Sharipov Alisher Romon o‘g‘li, Abu Rayhon Beruniy nomidagi UrDU magistranti, (Data Science), alish20001225@gmail.com Annotatsiya. Mazkur ilmiy ish o‘zbek tilidagi hujjatlar o‘xshashligini ularning qisqa mazmunlari (summaries) asosida aniqlashda qo‘llaniladigan mashinali o‘rganish modellarini tahlil qilinadi. Turli modelalrning afzalliklari, cheklovlari va matnlarni avtomatik taqqoslash jarayonidagi amaliy qo‘llanilishi ko‘rib chiqiladi. Kalit so‘zlar: O‘zbek tili, mashinali o‘rganish, hujjat o‘xshashligi, matn tahlili, vektorlash, NLP Аннотация. В данной статье анализируются модели машинного обучения, используемые для выявления сходства документов на узбекском языке на основе их аннотаций. Рассматриваются преимущества, ограничения и практическое применение различных моделей в процессе автоматического сравнения текста. Ключевые слова: Узбекский язык, машинное обучение, сходство документов, анализ текста, векторизация, NLP Abstract. This paper analyzes machine learning models used to identify document similarities in Uzbek based on their summaries. The advantages, limitations, and practical applications of various models in the process of automatic text comparison are considered. Key words: Uzbek language, machine learning, document similarity, text analysis, vectorization, NLP Introduction. The rapid growth of digital text data has created a strong need for efficient document similarity detection systems. Identifying similar documents is important in many fields such as plagiarism detection, information retrieval, recommendation systems, and academic research management. Traditional keyword-based methods often fail to capture semantic relationships between texts, especially when different wording is used to express similar ideas.Recent advances in natural language processing (NLP) and machine learning have enabled more accurate semantic comparison of documents. Instead of comparing full texts, using document summaries provides a compact and computationally efficient representation of content. Summaries preserve the core meaning of documents while reducing noise and processing cost. Therefore, summary-based similarity detection has become an attractive research direction. Machine learning models can learn hidden patterns in textual data and map documents into numerical vector spaces where similarity can be measured effectively. However, the performance of these models depends on feature representation, algorithm selection, and similarity metrics. The objective of this paper is to analyze classical machine learning approaches for document similarity detection using summaries and evaluate their suitability for practical applications. General pipeline is illustrated in the first figure. 252

Figure 1. General pipeline for summary-based document similarity detection Main body. Document similarity detection typically follows a supervised or unsupervised learning framework. First, document summaries are preprocessed through tokenization, stop-word removal, and normalization. Then, texts are transformed into numerical representations such as TF-IDF vectors or word embeddings. Finally, similarity between documents is computed using distance or similarity measures such as cosine similarity.Logistic Regression can be applied in supervised similarity classification tasks where document pairs are labeled as similar or not similar. It is simple, interpretable, and computationally efficient. However, Logistic Regression assumes linear separability and may struggle with complex semantic relationships in textual data. Support Vector Machines (SVM) are widely used for text classification and similarity detection. By using kernel functions, SVM can capture nonlinear relationships between document vectors. SVM models often achieve high accuracy in text similarity tasks. Nevertheless, they require careful parameter tuning and are less interpretable, which may limit their transparency in some applications. Decision boundaries example is shown in Figure 2. Figure 2. Decision boundaries of Logistic Regression and Support Vector Machine in text similarity space The k-Nearest Neighbors (k-NN) algorithm measures similarity based on distance between document vectors. It is intuitive and easy to implement, making it useful for baseline systems. However, k-NN becomes computationally expensive with large document collections and is sensitive to feature scaling and noise. Decision Tree models provide rule-based similarity classification that is easy to interpret. They help understand which textual features contribute most to similarity decisions. 253

The main limitation of Decision Trees is their tendency to overfit, especially when working with high-dimensional text data. The Decision tree model example is shown in Figure 3. Figure 3. Decision Tree structure and ensemble learning concept Ensemble learning methods, such as Random Forest and Gradient Boosting, combine multiple weak learners to improve prediction accuracy and robustness. Random Forest reduces overfitting through averaging, while Gradient Boosting focuses on correcting previous errors. These methods often outperform single models in document similarity tasks but increase computational complexity and reduce interpretability. In summary-based similarity detection, the quality of summaries significantly affects model performance. Well-generated summaries improve semantic matching, while poor summaries may lead to information loss. Therefore, integrating effective text summarization techniques with machine learning models is essential for building reliable systems. Machine learning models is demonstrated in Table 1. Table 1. Comparison of machine learning models for document similarity detection Conclusion This paper analyzed classical machine learning models for document similarity detection using summaries. The study shows that each algorithm has specific strengths and weaknesses depending on the task requirements. Models such as SVM and ensemble methods generally provide higher accuracy for semantic similarity detection, while Logistic Regression and Decision Trees offer better interpretability and lower computational cost. The choice of model should balance performance, efficiency, and explainability based on the intended application. Future research will focus on experimental evaluation of these models using real-world datasets and on integrating deep learning–based text embeddings to further improve semantic similarity detection. Developing reliable and scalable 254

document similarity systems remains an important direction for modern information retrieval and NLP applications. Literature references 1. Matlatipov, S.G., Rajabov, J., Kuriyozov, E., & Aripov, M. (2024, May). UzABSA: Aspect-Based Sentiment Analysis for the Uzbek Language. In Proceedings of the 3rd Annual Meeting of the Special Interest Group on Under-resourced Languages@ LREC-COLING 2024 (pp. 394-403). 2. Kutlimuratova, B., Kuriyozov, E., Urazbaev, A. & Rakhimova, G. (2025, November). Corpus-Based Error Analysis of Uzbek EFL Learners’ Academic Writing. In 2025 IEEE XVII International Scientific and Technical Conference on Actual Problems of Electronic Instrument Engineering (APEIE) (pp. 1-5). IEEE. 3. Mattiev, J., Davityan, M. & Kavsek, B. (2023). ACMKC: A compact associative classification model using K-modes clustering with rule representations by coverage. Mathematics, 11(18), 3978. 4. Madatov, K., Matlatipov, S. & Aripov, M. (2023). Uzbek text's correspondence with the educational potential of pupils: a case study of the School corpus. arXiv preprint arXiv:2303.00465. TERI SARATONINI DERMOSKOPIK TASVIRLAR ASOSIDA TASNIFLASHDA KO‘RISH TRANSFORMATOR (VIT) MODELLARINING SAMARADORLIGI TAHLILI Boboxonov Axmadxon Xolmirzoxon o‘g‘li, doktorant, Sharof Rashidov nomidagi SamDU b_akhmadkhon@mail.ru Annotatsiya. Sun’iy intellekt (AI) sohasida erishilgan yutuqlar teri saratoniga erta tashxis qo‘yishda muvaffaqiyatga erishdi. Tadqiqotda dermoskopik tasvirlaridan teri saratonini tasniflashning ViTga asoslangan yondashuvlari tahlil qilindi. Tadqiqot natijasida ViT teri saratonini tasniflashda mashinaviy o‘qitish (ML) modellariga nisbatan samarali ekanligi aniqlandi. Kalit so‘zlar: Teri saratoni, Kompyuter ko‘rish (CV), ML, DL, ViT. Аннотация. Достижения в области искусственного интеллекта (ИИ) привели к прогрессу в ранней диагностике рака кожи. В исследовании были проанализированы подходы на основе ViT для классификации рака кожи по дерматоскопическим изображениям. Исследование показало, что ViT оказался более эффективным, чем модели машинного обучения (МО) в классификации рака кожи. Ключевые слова: Рак кожи, Компьютерное зрение (CV), Машинное обучение (ML), Глубокое обучение (DL), Трансформация зрения. Abstract. Advances in artificial intelligence (AI) have led to advances in early diagnosis of skin cancer. The study analyzed ViT-based approaches for classifying skin cancer from dermoscopic images. The study found that ViT was more effective than machine learning (ML) models in classifying skin cancer. Keywords: Skin cancer, Computer Vision (CV), ML, DL, ViT. Kirish. Teri saratoni keng tarqalgan saraton kasalliklaridan biri bo‘lib, uning tarqalishi butun dunyo bo‘ylab yildan-yilga ortib borishi inson salomatligiga katta xavf tug‘dirmoqda. Keng tarqalgan terining xavfli o‘smalari asosan bazal hujayrali karsinoma (BCC), yassi hujayrali karsinoma (SCC) va melanomani o‘z ichiga oladi. BCC va SCC barcha teri xavfli o‘smalarining taxminan 95 %ni tashkil 255

qiladi, melanoma esa nisbatan kam uchraydigan bo‘lsa-da, uning xavflilik darajasi yuqori [1]. Jahon sog‘liqni saqlash tashkilotining hisob-kitoblariga ko‘ra, dunyoda har yili 2-3 million yangi teri saratoni holatlari qayd etiladi. Melanomaning tarqalishi yashirin va metastaz ko‘pincha erta bosqichda sodir bo‘ladi. Saratonning ushbu turi tufayli o‘lim holatlari soni har yili terining xavfli o‘smalaridan kelib chiqadigan umumiy o‘lim holatlarining 65 % –75 %ni tashkil qiladi [2]. Shuning uchun, terining xavfli o‘smalarini erta aniqlash, o‘z vaqtida to‘g‘ri tashxis qo‘yish davolash samaradorligini oshirish va bemorning hayot sifatini yaxshilash uchun juda muhimdir. So‘nggi yillarda dermatoskopiya teri kasalliklarini tashxislash uchun yangi invaziv bo‘lmagan vosita sifatida kirib keldi. Ammo ushbu vosita tibbiyotda qo‘llanilishda shifokorlarning klinik tajribasiga juda bog‘liqligi, inson omilining mavjudligi sababli sub’ektiv hisoblanadi. Sun’iy intellektning (AI) tibbiyotda keng qo‘llanilishi yuqoridagi muammolarni samarali hal qilish uchun yangi yondashuvlarni taqdim etadi. Teri tasvirlarini tahlil qilish uchun AI dan foydalanish orqali tashxis natijalari aniqroq va obyektivroq bo‘ladi. Teri saratonining dastlabki tasnifi an’anaviy mashinaviy o‘qitish (ML) yordamida teri saratoni dermatoskopik tasvirlaridan ajratib olingan shikastlanish shakli, tuzilishi, geometriyasi va boshqa qo‘lda ajratilgan informativ belgilar asosida amalga oshiriladi. Xususan, tadqiqotchilar K-eng yaqin qo‘shni (KNN) va Qo‘llabquvvatlovchi vektor (SVM) algoritmlari teri saratonini tasniflash va tashxislash uchun ishlatilgan [3]. Ushbu tadqiqotda xususiyatlarni ajratib olish uchun tezlashtirilgan mustahkam xususiyatlar va tasniflash uchun kvadratik SVM bilan birlashishi ishlab chiqildi. Shikastlanishni tashxislashda taklif qilingan yondashuv 0,85 s o‘rganish vaqtini, 100 % sezgirlikni, 85,7 % aniqlikni va 60 % o‘ziga xoslikni ko‘rsatadi. Teri tasvirlari asosida saratonni tasniflash va segmentatsiyalashning ishonchliligini oshirish uchun tadqiqotchilar o‘z ishlarida tasodifiy o‘rmon (RF) va qaror daraxtlari (DT) yondashuvlaridan foydalanishdi [4]. Taklif qilingan usul tasvirning fazoviy tafsilotlarini saqlashga yordam beradigan yuqori aniqlikdagi xususiyat xaritalarini yaratishi mumkin. Tadqiqotda assimetriyani tahlil qilish uchun yagona o‘lchov bo‘lib xizmat qilish uchun ko‘plab xususiyatlar o‘lchovlari yagona vektorga birlashtirilgan. Ushbu tadqiqotda ma’lumotlar to‘plamidagi DT algoritmidan foydalangan holda tasniflash samaradorligini oshirish uchun zarrachalar to‘dasini optimallashtirish (PSO) tavsiya etiladi. Ishlash ko‘rsatkichiga ko‘ra, algoritm 92,26 % aniqlikka erishgan. Tadqiqotchilar teri saratoni tasnifida to‘rtta asosiy bosqich mavjudligini ta’kidlashgan, jumladan: (1) oldindan ishlov berish (2) segmentatsiya (3) xususiyatlarni ajratib olish va (4) tasniflash (1-rasm). 1-rasm. Teri saratonini tasniflash jarayoni 256

An’anaviy ML klassifikatorlarining kamchiliklarini kamaytirish uchun tadqiqotchilar chuqur o‘qitish (DL) modellarini taqdim etdilar. DL ML klassifikatorining bir qismi bo‘lib, inson aralashuvi talab qilinmaydigan deyarli barcha turdagi vazifalar uchun ishlatiladi. DL modellari obyektni aniqlash, tasvir va signalni qayta ishlash, nutqni qayta ishlash, til tarjimasi, tasvir segmentatsiyasi, yuzni aniqlash, avtomatik xususiyatlarni ajratib olish, xususiyatlarni tanlash va deyarli barcha turdagi avtomatik qaror qabul qilish vazifalarida qo‘llanilishi mumkin. DL asosidagi modelning eng muhim afzalliklaridan biri shundaki, u kirish sifatida xususiyatlarni taqdim etishni talab qilmaydi, u xususiyatlarni avtomatik ravishda ajratib oladi. Shunday qilib, asosan chuqur neyron tarmoq modellari (DNN) xususiyatlarni ajratib olish vazifasidan qo‘lda va inson aralashuvini olib tashlaydi. Ma’lumki, to‘g‘ri xususiyatlarni ajratib olish asosan katta ma’lumotlar to‘plamida ko‘p vaqt talab qiladigan vazifadir. Hozirgi kunda DL ning rivojlanishi bilan AI texnologiyasi tibbiy tasvirlarni tahlil qilish sohasida katta yutuqlarga erishdi. Tibbiy tasvirni tahlil qilish sohasida konvolyutsion neyron tarmog‘i (CNN) keng qo‘llaniladi va yuqori aniqliklarni taqdim etadi [5]. An’anaviy ML ga asoslangan usullar bilan taqqoslaganda, DL asosidagi usullarning samaradorligi ancha yuqori. Bundan tashqari, DL asosidagi usulning axborotni birlashtirish usuli tibbiy diagnostikada ham katta ahamiyatga ega ekanligi isbotlangan [6]. Biroq, CNN larning tasvir lezyonlarining mahalliy xususiyatlariga bog‘liqligi, uni global kontekstual munosabatlarini modellashtirish uchun yetarli emas, bu esa ularning diffuz lezyonlarni boshqarish qobiliyatini cheklaydi. Ushbu muammoni hal qilish uchun tadqiqotchilar tasniflash samaradorligini oshirish uchun diqqat mexanizmlariga asoslangan ko‘rish transformerlarni (ViT) joriy qilishdi. ViT arxitekturasi tasvirlarni ketma-ket yamoqlarga bo‘lish va diqqat qilish mexanizmlari orqali global bog‘liqliklarni modellashtirish orqali ImageNet tasniflash vazifalarida sof transformator modellarining vizual tasvirlash qobiliyatini birinchi bo‘lib tasdiqladi. Bu yutuq uni tibbiy tasvirlashda qo‘llashni osonlashtirdi. Tadqiqotchilar Soumyya va boshqalar ViTning shikastlanish mintaqalariga e’tiborini yaxshilash uchun yumshoq diqqat mexanizmini joriy qilishdi, natijada global xususiyatlarga bog‘liqlik kamaydi va tasniflash aniqligi oshdi [7]. Zhao va boshqalar samarali xususiyatlarni o‘rganishga erishgan to‘liq transformatorga asoslangan FTN tarmog‘ini taqdim etishdi [8]. Li va boshqalar esa BC-FCU modulidan foydalangan holda ikki tarmoq o‘rtasidagi semantik tafovutlarni bartaraf etgan va ko‘p sinfli tasniflash aniqligini 87,6% ga ko‘targan CNN-transformator ikki tarmoqli modelini yaratdilar [9]. Shuningdek, ko‘plab tadqiqot natijalari transformatorga asoslangan modellar CNNlarning lokal cheklovlarini hal qilishda va uzoq masofali bog‘liqliklarni ushlashda sezilarli afzalliklarga ega ekanligini ko‘rsatadi. Asosiy qism. Vision Transformer (ViT) hozirda kompyuter ko‘rish sohasida eng zamonaviy va turli xil tasvirlarni aniqlash vazifalari uchun keng qo‘llaniladigan CNN larga raqobatbardosh muqobil sifatida paydo bo‘ldi. ViT modellari hisoblash samaradorligi va aniqligi jihatidan hozirgi eng zamonaviy 257

CNN lardan ancha ustun hisoblanadi. ViTlar tasvirlardan obyektni aniqlash, segmentatsiyalash, tasvirni tasniflash va harakatni aniqlash kabi mashhur tasvirni aniqlash vazifalarida keng qo‘llaniladi. Uning g‘oyasi tasvirni kichikroq bloklarga (yamoqlarga) ajratish, har bir blokni token kabi muomala qilish va keyin bu tokenlarni Transformer modeliga kiritishdan iborat. ViTlarda tasvirlar ketma-ketlik sifatida ifodalanadi va tasvir uchun sinf yorliqlari bashorat qilinadi, bu esa modellarga tasvir tuzilishini mustaqil ravishda o‘rganish imkonini beradi [10]. Kirish tasvirlari yamoqlar ketma-ketligi sifatida ko‘rib chiqiladi, bunda har bir yamoq yamoqdagi barcha piksellarning kanallarini birlashtirish va keyin uni kerakli kirish o‘lchamiga chiziqli ravishda proektsiyalash orqali bitta vektorga tekislanadi (2-rasm). 2-rasm. ViT arxitekturasi Quyida ViT arxitekturasini bosqichma-bosqich ko‘rib chiqilgan. 1. Tasvirni yamoqlarga bo‘lish Kirish tasviri sifatida 𝑋 ∈ 𝑅𝐻×𝑊×𝐶 berilgan bo‘lsin, bu yerda, H, W – balandlik va kenglik, C – kanallar soni (odatda 3, RGB). Tasvir P×P o‘lchamli kichik bloklar (yamoqlar) ga bo‘linadi [11-12]. Bloklar soni: 𝐻×𝑊 𝑁= 2 (1) 𝑃 Har bir blok esa quyidagicha ifodalanadi: 𝑥𝑝 ∈ 𝑅𝑃×𝑃×𝐶 (2) 2. Yamoqlarni vektorga tekislash Bloklar vektorga tekislanadi: 2 𝑓𝑙𝑎𝑡 𝑥𝑝 ∈ 𝑅𝑃 ×𝐶 (3) 3. Yassilangan (flatten qilingan) bloklardan past o‘lchamli chiziqli embeddinglar hosil qilish. Har bir blok quyidagicha embedding qilinadi: 𝑓𝑙𝑎𝑡 𝑧0𝑖 = 𝑥𝑝 𝐸 (4) 2 Bu yerda, 𝐸 ∈ 𝑅 (𝑃 𝐶)×𝐷 , D – embedding o‘lchami. Tasniflash uchun maxsus vektor qo‘shiladi: 𝑧0 = [𝑥𝑐𝑙𝑎𝑠𝑠 ; 𝑧01 ; 𝑧02 ; … ; 𝑧0𝑁 ] (5) 4. Pozitsion embeddinglarni qo‘shish Blok joylashuvini saqlash uchun pozitsion embedding qo‘shiladi: 𝑧0 = 𝑧0 + 𝐸𝑝𝑜𝑠 (6) 𝐸𝑝𝑜𝑠 ∈ 𝑅(𝑁+1)×𝐷 (7) 258