12th International Conference on Computer Science, Engineering and Applications (ICCSEA 2022)

November 26 ~ 27, 2022, London, United Kingdom

Accepted Papers


Depression Detection using Machine and Deep Learning Models to Assess Mental Health of Social Media Users

Smita Ghosh1, Sneha Ghosh2, Diptaraj Sen2, Pramita Das3, 1Department of Mathematics and Computer Science, Santa Clara University, California, USA, 2Department of Computer Science and Engineering, University of Engineering andManagement, Kolkata, India, 3Department of Electronics and Communication Engineering, National Institute of Technology Durgapur, Durgapur, India

ABSTRACT

During the COVID-19 pandemic millions of people were af ected due to quarantine and restrictions. With more than half of the worlds population active on social media, people resorted to these platforms as their outlet for emotions. This led to researchers analysing content on social media to detect depression by studying the patterns of content posting. This paper focuses on finding a data-driven metric called ‘Happiness Factor’ of a user to assess their mental health. Various models were trained to classify a post as ‘depressed’. A user’s ‘Happiness Factor’ was calculated based on the nature of their posts. This metric identifies degrees of depression of a user. The results show the ef ectiveness of the classifier in identifying the depression level. Also, a Mental Health Awareness Resource System is proposed which recommends mental health awareness resources to users on their social media interface based on their ‘Happiness Factor’.

KEYWORDS

Depression Detection, Machine Learning, Deep Learning, Universal Sentence Encoder, Social Media.


Query Optimization meets Reinforcement Learning

Enamul Haque, David R. Cheriton School of Computer Science, University of Waterloo, Ontario, Canada

ABSTRACT

Query optimization is one of the most important tasks of relational database management systems that, if improved properly, can help achieve higher performance in terms of time and resources. Configuration tuning of diverse database instances in distributed systems and optimization of query workloads for cloud databases are also important performance indicators to meet demands of the growing user base globally. But these problems are already NP-Hard, and solutions depend mostly on heuristics-based approximation or randomized algorithms. Here, in this work we explore how advances in Reinforcement Learning (RL) is contributing to this branch of computer systems research.

KEYWORDS

Machine Learning, Database, Query Optimization, Reinforcement Learning, Deep Learning.


Prediction of Genetic Disorders using Machine Learning

Sadichchha Naik1, Amisha Panchal1, Disha Nevare1 and Dr. Chhaya Pawar2, 1Student, Department of Computer Engineering, Datta Meghe College of Engineering, Navi Mumbai, India, 2Asst. Professor, Department of Computer Engineering, Datta Meghe College of Engineering, Navi Mumbai, India

ABSTRACT

A genetic disorder is a health condition that is usually caused by mutations in DNA or changes in the number or overall structure of chromosomes. Several types of commonly-known diseases are related to hereditary gene mutations. Genetic testing aids patients in making important decisions in the prevention, treatment, or early detection of hereditary disorders. With increasing population, studies have shown that there has been an exponential increase in the number of genetic disorders. Genetic disorders impact not only the physical health, but also the psychological and social well-being of patients and their families. Genetic disorders have powerful effects on families. Like many chronic conditions, they may require continual attention and lack cures or treatments. Low awareness of the importance of genetic testing contributes to the increase in the incidence of hereditary disorders. Many children succumb to these disorders and it is extremely important that genetic testing be done during pregnancy. In that direction, the project aims to predict Genetic Disorder and Disorder Subclass using a Machine Learning Model trained from a medical dataset. The model being derived out of a predictor and two classifiers, shall predict the presence of genetic disorder and further specify the disorder and disorder subclass, if present.

KEYWORDS

Genetic disorder, Machine Learning, Medical dataset.


Mining Movement Patterns to Separate Rugby Super League Players into Groups

Victor Elijah Adeyemo1,2,3,4 Anna Palczewska1, Ben Jones2,3,4,5,6 Dan Weaving2,4, 1School of Built Environment, Engineering and Computing, Leeds Beckett University, Leeds, UK, 2Carnegie Applied Rugby Research (CARR) Centre, Carnegie School of Sport, Leeds Beckett University, Leeds, UK, 3England Performance Unit, Rugby Football League, Leeds, UK, 4Leeds Rhinos Rugby League Club, Leeds, UK, 5School of Science and Technology, University of New England, Armadale, Australia, 6Division of Exercise Science and Sports Medicine, Department of Human Biology, Rica, Cape Town, South Africa

ABSTRACT

The application of pattern mining algorithms to extract movement/event patterns from sports big data enables the extraction of on-field activities, provides context to such activities and enhances training specificity. Currently, aggregated physical performance indicators (e.g., total distance covered) are used to separate players into positional groups but these indicators do not capture the sequential nature of match activities. As there are various types of pattern mining algorithms, this study aims to identify which one discovers the best set of movement patterns (on-field activities) to separate players into two playing positions and utilize classification algorithms to find the most accurate separation. Three pattern mining frameworks were implemented to extract movement patterns and five machine learning classification algorithms to separate groups of players via a case study (i.e., two Elite Rugby League players playing positions). The pattern mining frameworks are Sequential Movement Pattern-mining (SMP), l-length Closed Contiguous (LCC) and Apriori Closed (APR). Five classifiers were fitted on tabular datasets (whose independent variables are set of movement patterns) to determine which type of movement patterns accurately separates the groups. Extracted “LCC” and “SMP” on-field activities shared a 0.179 Jaccard similarity score (18%) as both are consecutive patterns. Extracted “APR” on-field activities shared no significant similarity with both extracted “LCC” and “SMP” on-field activities because it mined non-consecutive patterns. Multi-layered Perceptron algorithm fitted on the dataset whose independent variables were the extracted “LCC” on-field activities achieved the highest accuracy of 91.02% ± 0.02 and precision, recall and F1 scores of 0.91. Therefore, we recommend the extraction of closed contiguous (consecutive) over non-consecutive patterns for separating groups of players.

KEYWORDS

Pattern Mining, Performance Analysis, Rugby League, Sports Analytics, Machine Learning.


SERP Evaluation and Model Interpretation Based on Behavior Sequence

Xiaoyu Dong and Shen Shen and Jinkang Jia and Yifan Wang, Baidu Inc, Beijing, China

ABSTRACT

SERP (Search Engine Result Page) quality evaluation plays a vital role in industrial practice. With the rapid iterations of search engine, traditional page-level metrics like click ratio, dwell time can no longer evaluate user experience on various templates of results. To promote evaluation accuracy, we implement Transformer to capture the sequential patterns from behavior sequence data. In recent studies, approaches focusing on modeling behavior sequences have emerged. Some studies concentrate on feature engineering by extracting subsequence patterns, others focus on end-to-end deep learning models. While widely used, these two methods both have drawbacks, either a risk of distortion of true subsequence patterns or difficulty for interpretation. Here we implement Transformer to give considerations to both completeness of sequential patterns and model interpretation. To find the best way of modeling behavior sequence data with continuous features, we adopt two embedding methods to predict SERP quality evaluation, and the second one achieves good promotion. What’s more, we develop a novel interpretation method for transformer models and demonstrate its ability to make interpretations for subsequence patterns.

KEYWORDS

Behavior sequence, Transformer, Model Interpretation.


Web Application and Internet Security Threats

Vaibhav Katiyar, University Institute of engineering and technology, CSJMU, Kanpur

ABSTRACT

Computer and network security are one of the most challenging topics in the Information Technology research community. Internet security is a significant subject that may affect a wide range of Internet users. People that use Internet to sell, buy and even to communicate needs their communications to be safe and secure. This paper is discussing the different aspects of Internet and networking security and weakness. Main elements of networking security techniques such as the firewalls, passwords, encryption, authentication and integrity are also discussed in this paper. The anatomy of a web applications attack and the attack techniques are also covered in details. The security of high-speed Internet as the growth of its use has stained the limits of existing network security measures. Therefore, other security defense techniques related to securing of high-speed Internet and computer security in the real world are studied as well such as, DNS, One-Time Password and defending the network as a whole.This paper is also surveyed the worm epidemics in the high- speed networks and their unprecedented rates spread.

KEYWORDS

Web application Security, Network Security, Protection tools, SQL Injection, Firewall, and Intrusion Detection System.


Cloud Oriented Virtual Fragmented Database

Ahmad Shukri Mohd Noor, Nur F. Mat Zain, Rabiei Mamat and Noor hafhizah A. Rahim, Faculty of Ocean Engineering Technology and Informatics, Universiti Malaysia Terengganu

ABSTRACT

Fragmentation is way to replicate the data into multiple servers to increase the availability and accessibility. The existing technique Hierarchical Replication Scheme (HRS), Read One-write-All (ROWA) and Branch Replication Scheme (BRS), these three techniques are the most common techniques that used to replicate the data but they have their own drawbacks such as communication cost that is high. The proposed technique in this research is Mater-slave replica technique which is uses to distribute the data into servers, so it enables the user to access it from any server in case one server fail. This research about creates cloud environment and replicate the data into virtual servers to enables the user to access the data at any location and in any time. VirtualBox used in this research to create windows guested in the hard disk of the local host that used these windows as virtual servers and stored the data of the master replica server. by using three virtual servers, a series of experiments will be performed, and the results will be compared to the existing techniques, then the findings of the experimental will achieve the replication consistency of the database.

KEYWORDS

Virtual Machine, Database fragmentation, Database Replication, Mater-salve replica, Data Availability.


Artificial Intelligence to Analyze Mineral Resource Cost Estimation using Ordinary Kriging and Schlumberger Methods

Paryati1 and Krit Salahddine2, 1Department of Tehchnic Engineering, UPN “Veteran” Yogyakarta, Indonesia, 2Ibn Zohr University, Agadir City, Morocco

ABSTRACT

Artificial intelligence to analyze the estimated cost of mineral resources, mostly done in mining areas, to estimate the mining area which is an element of mineral reserves in nature. So to determine the commodity and the estimated cost for the mining area, a study of andesite estimation and a study of technical factors that influence the development of mining processes and estimates at PT. Varia Usaha Beton Mandalika, East Lombok Regency, West Nusa Tenggara, Indonesia. Thus, with an exploratory study, a description of the condition of andesite mineral commodities will be obtained at the location of the area. One approach used to estimate andesite is to use the Schlumberger method and use ordinary kriging so that it can simplify and speed up the work process, especially when carrying out andesite activities in a very large mining area. Research conducted in the mining area owned by PT. Varia Usaha Beton Mandalika, East Lombok Regency, West Nusa Tenggara, Indonesia. Mining companies use the Schlumberger configuration method and ordinary kriging. The results obtained by andesite mineral resource estimates are very good and relevant and valid. This mining method is used in the production operation process of andesite resource estimation. The selection model is based on the type of commodity and the contours of the mining area in the form of sand and stone. This method has the ability to detect the presence of inhomogeneous rock layers by comparing the apparent resistivity value with changes in the MN/2 electrode distance and is able to detect the presence of inhomogeneous rock layers. So this method is very suitable for estimating andesite resources in mining areas.

KEYWORDS

Artificial Intelligence, Estimation, Schlumberger, Vertical Sounding, Ordinary Kriging.


Agent-based Modeling and Simulation of Complex Industrial Systems: Case Study of the Dehydration Process Noureddine

Seddari1,2, Sohaib Hamioud3, Abdelghani Bouras4, Sara Kerraoui1 and Nesrine Menai1, 1LICUS Laboratory, Department of Computer Science, Université 20 Août 1955-Skikda, Skikda 21000, Algeria, 2LIRE Laboratory, Abdelhamid Mehri-Constantine 2 University, Constantine 25000,Algeria, 3LISCO Laboratory, Computer Science Department, Badji-Mokhtar University,Annaba 23000, Algeria, 4Department of Industrial Engineering, College of Engineering, AlfaisalUniversity,Riyadh 11533, Saudi Arabia

ABSTRACT

Agent-based modeling and simulation (ABMS) is a new approach to modeling autonomous systems and interacting agents. This method is becoming more and more popular for its efficiency and simplicity. It constitutes an approach in the field of modeling complex systems. Indeed, ABMS offers, contrary to other types of simulations, the possibility of directly representing the simulated entities, their behaviors, and their interactions without having to recourse to mathematical equations. This work is a contribution in this sense, the goal is to propose an agent-based model to simulate an industrial system. The latter presents the problem of complexity, which can be mastered with the Multi-Agent Systems (MAS) approach. Our model is validated by a case study of the natural gas dehydration process. The latter is consolidated by a simulation made in the multi-agent platform JADE (Java Agent DEvelopment Framework).

KEYWORDS

Agent-based modeling and simulation (ABMS), Industrial system, Multi-Agent Systems (MAS), Multi-agent platform JADE.


Generic Question Classification for Dialogue Systems

Marine Troadec, Matthis Houl`es, and Philippe Blache, LPL-CNRS, ILCB, Aix-en-Provence, France

ABSTRACT

We present in this paper a new classification approach for identifying questions in dialogue systems. The difficulty in this task is first to be domain-independent, reusable whatever the dialogue application and second to be capable of a real time processing, in order to fit with the needs of reactivity in dialogue systems. The task is then different than that of question classification usually addressed in question answering systems. We propose in this paper a hierarchical classifier in two steps, filtering first question/no-question utterances and second the type of the question. Our method reaches a f-score of 98% for the first step and 97% for the second one, representing the state of the art for this task.

KEYWORDS

Question classification, Dialogue systems, Hierarchical classification.


Global Language Positioning System

Xiaohui Zou1 and Shunpeng Zou2, 1Searle Research Center, No. 100, Renshan Road, Hengqin, Guangdong 519000, China, 2Peng Ge Learning Center,Csanady utca 4/b, 1132 , Budapest, Hungary

ABSTRACT

The aim is to explore a new method of Global Language Positioning System, that is, natural language processing by using GPS sequencing positioning. The method is: first, it is pointed out that Susuer distinguishes language and speech although it is good for Indo -European language, but it is not enough for Chinese; then, in the process of Chinese information processing that can be divided into Yan and Yu as the formal language from Tasky and Calnap, as well as the way of combining a single element set of pure mathematics to combine the metal group set at all levels to build a broad -language formal information processing model that combines Chinese and arithmetic; finally, compare the broad - language bilingual treatment described herein, translated by Google Machines Translate as typical representatives, narrow bilingual processing, such as English -Chinese or Chinese -English machine translation, distinguish between complete matching and partial matching, and completely not matching as three types, and then through human -computer interaction and collaboration methods, the reasons why the centralized inspection is not matched, which provides a reliable basis for the improvement or optimization of the five links of classification, matching, translation, prediction and decision -making. As a result, it not only improved the quality of machine translation significantly, but also provided a brand - new natural language processing methods for further data mining, semantic identification, information extraction, knowledge processing and software modeling, and the improvement of NLP & AI technology quality, by using handling paradigm, namely: Global Language Positioning System. Its significance is: this new paradigm not only practically ef icient and precise, especially Chinese information processing, but also, theoretically, with the existing natural language processing methods, and it is not only the same function which can be made in three aspects of natural language understanding, expert knowledge expression, and software model recognition.

KEYWORDS

Natural Language Processing, Natural Language Understanding, NLP & AI , Susuer, Tasky and Calnap.


The Idea System for Semantic & NLP

Jiawen QIU1 and Yezhen ZHAO1 and Xiaohui ZOU2, 3, 1China Mobile Communications Group Guangdong Co., Ltd. Zhuhai Branch, China, 2Searle Research Center, No. 100, Renshan Road, Zhuhai, Guangdong 519000, China, 3Peking University Alumni Association, Beijing 100871, China

ABSTRACT

The purpose of this paper is to put forward the concept of idea system and try to design its knowledge representation principle on the basis of academic discussion on the formation principle of the conceptual relationship of human knowledge. This method is an idea-based knowledge modeling method. The steps are: first, clearly define the concept as the mapping from sign to meaning; further, clarify that both sign and meaning are just a subclass of idea; finally, determine the separation of the sign relation network and the meaning relation network in the conceptual idea relation network, and modeling and computational processing are unified with the idea relationship network. It is characterized by: the knowledge representation method and principle of the idea relation network. The result is that the idea system, which is a new knowledge modeling tool system, systematically introduces its typical embodiments in combination with word, formula, figure, table, and compares and analyzes the research results of related knowledge graphs. Its significance lies in that: it is significantly different from the traditional method, that is, the knowledge representation model that uses symbols instead of meanings to model. This paper establishes a new method and new principle of knowledge construction model that is assisted by symbols and directly modeled with ideas, which is a new method for cognitive computing. It provides a new way to express knowledge based on meaning rather than symbols.

KEYWORDS

Semantic & NLP, Idea System, Human-Machine Collaboration, Knowledge Graph, Knowledge Modeling, Artificial.


Dynamic aspect ratio

Njal Borch 1 and Ingar Arntzen, NORCE Norwegian Research Institute

ABSTRACT

Media is to a large extent consumed on devices that have non-standard aspect ratios, both physically or while rendering content.For example, social media platforms often favour 1:1 ratios, TVs 16:9, iPad tablets 4:3 or 3:4, most Androids 16:9 or 9:16, PCs 16:9 or 16:10 and web pages tend to use responsive design and can therefore have almost any aspect ratio In order to ensure good experiences, it is often therefore necessary to create multiple versions of content, where the content is cropped to a more suitable format. Creating multiple encoded version of the content is a static solution though, and there are good reasons for solving this dynamically on the client side. In this paper we make the case for a client side dynamic aspect ratio solution, present work on implementation and experimentation, and finally provide some insights into how such a system could be implemented and provided in real world systems. Our solution was tested on a few different pieces of content from NRK, both drama series and typical TV debates.

KEYWORDS

Dynamic aspect ratio, Focus track, Multi-device, client side, AI analysis.


PRAGAN: Progressive Recurrent Attention Gan with Pre-Trained Vit Discriminator for Single Image Deraining

Bingcai Wei, Liye Zhang and Di Wang

ABSTRACT

Images collected under bad weather conditions are not conducive to the development of visual tasks. To solve this problem, it is a trend to use a variety of neural networks. The ingenious integration of network structures can make full use of the powerful representation and fitting ability of deep learning to complete low-level visual tasks. In this study, we propose a generative adversarial network(GAN) containing multiple attention mechanisms for the image deraining task. Firstly, to our best knowledge, we propose a pre-trained vision transformer(ViT) which is used for the discriminator in our GAN for single image deraining. Secondly, we propose a neural network training method, which can use a small amount of data for training while maintaining promising results and reliable vision quality. A large number of experiments have proved the correctness and effectiveness of our methods.

KEYWORDS

Deep Learning, Image deraining, Neural Network, Vision Transformer.


Deep Learning for the Classification of the Injunction in French Oral Utterances

Asma BOUGRINE, Philippe RAVIER, Abdenour HACINE-GHARBI and Hanane OUACHOUR, Prisme Laboratory, University of Orleans, 12 rue de Blois, 45067 Orléans, France

ABSTRACT

The classification of the injunction in French oral speech is a difficult task. The logarithmic energy can be a good indicator and our goal is to validate the predominance of this prosodic feature using SVM and K-NN. When applied to our first dataset, this feature indeed showed the best classification rates (CR) of 82% for SVM and 71.42% for K-NN. However, the energy was not the relevant feature when applied to our second, heterogeneous and wild, dataset. In order to improve the classification rates, we applied LSTM and CNN networks. With LSTM, a CR=96.15% was found using 6 prosodic features with the first dataset against 64% with the second dataset. The CNN, a network capable of automatically extracting the most relevant features, gave a better result on the second dataset with CR=73% largely exceeding SVM, K-NN and LSTM.

KEYWORDS

Injunction classification, prosodic features, CNN, LSTM, wild oral dataset.


Fruit Type Classification using Deep Learning and Feature Fusion

Harmandeep Singh Gill1, Baljit Singh Khehra2, 1Principal, Mata Gujri Khalsa College(Jalandhar) Punjab-144801, 2Principal, BAM Khalsa College, Garhshankar (Hoshiarpur) Punjab

ABSTRACT

Machine and deep learning applications play a dominant role in the current scenario in the agriculture sector. To date, the classification of fruits using image features has attained the researcher’s attraction very much from the last few years. Fruit recognition and classification is an ill-posed problem due to the heterogeneous nature of fruits. In the proposed work, Convolution neural network (CNN), Recurrent Neural Network (RNN), and Long-short Term Memory (LSTM) deep learning methods are used to extract the optimal image features, and to select features after extraction, and finally, use extracted image features to classify the fruits. To evaluate the performance of the proposed approach, the Support vector machine (SVM) unsupervised learning method, Artificial neuro-fuzzy inference system (ANFIS), and Feed-forward neural network (FFNN) classification results are compared, and observed that the proposed fruit classification approach results are quite efficient and promising.

KEYWORDS

Image classification, Feature Extraction, Deep Learning, Feature Fusion, CNN, RNN, LSTM.


A Soft System Methodology Approach to Stakeholder Engagement in Water Sensitive Urban

Lina Ntomene Lukusa and Ulrike Rivett

ABSTRACT

Poor water management can increase the extreme pressure already faced by water scarcity. Water quality and quantity will continue to degrade unless water management is addressed holistically. A holistic approach to water management named Water Sensitive Urban Design (WSUD) has thus been created to facilitate the effective management of water. Traditionally, water management has employed a linear design approach, while WSUD requires a systematic, cyclical approach. In simple terms, WSUD assumes that everything is connected. Hence, it is critical for different stakeholders involved in WSUD to engage and reach a consensus on a solution. However, many stakeholders in WSUD have conflicting interests. Using the soft system methodology (SSM), developed by Peter Checkland, as a problem-solving method, decision-makers can understand this problematic situation from different world views. The SSM addresses ill and complex challenging situations involving human activities in a complex structured scenario. This paper demo nstrates how SSM can be applied to understand the complexity of stakeholder engagement in WSUD. The paper concludes that SSM is an adequate solution to understand a complex problem better and propose efficient solutions.

KEYWORDS

Co-design, ICT Platform, Soft Systems Methodology, Water Sensitive Urban Design.


A Soft System Methodology Approach to Stakeholder Engagement in Water Sensitive Urban

Lina Ntomene Lukusa and Ulrike Rivett

ABSTRACT

Poor water management can increase the extreme pressure already faced by water scarcity. Water quality and quantity will continue to degrade unless water management is addressed holistically. A holistic approach to water management named Water Sensitive Urban Design (WSUD) has thus been created to facilitate the effective management of water. Traditionally, water management has employed a linear design approach, while WSUD requires a systematic, cyclical approach. In simple terms, WSUD assumes that everything is connected. Hence, it is critical for different stakeholders involved in WSUD to engage and reach a consensus on a solution. However, many stakeholders in WSUD have conflicting interests. Using the soft system methodology (SSM), developed by Peter Checkland, as a problem-solving method, decision-makers can understand this problematic situation from different world views. The SSM addresses ill and complex challenging situations involving human activities in a complex structured scenario. This paper demo nstrates how SSM can be applied to understand the complexity of stakeholder engagement in WSUD. The paper concludes that SSM is an adequate solution to understand a complex problem better and propose efficient solutions.

KEYWORDS

Co-design, ICT Platform, Soft Systems Methodology, Water Sensitive Urban Design.


Evaluation of Machines Learning Algorithms in Detection and Mitigation of DNS-Based Phishing Attacks

Kambey L. Kisambu1 and Eng. Gilberto Makero2, 1Msc, Cyber Security, Department of Computer Science, University of Dodoma, 2Network Engineer, e-Government Authority of Tanzania

ABSTRACT

DNS-based phishing attacks are among the major threats to Internet users that are difficult to defend against because they do not appear to be malicious in nature.Users have been primarytargets for these attacks that aim to steal sensitive information. The DNS protocol is one ofthe approaches that adversaries use to transfer stolen data outside the institutions networkusing various forms of DNS tunneling attacks. This study deals with evaluation of ML algorithms indetection of DNS-based phishing attacks for securing networks.It deeply evaluate the efficacy of the algorithms when integrated with other solution. The main classifiers used such as SVM, KNN, Logistic Regression and Naïve Bayes were evaluated using performance metrics namelyaccuracy, precision, recall and f-score. Based on the findings, the study proposed improvement for securing systems and networks against DNS-based phishingattacks using the best performing ML algorithm to keep pace with attacks evolution.

KEYWORDS

Malware, DNS-based, Phishing attacks, Machine learning, algorithms, DNS filters, DNS Poisoning, Detection, Mitigation techniques.


Radio Map Construction based on BERT for Fingerprint-based Indoor Positioning System

Zhuang Wang, Liye Zhang, Qun Kong, Cong Liu, Aikui Tian, Computer Science and Technology, Shandong University of Technology, Zibo, China

ABSTRACT

Due to the heavy workload of RSS collection and the complex indoor environment, the WLAN signal is easy to disappear. Therefore, it is time-consuming and laborious to build a WLAN fingerprint indoor positioning system based on Radio Map. To quickly deploy an indoor WLAN positioning system, the Bidirectional Encoder Representation from Transformers (BERT) model is used to populate the missing signal in the Radio Map to quickly build the Radio Map. Because the number of input data in the BERT model cannot exceed 512 words. Therefore, we redefine the model structure based on the original BERT model and fill the missing signals in the radio map in parallel. In addition, the sum of each fragment loss function is defined as the total loss function. The experimental results show that using the improved BERT model to fill the missing signals in Radio Map has higher accuracy and shorter time.

KEYWORDS

Indoor Positioning System, WLAN, Radio Map, BERT Model, Loss Function.