TR2023006691A2 - WORD AND GRAPH BASED DISTRIBUTED EMBEDDING SYSTEM AND METHOD BASED ON AGING FACTOR AND VISIT FREQUENCY IN DIGITAL FOOTPRINTS FOR USER BEHAVIOR PREDICTION - Google Patents

Abstract

Innovation, digital (online/mobile) banking and finance, e-commerce, e-health, etc. It is about a system and a method related to the system that predicts the next behavior of the customer, makes suggestions and makes decisions in a distributed manner, by developing an effective embedding methodology that can capture the navigational behavior of customers or social media application end users (100). The invention aims to gain insight into customer/end user (100) preferences and future financial, healthcare, retail, etc. It will help predict product choices or interests.

Description

DESCRIPTION WORD AND GRAPH BASED DISTRIBUTED EMBEDDING SYSTEM AND METHOD BASED ON AGING FACTOR AND VISIT FREQUENCY IN DIGITAL FOOTPRINTS FOR USER BEHAVIOR PREDICTION Technical Field Invention, digital (online/mobile) banking and finance, e-commerce, e-health, etc. By developing an effective embedding methodology that can capture the navigational behavior of customers or social media application end users, a system that predicts the next behavior of the customer in a distributed manner, makes suggestions and makes decisions, and a method related to the system. State of the Art Digital banking, finance, e-health, e-commerce and social media applications have significantly improved the customer/end user experience by providing ease and efficiency of transacting and communicating without the need to go to a branch, hospital or store or meet with others face to face. Additionally, customers/end users generate valuable data as they browse and transact on these platforms. To make sense of this data, it is crucial to transform and analyze these customer footprints into structured datasets. Banks and financial institutions, hospitals and healthcare institutions, e-commerce and social media sites can better understand customers' behaviors and tendencies towards different banking and finance, e-commerce, e-health products and predict their next moves by taking advantage of their digital banking footprints. and can use this valuable information to benefit customers, a bank or financial institution, a healthcare institution, an e-commerce store or marketing organization. However, digital banking and finance, e-commerce, e-health, social media, etc. Compact and dynamic representations of customers are needed to better predict customer activities based on the time-based dynamics of their behavior in applications. Embedding techniques offer a solution to this problem by digitizing customer transactions and transforming their behavior into numerical representations that can accurately predict their future intentions [1]. By using embedding techniques, it is possible to make sense of customers' navigation in digital applications and predict their future actions, thus providing a more personalized customer/end user experience. Word embedding techniques are widely used in Natural Language Processing (NLP) and text processing [2, 3]. Word embedding is the representation of words and phrases as vectors. However, there is a paucity of comparative analyzes of embedding methods in digital footprint studies, and embedding methods that consider time-based datasets are inadequate. This patent proposal addresses this lack of techniques that can accurately represent customers' circulatory footprints, taking into account both time and frequency dimensions. To achieve this, a system is proposed that includes a new methodology for creating customer-based representations using the time-based and frequency-based dynamics of the customer's circulatory behavior in digital applications. Word2Vec [5], DeepWaIk [4], and StrucZVec are dependent on varying financial, health, etc. It accurately models needs, interests and investment intentions. In burial screenings, aging and visitation frequency factors are addressed. Embedding techniques have developed over time. It has branched into subfields such as word embedding, which includes Word2Vec and Glove, and graph embedding, which includes DeepWaIk, Graph2vec, Node2Vec and Struct2vec [4, 5, 7-10]. Word embedding is a powerful tool used in natural language processing that provides vector representation of words in a corpus, capturing both semantic and syntactic meaning [2]. This technique provides similarity analysis, speech tagging, clustering and syntactic parsing in NLP applications [6]. Word embedding has facilitated numerous analyzes and predictions in supervised, semi-supervised and unsupervised machine learning algorithms by converting words into vectors [2]. The Word2Vec model consists of two approaches: Continuous Bag of Words (CBoW) and skip-gram, which predicts a specific word by transforming it and feeding it into a neural network, using the predicted word as output. In contrast, skip-gram predicts the neighboring words around a given word, with the predicted neighbor words as output. Graph embedding, unlike word embedding, is a technique that uses graphs to represent nodes as vectors in a low-dimensional space. It includes decomposition techniques that are useful in a variety of applications, including graph-based embedding, node classification, and community detection [14]. Graph embeddings are powerful techniques that effectively solve problems related to social networks and community detection. These methods help create vectorized features for graph attributes such as nodes, edges, and subgraphs [15]. Some common graph embedding methods are: Node2vec, DeepWalk, Struchec, Graph Convolutional Networks (GCNs), GraphSAGE, LINE and TransE. Node2vec generates random walks on a graph and uses a skip-gram model to learn node embeddings from the generated walks. Similarly, DeepWalk generates random walks on a graph, but uses a neural network to learn node embeddings instead of a skip-gram model. GCNs use a convolutional neural network (CNN) architecture to learn node embeddings by collecting features of neighboring nodes. GraphSAGE extends the GCN idea by using a differentiable aggregator function to combine properties of neighboring nodes. While LINE preserves local and global network structures by optimizing the first- and second-order proximity of nodes, TransE is a method for embedding directed graphs that uses a translation-based scoring function to learn embeddings. DeepWalk is a recently developed popular and effective graph embedding technique [14]. This technique maintains a higher degree of closeness by maximizing the probability of tracking the last node and subsequent nodes in each random walk [16]. It works in two stages: The first stage involves generating random walks in the graph to optimize the embedded representations of nodes by randomly walking around neighbors [17]. In the second stage, the Word2Vec jump-gram algorithm is applied to each generated walk to vectorize the nodes. DeepWalk is applied to successful word embedding to adjust nodes such that the frequency of occurrence of pairs is preserved in short random walks [14]. There are different implementations of DeepWalk graph representations, such as Hierarchical Softmax, which uses the Huffman binary tree as an alternative representation of the vocabulary [18]. DeepWalk has been used successfully in many applications such as community detection, link prediction and recommendation systems [14, 19]. Struc2Vec is a graph embedding method that has attracted attention recently [10]. Struc2Vec is designed to capture structural information of a graph by exploiting node-to-node structural similarity. The basic idea behind Struc2Vec is to define a set of context nodes for each node in the graph based on the structural similarity between nodes. These context nodes are then used to train a skip-gram model to learn node embeddings. To calculate the context nodes of a given node, Struc2Vec uses a random walk strategy. Struc2Vec performs multiple random walks of fixed length, starting from each node in the graph. The resulting random walks form a context tree for each node. The context tree represents the structural similarity of the node and its neighbors based on observed random walks. Once context trees are created, Struc2Vec uses them to create context nodes for each node in the graph. Context nodes are nodes that are most similar to a given node based on context trees. Specifically, context nodes are defined as the nodes that are most similar to the given node based on the paths shared between the relevant context trees. The resulting context nodes are then used to train a skip gram model to learn node embeddings. The skip gram model is trained to predict context nodes given the target node, and the learned embeddings are used as final node representations. Overall, Struc2Vec is a powerful method for graph embedding that is particularly effective for capturing structural information of a graph [10]. Various embedding methodologies have been proposed and examined in the literature [19]. There are studies that use word-based embedding techniques [22-25] to understand customer behavior and create graphical customer interface test scripts. Moreover, pattern-based embedding approaches [20, 21] are widely used to analyze customers' browsing behavior. Some of the applications in the literature are summarized below. It relates to an authentication system for digital information embedded in the bit stream. It embeds only for authentication purposes and does not predict the next action by analyzing customer behavior. The word embedding neural network provides a mechanism for language models. The mechanism configures the data processing system with an external word-embedded neural network language model that accepts a string of words as input and predicts a valid word based on the word string. However, it does not predict the next step the customer will take by taking into account time-sensitive data. US patent U89659560B2 is about software that trains an artificial neural network to create vector representations for natural language text. This software takes a set of natural language texts for use in education or produces a set of metadata using supervised/unsupervised learning methods. However, the proposed software does not analyze customer behavior and predict the next step. It includes methods and systems for embedding n-grams from the text into a latent space, embedding the input text into the latent space based on the embedded n-grams, and weighting said n-grams according to the spatial evidence of the relevant n-grams. However, it only addresses document classification and does not predict the next step by analyzing customer behavior. US patent number US7620565B2 relates to a system that provides personalized products or services. The system collects information from the user, analyzes it, and utilizes new pre-programmed interactions from the vendor for future interactions with the user. Here, a customized service is offered to the customer, but the future interaction options in the method used have static values. However, it is not possible to predict the next step the customer will take by analyzing time-sensitive data with artificial intelligence techniques. US patent number US10489792B2 relates to a system that maintains the service quality level of messages sent to customers by the customer representative. Messages are modified by word embedding with a neural network and delivered to the customer. However, only the improvement of the messages delivered to the customer is discussed here and it does not predict the action that will take place in the next step by analyzing customer behavior. US patent number US10909459B2 is about training a neural network with a set of documents to create an embedded space. However, using time-sensitive data, the next step the customer will take is not predicted. Additionally, a distributed embedding method or system cannot be recommended. and techniques for creating a customized knowledge graph for a particular knowledge domain (for example, for a particular customer or a particular bot) based on large scale. Knowledge graphs, e.g., search, question answering, conversational interfaces (e.g., chatbots), recommendation systems, etc. Can be used in areas. Customized knowledge graphs can be used, for example, to improve intent classification in a chatbot based on knowledge graph embedding techniques. However, using time-sensitive data, the next step the customer will take is not predicted. Additionally, a distributed embedding method or system cannot be recommended. US patent number US11216510B2 is about processing the text of an incomplete message entered by the user using a word embedding technique with a neural network to suggest messages similar to the message the user is entering. However, only the next message content is predicted and the next action the customer will take is not predicted using time-sensitive data. In the US patent numbered US10650311BZ, messages in the message exchange between users can be processed to recommend a resource to one of the users. For example, the initial user might be a customer representative assisting a customer, and the suggested resources might include the text of a message to be sent to the customer. Resources can be suggested by calculating the semantic representation of messages in the session, calculating a context vector that defines the context of the session, calculating a context hash vector from the context vector, and retrieving one or more resources from a data store. The proposed structure includes artificial intelligence-based word or source embedding methods. However, using time-sensitive data, the next step the customer will take is not predicted. A method for generating a product recommendation based on zero shot learning and multimodal knowledge graph representation is described. Relationships between multiple products and customers are created and stored in a knowledge graph. A representation of the nature and relationship between multiple products and customers stored in the knowledge graph is learned and used to make product recommendations to customers. However, the described system only recommends products and does not predict the next step the customer will take using time-sensitive data. US patent number US11328221B2 relates to a system that classifies short text with unbalanced data. Applies the weights of the regression layer to the text embedding vector to create an embedding vector representing a text sample and a data model output vector. The data model output vectors are assigned a specific classification label to the text instance based on its weight value. So the patent is only about short text classification and using time sensitive data the client's next action is not predicted. It is about collecting experience analytics data and creating a customer model using techniques such as deep learning of customer behavior according to the collected customer experience analytics data, thus predicting customer behavior. By predicting customer behavior, the customer's future interaction can be used by the customer representative to take appropriate action. However, time sensitive data is not used to predict the customer's next step. In the US patent number USO10755293B2, customer journey prediction and resolution is accomplished through a predictive model where each user is mapped to all available user journey information corresponding to a particular job. The predictive model is analyzed to understand the user's characteristics and preferences regarding the services he or she has subscribed to, and the lowest effort solution for the user. The predictive model is analyzed to predict personalized services. The system automatically optimizes itself through continuous feedback and performance improvement. However, time sensitive data is not used to predict the customer's next step. As a result, due to the negativities described above and the inadequacy of existing solutions on the subject, it has been deemed necessary to make an improvement in the relevant technical field. Purpose of the Invention Invention is used in digital banking and finance, e-commerce, e-health, etc. It aims to develop an effective embedding methodology that can capture the navigational behavior of customers or social media application end users. The invention describes a system and method that predicts the next behavior of the customer, makes suggestions and makes decisions in a distributed manner. Allows us to gain insight into customer/end user preferences and future financial, healthcare, retail, etc. It allows predicting product choices or interests. For this purpose, banking and finance, retail, healthcare, social media, etc. A system that optimizes the use of customers' digital footprints for multiple forecasting tasks in the system is proposed. Banking and finance, e-commerce, e-health, etc. digital banking and finance, e-commerce, social media, etc. to accurately represent their customers or social media users. A new embedding system is described that takes advantage of the temporal and repetitive dynamics of circulatory footprints created in applications. The system uses state-of-the-art embedding algorithms such as Word2Vec, DeepWalk and Struchec and combines the aging factor and overall customer/end-user trend based on visit frequency to improve the generated embedding impressions. Thus, it is possible to make accurate predictions about customers' intentions, even when limited transaction history and contextual data are available. However, the focus is mainly on time-sensitive graph-based embedding strategies to better understand customer behavior. The proposed system shows that customer placement methods are more efficient when applied at lower dimensions and can be used to increase the accuracy of customer behavior prediction by capturing time dependence and repetitiveness of customer behavior by using footprints in digital banking. The invention provides solutions to the following technical problems: 1) How can we design an embedding framework to better represent customers' digital footprints? How can we use factors like visit frequency to provide better embedding representation of customer behavior? 2) How can we extend data representations of customers' digital footprints generated by state-of-the-art algorithms such as Word2Vec, DeepWalk, and Struchec to leverage factors such as time dependence to provide better embedding representation to represent customer behavior? 3) How can we understand the performance of the proposed embedding framework by comparing and contrasting it with existing embedding methodologies using traditional machine learning algorithms? How can we increase the number of labeled navigational browser datasets that can be used to train machine learning algorithms? 4) How can we evaluate the effectiveness of the proposed embedding framework for predicting customers' financial behavior, including their likelihood of purchasing certain financial products or participating in certain transactions? The structural and characteristic features and all the advantages of the invention will be understood more clearly thanks to the figures given below and the detailed explanation written by making references to these figures. Description of the Figures Figure 1 shows the flow diagram of the method of the invention. Figure 2 shows the high-level computing system architecture of the word and graph-based embedding system based on the aging factor and visit frequency in the digital footprints of the invention. Figure 3 shows the detailed architecture of the core cloud server in the system subject to the invention. Drawings do not necessarily need to be scaled and details that are not necessary to understand the present invention may be omitted. Description of Part References Retrieving user interaction data 2. Data cleaning, data labeling, training set creation and customer navigation behavior extraction 3. Converting customer navigation behavior into user movement sequence Applying graph embedding and word embedding on pre-processed data Customer navigation by applying aging and normalization based on visit frequency Creating embedding vectors Applying frequency and time-based normalization Applying an aging-based transformation to the embedding data set representing the customer's navigational behavior and creating an embedding vector for each customer Dividing the labeled embedding data set into training and test sets Customer intent created with machine learning algorithms Training the training dataset using embedding vectors. Using the created machine learning models to predict the next navigation behavior of new users 100. User 101. User device 102. Edge server 103. Core cloud server 104. Core network 200. Communication module 201. Word embedding database 2011. Word embedding models 202 . Graph embedding database 2021. Graph embedding models 203. Aging model database 2031. Aging models 204. Memory unit 2041. Device data 2042. User data 2043. User behavior prediction data 2044. Training data 2045. Test data 205. Edge server database 2051 . Edge server configuration data 206. Word embedding module 207. Graph embedding module 208. Justification module 209. Decision support system 210. recommendation system 211. Data analytics module 212. Rule engine 213. Validation module 214. Query module 215. Processing module 216 Sending device 217. Receiving device 218. End server management module 219. Management module Detailed Description of the Invention In this detailed description, the preferred embodiments of the invention are explained only for a better understanding of the subject and without creating any limiting effect. The invention involves various stages, including client embedding framework, data pre-processing, embedding, post-processing and prediction. Figure 1* shows the flow diagram of the invention. The framework takes digital footprints of customers obtained from mobile application platforms as input. The raw data is cleaned by removing irrelevant features by going through the pre-processing phase. The cleansed data is then fed into the embedding phase, where an appropriate embedding technique is applied to represent customer behavior in a high-dimensional vector space. Post-processing, the embeddings are further processed to eliminate noise and improve embedding quality. Finally, the resulting embeddings are used in the forecasting phase to predict customer behavior and preferences. Using this framework, customers' digital footprints can be better represented, gaining insight into their behavior that can be used to improve the banking experience and develop personalized financial products. The framework consists of multiple stages (Figure 1). The processes that occur in these stages are described below. During the preprocessing phase, login, logout, and rarely visited pages that customers visit by default are eliminated. The time customers spend on each page is calculated. When working with smaller labeled datasets in supervised machine learning, one of the techniques that can be used is data augmentation. This involves creating new training samples from existing ones by applying various transformations. In this work, we use one of the generative deep learning approaches to generate secret shopper circulatory behaviors, as shown in Figure 1. In the case of data preprocessing, the proposed framework transforms customer navigation behaviors from log files into user action sequences. We then apply an embedding approach to convert the user strings into an input that can be fed into the LSTM algorithm to create an LSTM-based model. We then use this model to create new arrays. These newly created arrays are checked against the initial real user arrays. Only sequences that differ from actual user action sequences and are possible considering digital banking mobile application user interfaces are filtered as final output. Finally, the original and newly created user action sequences become the output of the data preprocessing phase. In the embedding phase, preprocessed data goes through word embedding and graph embedding data representation methods. The word embedding method used in this study is Word2Vec (CBoW), and the graph embedding methods used are DeepWalk and StrucZVec algorithms. The post-processing phase involves creating customer embedding vectors using normalization techniques based on aging and visit frequency. Within the scope of the invention, DeepWalk, Struct2vec, etc., which embed algorithms to capture the trends of mobile banking customers towards certain products. An embedding system that uses graph embedding and word embedding methods together is proposed. The system uses the time spent by customers on a digital (mobile/web) application or pages and the frequency of visits of each customer/user's pages as input. In the post-processing phase, a frequency and time-based normalization approach is used. The frequency-based approach involves calculating a frequency weight for each page based on the frequency of visits by each customer. This approach gives more importance to the most visited pages. In contrast, the inverse frequency-based approach gives less importance to the most visited pages by taking the inverse of the frequency of the most visited pages by each customer. In general, the proposed approach aims to better represent customers' digital footprints and provide more accurate prediction of their behavior in mobile banking applications. The time-based embedding transformation approach considers time-sensitive digital footprints to better capture the circulatory behavior of mobile banking customers. By considering the total time each customer spends on relevant pages while browsing, this normalization technique can provide more precise information about customer behavior. Once the weight vectors are calculated for both frequency-based and time-based embedding transforms, they are further normalized by multiplication with the weight vectors. This process ensures that the resulting embedding vectors are appropriately weighted and optimized for accurate prediction tasks. In the aging phase, more attention is given to the final circulatory behaviors. Using this approach, an aging-based transformation is applied to embed datasets representing customer browsing behavior. For customer-based sequence vectors, the proposed module aggregates session embedding vectors to create an embedding vector for each customer. This module considers all sessions belonging to a client and applies a aging coefficient to give more weight to recent sessions while reducing the weight of older sessions. Additionally, to further investigate the effectiveness of the aging-based embedding approach, the overall navigation behavior of customers is calculated by averaging the session embedding vectors. This method creates customer embedding vectors by taking the arithmetic average of all session embedding vectors that represent the general behavior of each customer in the mobile banking application. DeepWaIk, Stru02vec, etc. Session embedding vectors are constructed using algorithms and aging and frequency-based normalization techniques are applied. Then, customer-based vectors are calculated to represent each customer with an embedding vector. Customer embedding vectors represent the general browsing behavior of the customer. This is done by calculating the average of all session embedding vectors for each client, resulting in an n-dimensional client-based vector, where n is a framework parameter that determines the size of the embedding vector. The customer-based vectors obtained through this process are used as input for clustering algorithms to segment customers based on their circulatory behavior. In the training phase, the labeled embedding dataset is divided into training and testing sets. It is trained on the training dataset using customer embedding vectors created to predict the customer's intent with machine learning algorithms. Isolation Forest and XGboost, etc. to evaluate the performance of the framework. machine learning algorithms are used. Isolation Forest is a machine learning technique that is effective in detecting anomalies by assuming that samples moving away from the central data set are abnormal. It uses random sampling to create binary trees and these are summed to create a model for a given data set. The main purpose of the Isolation Forest is to identify unusual samples that deviate from most of the data. Since the dataset used in this study is unbalanced, Isolation Forest was chosen as the machine learning algorithm to test the proposed framework [36]. XGboost, a supervised machine learning algorithm, has also been used in the framework. XGboost creates accurate models by using boosting and creating many models in turn, with each model trying to address the weaknesses of the previous one. Tree hardening is achieved by adding each model to the ensemble as a decision tree. XGboost is widely used in various industries due to its high accuracy and speed in large-scale and complex data analysis [37]. In this stage, labeled embedding datasets are trained on a training dataset (i.e. the training dataset to create learning models of the data. These models are then validated using the test dataset. The output of this phase is the creation of machine learning models. In the prediction phase, the created machine learning models are used to predict the next browsing behavior of new users. Clickstream data may be used to analyze customer behavior. These may include the following online activities of digital banking customers: 0 Pages visited during the session, page rank and time spent on pages viewed, 0 Financial transactions, e.g. loan applications, account openings, etc., o Unique customer number, 0. Session date and time, 0 Whether the customer performed the target event. In the proposed system, the data preprocessing phase consists of several steps. In addition, low frequency pages that are rarely visited by customers are removed from the clickstream dataset. Once these pages are removed, the duration of each page in each session is calculated. If there are missing values for page times, the average page time for sessions is calculated and used to replace null values. During the embedding phase, Word2Vec CBoW, DeepWalk, Struchec, etc. are used to create session-based representations. embedding methodologies are used. Word2Vec CBoW is a widely used word embedding methodology, while DeepWalk and Struchec are graph embedding methodology. These methodologies effectively capture customer behavior and intent. New frequency and time-based normalization techniques can be used at a post-processing stage to improve the generated embedding vectors. Specifically, page rank data can be used to calculate the frequency of each page a customer visits to create frequency-based normalization. Additionally, page times can be used to calculate client-based time duration normalizations for embedding vectors. The invention mentioned in this patent proposal may be or include a system, method and/or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions thereon to cause a processor to perform features of the present invention. The computer-readable storage medium may be a tangible device capable of storing and storing instructions for use by an instruction execution device. The computer-readable storage medium may be, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. A non-exhaustive list of more specific examples of computer-readable storage media includes: a portable computer floppy disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or A device such as flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital versatile disk (DVD), memory stick, floppy disk, and any suitable combination thereof. As used herein, a computer-readable storage medium is a computer-readable storage medium, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (for example, pulses of light passing through a fiber optic cable), or electrical signals transmitted through a wire. They should not be interpreted as temporary signals per se. The computer-readable program instructions described herein may be downloaded from a computer-readable storage medium to associated computing/processing devices or sent over a wireless network, such as the Internet, a local area network, a wide area network, and an external computer or external storage device. The network may include copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers, and/or edge servers. A network adapter card or network interface in each computing device receives computer-readable program instructions from the network and transmits the computer-readable program instructions for storage on a computer-readable storage medium within the corresponding computing device. The invention may be computer-readable program instructions to perform operations, assembler instructions, instruction set architecture instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status determination data, or source code or object. It may contain code written in any combination of one or more programming languages, including an object-oriented programming language such as Python, C++, or a functional language such as the "C" programming language or R, or similar programming languages. Computer-readable program instructions may be executed entirely on the user's computer, partly on the user's computer, as a standalone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server. In the second scenario, the remote computer can connect to the user's computer through any type of network, including a local area network or wide area network, or the connection can be made to an external computer. In some cases, electronic circuitry, including, for example, programmable logic circuitry, field programmable gate arrays (FPGA), or programmable logic arrays (PLA), can execute computer-readable program instructions using the state information of the computer-readable program instructions for customization. The high-level system architecture of the word and graph-based embedding system based on the aging factor and visit frequency in digital footprints for user behavior prediction is shown in Figure 2. The proposed system supports edge computing architecture and distributed decision making. M users (100), N user devices (101), P end servers (102) can receive service. It is assumed that there is at least one core cloud server (103) in the core network (104) and that a server (103) is directly connected to other servers in the core network (104) via wired or wireless. The core cloud server 103 is a word and graph based embedding server based on the aging factor and visit frequency in digital footprints for user behavior prediction. Here, the end server (102) is a web/mobile/metaverse application server, government institution, banking service provider, bank channel (POS device/office website, ATM, etc.), IoT gateway, authenticating institution or user device (101). ) could be himself. The system structure shown in Figure 2 is only an example. The system is also valid without having an edge computing architecture. In this case, the user device (101) will not be connected to the end server (102), but will be directly connected to the main server (103). Figure Site, core cloud server (103) architecture is shown. In the architecture, there is a communication module (200) that provides internal communication/integration within the core network server (103). The word embedding database (201) is accessed by the word embedding module (206) and word embedding analysis models are stored. The graph embedding database (202) is accessed by the graph embedding module (207) and graph embedding analysis models are stored. The aging model database (203) is accessed by the aging module (208) and the aging models are stored. Device data, user data, user behavior prediction data, training and test data are stored in the memory unit (204). The end server database (205) is accessed by the end server (102). Depending on its functionality, the data held may vary, for example, some or all of the data types stored in the core cloud server (203) may be stored there. Data sharing, data backup, etc. to the core cloud server (203). Data synchronization may occur for certain purposes. The first stage of the method of the invention is as follows: 1. The user (100) accesses the end server (102) through the user interface (AR/VR/MR device, mobile application) on the user device (101). In another flow, the end server 102 may be the user device 101 itself. The user device (101) directly accesses the core cloud server (103) and skips step 2 and moves to step 3. It transmits the content to the core cloud server (103) for processing. 3. The core cloud server (103) executes the workflow shown in Figure 1 with the data analytics module (211) on the received data: It pre-processes the data and transforms customer navigation behaviors into user movement sequences (3). It applies graph embedding and word embedding on pre-processed data (4). It creates customer placement vectors by applying normalization based on aging and visit frequency (5). It applies normalization based on frequency and time (6). It applies a leaning-based transformation to the embedding data set representing the customer's navigation behavior and creates an embedding vector for each customer (7). It divides the labeled embedding data set into training and test sets (8). The created machine learning models are used by the decision support system (209) to predict the next navigation behavior of new users (10). As a result of the behavior prediction, suggestions may be made to the system and/or user by the recommendation system (210) and/or the decision support system (209) may make a decision. 4. According to the configuration determined by the end server management module (218), some of the functionalities performed by the core cloud server (103) can be performed by the end server (102). o The decision taken by the core cloud server (103) is transmitted to the edge server (102). The end server (102) makes the relevant guidance in the user interface and/or executes the workflow in line with the suggestion provided and/or the decision taken. The said decision support system (209) is the unit that predicts user behavior in line with the data obtained as a result of graph and word embedding and decides the next action or step. The recommendation system (210) is the system that provides suggestions to the user (100) as a result of user behavior prediction and/or to decide the next action or step for the user (100) by the decision support system (209). The data analytics module (211) performs and visualizes all artificial intelligence / machine learning and data analytics operations in the system. In order for the system to predict user behavior, the user device (101), the user interface, are configured in accordance with security, legislation, policies and rules according to the user type, and the rule engine (212) ensures the execution of the necessary workflow. The verification module (213) verifies the data format in the system, compliance with the radio communication/security/data transmission protocol for data transmission, authentication, etc. It is the module that performs all types of verification operations. The query module (214) is used to query the result of user behavior prediction in the system, query the user type, query whether the prediction process has been completed, etc. Performs all query operations. The processing module (215) performs data pre-processing, post-processing, etc. in the system. It is the module that performs all processing operations. Sender device (216) is the device that sends data to external systems (API gateway, IoT gateway, etc.) by the core cloud server (103). The receiving device (217) is the device that receives the data sent by external systems to the core cloud server (103). While the end servers (102) in the system are managed by the end server management module (218), the management module (219) enables the management of the system (system configuration management, work distribution between servers, user management, device management, etc.). The system structure will also be blockchain-based. In this case, the end servers (102) are the blockchain participant node; The core cloud server (103) will be the blockchain node and the core network (104) will be the blockchain network. Blockchain network participant can be any institution (banking service provider, fintech, government agency), mobile/web/metaverse, etc. It can be an application server or a device (user device, ATM, POS, etc.). Workflow of predicting the next navigation behavior of the user (100) based on blockchain: The user (100) accesses the blockchain participant node through the user interface (AR/VR/MR device, mobile application) on the user device (101). In another flow, the blockchain participant node may be the user device 101 itself. In this case, it directly accesses the blockchain node. The blockchain participant node receives the user (100) and/or user device (101) information and transmits the content to the blockchain node for processing. In another flow, user device 101 directly accesses the blockchain node. According to the rules determined in the blockchain node management module (219) and with the configuration provided by the rule engine (212), some or all of the transactions performed can be carried out distributedly by different blockchain nodes over the blockchain network. Different data can be transmitted by the same or different blockchain participant nodes in the blockchain nodes in the blockchain network. The data obtained by blockchain nodes can be used to make collective predictions, recommendations and/or decisions during the prediction phase of the above workflow of users' next navigation behavior. Therefore, each blockchain node can have data and/or functionality independent of the others. The predictions obtained by each blockchain node are shared through the blockchain network and a joint proposal is made and/or a decision is made. The decision-making process is carried out by the decision support system (209) of the blockchain node. The decision support system (209) carries out the decision mechanism according to the work sharing between nodes determined by the management module (219). The decision mechanism can be majority voting, OR rule or AND rule or weighted average, etc. It may include method or be realized based on machine learning. Decisions made are recorded on the blockchain; Decisions or predictions made for each user are added as transactions. Blockchain node(s) execute the workflow shown in Figure 1 with the data analytics module (211) on the received data. In another flow, some of the operations performed by the blockchain node, such as data preprocessing, may be performed by the blockchain participant node. According to the configuration determined by the blockchain participant management module, some of the functionality performed by the blockchain node can be performed by the blockchain participant. The decision taken by the blockchain node is transmitted to the blockchain participant node. The blockchain participant node makes the relevant guidance in the user interface and/or executes the workflow in line with the suggestion provided and/or the decision taken. In another flow, the blockchain node directly directs the user interface and/or executes the workflow. Predicting the next navigation behavior of the user (100) with a centralized architecture. The user (100) accesses the server (102) through the user interface (AR/VR/MR device, mobile application) on the user device (101). In this scenario, there is no distinction between the edge server (102) and the core cloud server (103) (there is only one server). Additionally, there is no end server management module (218). The server (102) receives the user (100) and/or user device (101) information and executes the workflow shown in Figure 1 with the data analytics module (211) on the received data. It makes the relevant guidance and/or executes the workflow in the user interface (101) in line with the suggestion and/or decision taken by the server (102).TR TR

Claims (2)

STEMS 1. It is a method that captures, predicts, makes suggestions and makes decisions about users' navigational behavior in digital applications and pages. o the user (100) accesses the end server (102) through the interface on the user device (101), that end server (102) pre-processes the user (100) and/or user device (101) information with the data processing module (215) and sends the user information to the end server (102). (100) converting navigation behaviors into user (100) movement sequences (3), graph embedding on pre-processed data with the graph embedding module (207), word embedding application with the word embedding module (206), word embedding application with the justification module (208). creating user (100) placement vectors by applying normalization based on aging and visit frequency (5), applying frequency and time-based normalization to the placement vectors by determining a frequency and time weight for each page according to the visit frequency of each user (100) and the time spent during the visit (6). ), applying a transformation based on aging to the embedding data set representing user (100) navigation behaviors and dividing a labeled embedding data set for each user (100) into training and test sets (8), to predict user (100) intent with machine learning algorithms. To train the training data set using the user (100) embedding vectors created for the user (9), to predict the next navigation behavior of new users (100) with the machine learning models created (10), to provide relevant guidance in the user interface in line with the suggestion provided and/or the decision taken. and/or includes the process steps to carry out the workflow. 2. It is a method according to claim 1 and its feature is; The operations performed by the core cloud server (103), which manages the edge servers (102) and has some or all of the functionality of the edge servers (102), are carried out distributedly by different core cloud servers (103) over a core network (104) and the resulting It is the use of data to make collective predictions and/or decisions at the stage of predicting the users' next navigation behavior. It is a system that captures, predicts, makes suggestions and makes decisions about users' navigational behavior in digital applications and pages. users (100) access using a user device (101), takes user (100) and/or user device (101) information, divides the labeled embedding data set into training and test sets, to predict user (100) intent with machine learning algorithms. At least one end server (102), which trains the training data set using the created user embedding vectors, makes the relevant guidance in the user interface and/or executes the workflow in line with the suggestion provided and/or the decision taken, pre-processes the data and monitors the navigation behaviors of the user (100). processing module (215) that converts it into a string, word embedding module (206) that performs word embedding operations on pre-processed data, graph embedding module (207) that performs graph embedding operations on pre-processed data, user (100) navigation behaviors and final navigational behaviors. aging module (208), which applies an aging coefficient to weight, applies an aging-based transformation to the embedding data set representing user (100) navigation behaviors normalized by frequency and time, and creates an embedding vector for each user (100), graph and word The decision support system (209) that predicts user behavior in line with the data obtained as a result of embedding, makes suggestions to the user (100) as a result of the user behavior prediction and/or offers suggestions for the user (100) to decide on the next step by the decision support system (209). It includes a recommendation system (210). It is a system according to claim 3 and its feature is; It contains at least one core cloud server (103) that manages the end servers (102), receives user (100) and/or user device (101) information from the end servers (102), and has some or all of the functionality of the end servers (102). . It is a system according to claim 4 and its feature is; It contains the core network (104) that provides communication between the core cloud servers (103). It is a system according to claim 5* and its feature is; Valid for blockchain-based, the edge server (102) is the blockchain participant node, the core cloud server (103) is the blockchain node, and the core network (104) is the blockchain network. It is a system according to claim 3 and its feature is; It contains a word embedding database (201) where word embedding analysis models are stored, accessed by the word embedding module (206). It is a system according to claim 3 and its feature is; It contains a graph embedding database (202) where graph embedding analysis models are stored, accessed by the graph embedding module (207).