Method and apparatus for ranking trainers based on big data and matching recommendations using ai with preference tags

Abstract

A big data-based trainer rating calculation and propensity tag-based AI recommendation matching method and device are disclosed. A big data-based trainer grade calculation and preference tag-based AI recommendation matching method according to one embodiment of the present disclosure may include the steps of: receiving information on a trainer's qualifications and career, calculating a trainer's grade score according to a trainer grade calculation algorithm, determining a trainer's grade based on the grade score, and registering a trainer whose calculated grade score is above a preset standard in a platform database; receiving a request from a service user member to obtain member profile information and at least one preference tag; filtering trainer candidates that meet the member's preference tag and profile conditions; applying a machine learning-based AI recommendation model to the filtered trainer candidates to calculate a suitability score between the member and each trainer candidate, ranking two or more trainer candidates according to the suitability score, and providing them as a recommendation list to the member's user terminal; upon receiving a signal from the member to select one of the recommended trainers and a signal to select a desired affiliated exercise space, transmitting a matching request to the selected trainer to confirm the reservation of a personal training session between the trainer and the member; and collecting and storing feedback data about the trainer from the member after the personal training session is conducted.

Description

Technology Field

[0001] The present disclosure relates to a big data-based trainer rating and a preference tag-based AI recommendation matching method and apparatus that effectively matches a personal training service provider (trainer) and a user (member) through a trainer rating algorithm utilizing big data and an AI recommendation matching algorithm based on preference tags. Background Technology

[0002] In the health and fitness industry, the matching of members with trainers has traditionally relied on offline gyms. Traditionally, the structure was such that upon registering, members would only conduct personal training (PT) with a trainer designated by that gym. This has led to various problems.

[0003] First, there are difficulties in managing trainer qualifications and quality. Members find it difficult to directly verify a trainer's certification status or experience, and there was a risk of injury during workouts due to unverified trainers.

[0004] Second, there are issues of information asymmetry and price opacity. Personal training (PT) prices vary from gym to gym and trainer to trainer, making it difficult for members to determine appropriate price levels, which leads to information asymmetry regarding pricing. Furthermore, in some cases, PT prices are set opaquely, causing distrust among members.

[0005] Third, the inefficiency of the matching method is a problem. In traditional gyms, it is difficult for members to choose their preferred trainer; assignments are typically random or designated by the gym manager, making it difficult to adequately reflect the member's individual preferences or goals. This has led to lower member satisfaction and has also affected service quality.

[0006] Fourth, there are structural constraints between trainers and gyms. Since trainers are often affiliated with specific gyms, they face difficulties in securing fair compensation commensurate with their capabilities or a sufficient client base, and their operational space is limited to offline locations. At the same time, gyms face significant operational burdens and labor costs for hiring and training trainers as full-time employees; however, due to high turnover rates, it is difficult to achieve stable returns on investment. These structural conflicts have led to persistent inefficiencies in the personal training (PT) service market. For instance, due to the conflict between monthly subscription models and PT service models, a paradoxical situation exists where gyms profit when members use facilities less, whereas PT trainers profit when they conduct more sessions. Consequently, gym managers hire trainers to generate PT revenue, but if profits do not exceed labor costs, deficits occur, leading to a vicious cycle of declining service quality and member dissatisfaction. Furthermore, as the number of gyms closing due to financial deterioration increases, customer management issues are also emerging, where trainers lose existing member data and clients when their affiliated gym shuts down.

[0007] Against this backdrop, while trainer-member matching services have emerged through some online platforms or apps, they still face limitations such as a lack of trainer quality verification, insufficient personalized recommendations, and inadequate improvements to existing structural issues. For instance, they often remain at the level of simply introducing trainers based on distance or time slots, or fail to resolve fundamental problems by maintaining the inefficient structures of gyms (such as fixed prices and restrictions on trainer affiliation). Therefore, there is a need for a platform that enhances service quality and benefits trainers, members, and gyms alike by introducing a grading system based on trainer expertise and performance, as well as intelligent recommendation matching tailored to member preferences.

[0008] The aforementioned background technology is technical information that the inventor possessed for the derivation of the present invention or acquired during the process of deriving the present invention, and it cannot be considered as prior art disclosed to the general public prior to the filing of the present invention. Prior art literature

[0009] Prior Art 1: Korean Published Patent Application No. 10-2024-0048651 (April 16, 2024) The problem to be solved

[0010] One objective of the embodiment of the present disclosure is to objectively evaluate and grade the qualifications and abilities of trainers on a personal training matching platform, and to recommend and match trainers optimized for each member's exercise goals and tendencies using artificial intelligence.

[0011] One objective of the embodiment of the present disclosure is to verify the minimum qualifications of trainers through a trainer grading algorithm and assign grades by scoring career and competency data, thereby ensuring that only verified trainers whom members can trust are matched, and thereby preventing member damage caused by unqualified trainers and establishing a fair compensation system for the trainers' expertise.

[0012] One objective of an embodiment of the present disclosure is to provide a customized trainer recommendation that takes into account a member's exercise goals, preferred training style, and personality traits through a preference tag-based recommendation matching AI. By utilizing a model learned based on user behavior logs and group statistics, this aims to achieve personalized optimal matching rather than matching based on simple distance or price, and to increase member satisfaction and continuous usage rates (re-registration rates).

[0013] One objective of an embodiment of the present disclosure is to implement a space-sharing matching mechanism for freelance trainers, thereby enabling trainers to freely provide PT sessions at multiple affiliated gyms without being affiliated with a specific gym. Through this, gyms can generate revenue through space rental without separate employment, trainers can expand their business scope online and meet members in various regions, and members can gain the flexibility to receive PT by selecting a desired location (such as a convenient location near their residence).

[0014] In other words, the present disclosure has the objective of implementing a personal training brokerage platform that benefits trainers, members, and gyms alike.

[0015] The purpose of the embodiments of the present disclosure is not limited to the problems mentioned above, and other unmentioned purposes and advantages of the present invention may be understood from the following description and will be more clearly understood by the embodiments of the present invention. Furthermore, it will be understood that the purposes and advantages of the present invention can be realized by the means and combinations thereof set forth in the claims. means of solving the problem

[0016] A big data-based trainer grade calculation and preference tag-based AI recommendation matching method according to one embodiment of the present disclosure is a big data-based trainer grade calculation and preference tag-based AI recommendation matching method in which each step is performed by a processor, comprising: receiving qualification and career information of a trainer and calculating a grade score of said trainer according to a trainer grade calculation algorithm, determining the grade of said trainer based on said grade score, and registering said trainer, whose calculated grade score is greater than or equal to a preset standard, in a platform database; receiving a request from a service user member and obtaining member profile information and at least one preference tag; filtering trainer candidates that match said member's preference tag and profile conditions; applying a machine learning-based AI recommendation model to said filtered trainer candidates to calculate a suitability score between the member and each trainer candidate, and ranking two or more trainer candidates according to said suitability score and providing them as a recommendation list to said member's user terminal; and, upon receiving a signal from said member to select one of the recommended trainers and a signal to select a desired affiliated exercise space, transmitting a matching request to the selected trainer to confirm the reservation of a personal training session between the trainer and the member. and may include the step of collecting and storing feedback data about the trainer from the member after the personal training session.

[0017] In some examples, during the step of registering the trainer in the platform database, the trainer rating algorithm may be configured to calculate a total score by summing a basic score assigned based on whether the trainer meets the qualification requirements and a weighted score set based on the trainer's ability verification information.

[0018] In some examples, the trainer rating algorithm is configured to assign a base score based on whether the trainer meets minimum qualification requirements, and to sum predefined weighted scores for certifications held, experience, and areas of expertise. Additionally, it may be configured to score and add or subtract quality indicators based on past or external member ratings, reviews, and the number of claims collected for the trainer, or to apply a trainer-member matching success probability or satisfaction prediction value predicted by a machine learning model trained on big data within the platform as a weight, or to add or subtract a demand index considering the trainer's activity area, the number of affiliated gyms, and infrastructure suitability.

[0019] In some examples, the step of registering the trainer in the platform database may include the step of setting and storing service rate information and PT session prices corresponding to the determined grade based on the pre-set grade-specific service rates and PT session pricing policy.

[0020] In some examples, the method further includes a step of periodically or event-triggeringly updating the trainer's grade determined in the step of registering the trainer in the platform database according to a preset condition, wherein the preset condition may include at least one of the following: when the trainer's cumulative number of PT sessions exceeds a preset threshold; when the trainer's expertise is enhanced by acquiring a new certification or degree; when the rating, re-registration rate, and satisfaction indicators collected from members are above or below a preset standard; when the trainer's platform contribution indicator is above or below a preset standard; when the risk indicator is above or below a preset standard; when the frequency of recent activity or dormancy period is above or below a preset standard; or when the deviation indicator between the matching suitability predicted by the AI ​​recommendation model and the actual feedback is above or below a preset standard.

[0021] In some examples, the member’s preference tag obtained in the step of obtaining the preference tag consists of keywords expressing the exercise goals, preferred coaching style, preferred trainer characteristics, or other requirements set by the member, and the member may directly select and input one or more tags, or the member’s preferences may be automatically inferred from the member’s past behavior logs and assigned as tags.

[0022] In some examples, in the step of providing a recommendation list to the user terminal of the member, the machine learning-based AI recommendation model is implemented as a deep learning algorithm and is trained to include content-based matching that considers the similarity between the member's propensity tag and the trainer profile tag, and collaborative filtering that considers past matching success data of multiple members, and can be trained to predict a matching suitability score by considering collective statistical data including the matching success rate and satisfaction rating of multiple members together with the behavior log of an individual member.

[0023] In some examples, in the step of providing a recommendation list to the user terminal of the member, the machine learning-based AI recommendation model may be a hybrid recommendation algorithm that combines a content-based filtering module that performs cosine similarity or inner product operations between trainer profile tags and member propensity tags, and a collaborative filtering module that learns latent factors by decomposing a matching rating matrix of multiple users or predicts preferred trainers for similar member groups.

[0024] In some examples, the method may further include a step of updating the feedback data and member-trainer matching logs as training data for the machine learning-based AI recommendation model to improve the accuracy of future trainer recommendations.

[0025] In addition, a big data-based trainer grade calculation and propensity tag-based AI recommendation matching device according to one embodiment of the present disclosure comprises: a memory; and at least one processor connected to the memory and configured to execute computer-readable commands included in the memory, wherein the at least one processor receives qualification and career information of a trainer, calculates a grade score of the trainer according to a trainer grade calculation algorithm, determines the grade of the trainer based on the grade score, and registers the trainer whose calculated grade score is greater than or equal to a preset standard in a platform database. The system may be configured to receive a request from a service user, obtain member profile information and at least one preference tag, filter trainer candidates that match the member's preference tag and profile conditions, apply a machine learning-based AI recommendation model to the filtered trainer candidates to calculate a suitability score between the member and each trainer candidate, rank two or more trainer candidates based on the suitability score and provide them as a recommendation list to the member's user terminal, and upon receiving a signal from the member to select one of the recommended trainers and a signal to select a desired affiliated exercise space, transmit a matching request to the selected trainer to confirm the reservation of a personal training session between the trainer and the member, and collect and store feedback data regarding the trainer from the member after the personal training session.

[0026] In addition to this, other methods for implementing the present invention, other systems, and computer-readable recording media storing a computer program for executing said methods may be further provided.

[0027] Other aspects, features, and advantages other than those described above will become clear from the following drawings, claims, and detailed description of the invention. Effects of the invention

[0028] According to an embodiment of the present disclosure, by applying a trainer rating algorithm to filter out unqualified trainers and providing objective rating information based on experience and qualifications, it is possible to enhance member trust and promote fair compensation for trainers.

[0029] According to an embodiment of the present disclosure, by performing preference tag-based AI recommendation matching, a trainer matching the member's exercise goals and preferences can be automatically connected, thereby improving matching accuracy and member satisfaction.

[0030] According to an embodiment of the present disclosure, by operating a big data-based recommendation model that learns large-scale user behavior logs, a self-reinforcing effect can be obtained in which the quality of recommendations is automatically enhanced as the number of users increases.

[0031] According to an embodiment of the present disclosure, by introducing a space-sharing-based matching mechanism with freelance trainers, a win-win ecosystem can be established that increases the revenue and convenience of trainers, members, and gyms alike.

[0032] The effects of the present disclosure are not limited to those mentioned above, and other effects not mentioned will be clearly understood by a person skilled in the art from the description below. Brief explanation of the drawing

[0033] FIG. 1 schematically illustrates a personal training brokerage system that provides a big data-based trainer rating calculation and a propensity tag-based AI recommendation matching service according to one embodiment. FIGS. 2 and FIGS. 3 schematically illustrate a personal training brokerage platform according to one embodiment. FIG. 4 is schematically illustrated to explain an AI recommendation matching model according to one embodiment. FIG. 5 is a block diagram illustrating a personal training intermediary device according to one embodiment. FIG. 6 is a flowchart illustrating a personal training brokerage method according to one embodiment. FIG. 7 is a trainer grade classification table according to one embodiment. FIGS. 8 to 13 are exemplary diagrams illustrating interface screens of a personal training brokerage platform according to one embodiment. FIG. 14 is an application page configuration of a personal training brokerage platform according to one embodiment. Specific details for implementing the invention

[0034] The advantages and features of the present disclosure and the methods for achieving them will become clear by referring to the embodiments described in detail together with the accompanying drawings.

[0035] However, the present disclosure is not limited to the embodiments presented below, but can be implemented in various different forms and should be understood to include all modifications, equivalents, and substitutions that fall within the spirit and scope of the present disclosure. The embodiments presented below are provided to make the present disclosure complete and to fully inform those skilled in the art of the scope of the disclosure. In describing the present disclosure, detailed descriptions of related prior art are omitted if it is determined that such detailed descriptions may obscure the essence of the present disclosure.

[0036] The terms used in this application are used merely to describe specific embodiments and are not intended to limit the disclosure. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this application, terms such as “comprising” or “having” are intended to specify the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof. Terms such as “first,” “second,” etc., may be used to describe various components, but the components should not be limited by such terms. Such terms are used solely for the purpose of distinguishing one component from another.

[0037] Hereinafter, embodiments according to the present disclosure will be described in detail with reference to the attached drawings. In describing with reference to the attached drawings, identical or corresponding components are given the same reference numerals, and redundant descriptions thereof will be omitted.

[0038] FIG. 1 schematically illustrates a personal training brokerage system that provides a big data-based trainer rating calculation and a propensity tag-based AI recommendation matching service according to one embodiment, FIG. 2 and FIG. 3 schematically illustrate a personal training brokerage platform according to one embodiment, and FIG. 4 schematically illustrates an AI recommendation matching model according to one embodiment. In the present invention, the term 'platform' may comprehensively refer to a service-based infrastructure in which personal training matching, reservation, payment, and feedback functions are integrally implemented, that is, a software and hardware convergence system designed to drive and link the entire process.

[0039] Referring to FIG. 1, the personal training brokerage system (1) may include a personal training brokerage device (100), a user terminal (200), a server (300), and a network (400).

[0040] In some examples, the personal training brokerage system (1) relates to a freelance trainer brokerage platform, and more specifically, to a software system and a method of implementing the same that effectively matches a personal training service provider (trainer) and a user (member) through a trainer rating algorithm using big data and an artificial intelligence recommendation matching algorithm based on personality tags.

[0041] The trainer matching system and method according to the present invention can provide various effects by solving the aforementioned problems. First, the introduction of a trainer rating algorithm allows the expertise and experience of trainers participating in the platform to be managed using objective indicators. Trainers who do not meet minimum qualification requirements are filtered out, and a pool of verified trainers can be formed through scoring and rating based on experience and qualifications. Accordingly, members can select trainers by referring to rating information that quantifies reliability, and service quality issues caused by unqualified trainers can be significantly reduced. Furthermore, trainers receive fair compensation based on their capabilities (e.g., profit distribution based on rating), which enhances their motivation. Additionally, by providing a career ladder that allows them to advance their careers and move up to higher ratings, the trainer turnover rate can be lowered, and stable long-term participation in the platform can be encouraged.

[0042] Second, member satisfaction and convenience can be enhanced through AI recommendation matching based on personality tags. The recommendation system of the present invention collects individual members' exercise goals (e.g., weight loss, muscle gain, etc.) and preferred training methods or trainer personalities (e.g., strict coaching vs. friendly coaching, motivational style, etc.) in the form of tags, and a machine learning model combines this information with past successful matching data to identify the most suitable trainer. As a result, members can easily discover a trainer who fits them well through personalized recommendations, thereby reducing dissatisfaction caused by random assignments and trial-and-error during the consultation process. In fact, according to the present invention, effects such as improved trainer-member matching accuracy and increased member re-registration rates are expected. Furthermore, since the recommendation system learns from large-scale user behavior logs and collective statistics to become more sophisticated over time, it also achieves a big data-based self-reinforcement effect, where recommendation quality and satisfaction further improve as the number of platform users increases.

[0043] Third, through a matching mechanism based on freelance trainers and space sharing, it is possible to realize an ecosystem that benefits all platform participants (also known as the 'Moduhwa' concept). Trainers can work as freelancers without being employed by a specific gym and conduct personal training (PT) at various affiliated gyms (spaces), thereby gaining opportunities to expand their business scope and increase income. Members can choose a gym in their preferred location to receive PT, enjoying the freedom and convenience of location selection. Gyms generate additional revenue by renting out their facilities without directly hiring trainers, and operational efficiency is enhanced as they can provide PT services without separate labor or training costs. Furthermore, gyms can attract new members through the platform and maximize profits by utilizing space during idle hours. Consequently, the platform of this invention eliminates the inefficiencies arising from the existing gym-trainer employment structure and achieves a win-win structure among trainers, members, and gyms.

[0044] In summary, the present invention has the effect of enhancing service quality and reliability through big data-based trainer grading and AI matching intelligence, and driving innovation in the market structure through flexible operations based on freelancers.

[0045] Referring to FIG. 2, the personal training brokerage platform of the present invention can be bidirectionally connected to trainers, members, gyms, big data repositories, and machine learning and deep learning modules, respectively, as a cloud-type central hub. Each external entity transmits and receives various data in real time, such as profile information, reservation requests, facility usage information, and activity logs, through communication with the platform, and this data flow is illustrated by the arrows in the drawing.

[0046] In some examples, a personal training brokerage platform loads human and behavioral data collected from trainers' and members' terminals into a big data repository and can additionally integrate facility availability times, location, and equipment information provided by gym managers' terminals. The accumulated large-scale data undergoes a preprocessing stage and is transmitted to machine learning and deep learning modules, which can train and update personalized matching models by applying various algorithms such as collaborative filtering, content-based filtering, and reinforcement learning.

[0047] The trained model can generate recommendations for trainers, programs, and gyms optimized for each member's goals, preferences, and past history. The recommendations derived by the AI ​​module are returned to the platform, which provides them to members and trainers via the user interface and integrates with gym reservation systems to display them in an immediately actionable format. At this stage, member feedback and actual class attendance and usage records are reflected back into the big data repository, completing the virtuous cycle of the data model service.

[0048] As such, FIG. 2 intuitively visualizes a platform-centric data-intensive and machine learning-based recommendation ecosystem, thereby clearly explaining the functional linkages and data flow between each component.

[0049] Referring to FIG. 3, the concept of the personal training brokerage platform of the present invention is visually represented, in which consumers, trainers, and gyms are distinguished as independent entities yet integrated within a single service flow. For example, consumers can use personal training and make payments, trainers provide services and receive approximately 60% (example) of the total transaction amount, and gyms rent out space and receive approximately 30% (example). In other words, the personal training brokerage platform of the present invention can clearly define a compensation system for each stakeholder.

[0050] Furthermore, the personal training brokerage platform of the present invention applies a preference-based recommendation AI and a capability-based pricing grading system. Through this, the platform multidimensionally evaluates user preferences and trainer capabilities to grade them, and by reflecting the results in its pricing policy, enables selection based on verified information. In other words, consumers can select the trainer they desire after checking the grade and price information presented by the AI, and trainers also gain credibility as their capabilities are objectively evaluated and displayed. Additionally, the platform of the present invention does not combine or fix trainers and gyms, but rather supports consumers in selecting and combining trainers and spaces according to their personal needs. That is to say, trainers can secure class spaces by flexibly entering into contracts with multiple gyms, and gyms can attract a large number of trainers to utilize their facilities, thereby expanding the scope of value exchange centered on the platform.

[0051] As such, Figure 3 integrates and visualizes the revenue distribution structure by stakeholder, the AI-based rating and pricing mechanism, and the option for consumers to freely select and combine trainers and gyms, thereby intuitively explaining the core functions and service flow provided by the platform.

[0052] Referring to FIG. 4, the AI ​​recommendation system of the present invention may be composed of four input information and a recommendation model. For example, the first input is profile data including the member's gender, age, exercise goals, preferred time slot, etc., and the second input is profile data listing the trainer's career, qualifications, field of expertise, available hours, etc. The third input is history data consisting of past class history, feedback left by the member, star ratings, retake status, etc., and the fourth input is environmental context information such as the member's real-time location, gym accessibility, and time slot.

[0053] The central arrow visually represents the flow of each input into the AI ​​recommendation model. Inside the rectangular block on the right, an icon in the form of a multilayer neural network is depicted, symbolizing that machine learning and deep learning models combining collaborative filtering, content-based filtering, and reinforcement learning integrate and process multiple feature vectors. The recommendation model matches input member characteristics with trainer characteristics in a multidimensional space and calculates the suitability by reflecting history and contextual information as weights.

[0054] The deep learning-based AI recommendation model of the present invention may adopt a dual-path structure to simultaneously process large-scale behavior logs and static profile data. One path generates latent representations of individual features by encoding member, trainer, and context information into an embedding layer and a multilayer perceptron, while the other path can extract short-term and long-term patterns by analyzing time-series data, such as class history, feedback, and retake status, using a recurrent neural network or a transformer-series network. The latent vectors produced from the two paths are readjusted in importance in an attention module, and then linear and non-linear combinations are performed in the final integration layer. At this time, a reinforcement learning-based ranking loss function is applied to learn to predict the actual selection and retake behavior of recommendations as accurately as possible.

[0055] Model training is bifurcated into offline batch training and online fine-tuning. In the offline phase, large-scale parallel training is performed based on logs accumulated over several months to enhance overall recommendation quality, while in the online phase, real-time click, reservation, and cancellation events are collected in a small buffer and periodically fine-tuned to maintain personalization sensitivity.

[0056] The checklist icon positioned to the right of the model represents a ranked list of trainers. The AI ​​recommendation system sorts multiple trainers in descending order based on optimized suitability scores and displays the results in the output section, enabling members to make immediate selections. This generated ranked list is updated in real-time and can be automatically reordered whenever additional member feedback or changes in context are reflected. This structure implements a virtuous cycle recommendation ecosystem that achieves both accuracy and reliability by continuously absorbing interaction data between members, trainers, and gyms. A more detailed explanation will follow.

[0057] In some examples, users may access an application or website implemented on a user terminal (200) to request a personal training brokerage service, use the service at an interface provided by a personal training brokerage device (100), and perform support processes for providing the personal training brokerage service.

[0058] In some examples, the user terminal (200) may include a first user terminal (210), a second user terminal (220), and a third user terminal (230). In some examples, the first user terminal (210) may be a trainer terminal conducting personal training, the second user terminal (220) may be a member terminal requesting personal training, and the third user terminal (230) may be an administrator terminal managing a personal training brokerage system. Additionally, in some examples, although not illustrated in the drawings, a gym administrator terminal providing a gym may be additionally included. Each terminal is designed to access the platform of the present invention through the same app or web-based interface to perform various functions such as personal training matching, reservation, content management, and statistical monitoring.

[0059] However, in one embodiment, user terminals are described separately for convenience of explanation, but this is not limited thereto. That is, in one embodiment, the user terminal (200) may refer to different devices such as the first user terminal (210), the second user terminal (220), and the third user terminal (230), but may also refer to the same terminal. Since the distinction between user terminals is for logical convenience, a single device may perform multiple roles simultaneously. For example, if a trainer operates a small private studio, the same tablet may serve as both the trainer terminal and the gym manager terminal, and may remotely access the system management console to change settings when necessary. Conversely, in a large gym chain, separate dedicated terminals may be provided to distribute tasks by role.

[0060] The trainer terminal (210) is used by the trainer to register their profile, area of ​​expertise, and available time, manage session schedules, check member feedback in real time, or send and receive messages. In some embodiments, the trainer terminal may be linked with a wearable camera or smartwatch to automatically collect heart rate and exercise intensity data and transmit the data to a platform to improve the quality of personalized instruction. In another embodiment, the trainer may be configured to analyze the member's posture in real time and display corrective feedback through a smart mirror or AR glasses.

[0061] The member terminal (220) can be used to receive personalized recommendations by inputting goals, preferred time slots, exercise history, health data, etc., and to check information on recommended trainers and gyms, or to make payments, make reservations, or write reviews. In some embodiments, the member terminal (220) is implemented not only on smartphones but also on smart TVs, in-vehicle infotainment systems, smart speakers, etc., so that training video streaming and reservation notifications can be received at home or while on the move. In another embodiment, biosignals collected from a fitness band or smart ring worn by the member can be transmitted to the platform in real time and utilized as additional features of the AI ​​recommendation model.

[0062] The system administrator terminal (230) can be used for back-office functions where a platform operator performs content inspection, database maintenance, policy setting, and service quality monitoring. In some examples, the system administrator terminal (230) may be connected to a threat detection algorithm to detect abnormal patterns in real time and to immediately perform account locking and notification sending. In another embodiment, a natural language query interface may be provided to increase operational efficiency, allowing the administrator to view complex statistics using only voice and text commands, such as "Show me the status of gyms with high reservation volume today."

[0063] A gym manager terminal (not shown) can update facility equipment status, available reservation times, and fee information, and manage trainer and member entry records. In some embodiments, the gym manager terminal is linked with exercise equipment equipped with IoT sensors to automatically report equipment usage rates and malfunctions, and the platform can incorporate this into member recommendation logic to prioritize recommending gyms where specific equipment is readily available. In other embodiments, a gym unmanned kiosk doubles as the gym manager terminal, transmitting real-time headcount data to the platform simultaneously with entry authentication based on QR code and facial recognition.

[0064] Such user terminals (200) can receive services through an authentication process after accessing a personal training brokerage application or website. The authentication process may include, but is not limited to, authentication of user information such as membership registration and authentication of the user terminal, and may also be performed by simply accessing a link transmitted from a personal training brokerage device (100) and / or a server (300). All terminals undergo an authentication procedure when accessing the service, and this process may be performed by membership registration based on email and mobile phone number, social login, biometric recognition, authentication based on terminal identification value, or simply by accessing a link or QR code sent by the platform. In some embodiments, differential privacy techniques or federated learning may be applied to ensure that local terminal data is not directly stored on the server while still contributing to AI model training.

[0065] In one embodiment, the user terminal (200) may be, but is not limited to, a desktop computer, smartphone, laptop, tablet PC, smart TV, mobile phone, PDA (personal digital assistant), laptop, media player, micro server, GPS (global positioning system) device, e-book reader, digital broadcasting terminal, navigation, kiosk, MP3 player, digital camera, home appliance, and other mobile or non-mobile computing devices operated by the user. Additionally, the user terminal (200) may be a wearable terminal such as a watch, glasses, hair band, and ring equipped with communication functions and data processing functions. In some examples, the user terminal (200) is not limited to the above description, and any terminal capable of web browsing may be used without restriction. For example, a smartwatch indicates whether a reservation is confirmed with a short vibration and a notification message, AR glasses display a virtual guidance arrow while the trainer observes the member's movements, and a smart mirror projects posture correction feedback to the member in real time. By accommodating these diverse hardware environments, the platform realizes a multi-device ecosystem accessible anytime and anywhere, and can effectively expand the user experience and service scope.

[0066] In some examples, the personal training brokerage system (1) may be implemented by a personal training brokerage device (100) and / or a server (300). In other words, the personal training brokerage device (100) may be implemented in part or in whole by a server (300).

[0067] That is, the personal training brokerage device (100) may be implemented in a server (300), and the server (300) may be a server for operating a personal training brokerage system (1) that includes the personal training brokerage device (100), or a server that implements a part or the whole of the personal training brokerage device (100).

[0068] In some examples, the server (300) may be a server that controls the operation of the personal training brokerage device (100) for the overall process of receiving information from the user terminal (200) and providing a personal training brokerage service to the user terminal (200).

[0069] Additionally, the server (300) may be a server responsible for the operation of various devices so that service provision and management on the personal training brokerage platform can be performed smoothly, by managing and overseeing not only the configurations of the personal training brokerage device (100) but also the new registration of related devices and network environments.

[0070] Additionally, the server (300) may be a database server that provides data for operating the personal training brokerage device (100). Furthermore, the server (300) may include a web server, an application server, or a deep learning network providing server.

[0071] And the server (300) may include a big data server and an AI server required to apply various artificial intelligence algorithms, and a computation server that performs computations of various algorithms.

[0072] Additionally, in some examples, the server (300) may include the servers described above or network with such servers. That is, in the present embodiment, the server (300) may include the web server and AI server mentioned above or network with such servers.

[0073] In some examples, the personal training brokerage device (100) and the server (300) in the personal training brokerage system (1) may be connected by a network (400). Such a network (400) may include wired networks such as LANs (local area networks), WANs (wide area networks), MANs (metropolitan area networks), and ISDNs (integrated service digital networks), or wireless networks such as wireless LANs, CDMA, Bluetooth, and satellite communication, but the scope of the present disclosure is not limited thereto. Additionally, the network (400) may transmit and receive information using short-range communication and / or long-range communication.

[0074] Additionally, the network (400) may include connections of network elements such as hubs, bridges, routers, switches, and gateways. The network (400) may include one or more connected networks, such as a multi-network environment, including a public network such as the Internet and a private network such as a secure corporate private network. Access to the network (400) may be provided through one or more wired or wireless access networks. Furthermore, the network (400) may support an Internet of Things (IoT) network and / or 5G communication that exchanges and processes information between distributed components, such as objects.

[0075] FIG. 5 is a block diagram illustrating a personal training intermediary device according to one embodiment.

[0076] Referring to FIG. 5, the personal training intermediary device (100) may include a communication interface (110), a user interface (120), a memory (130), and a processor (140).

[0077] The communication interface (110) may provide a communication interface necessary to provide transmission and reception signals between external devices in the form of packet data in conjunction with the network (400). Additionally, the communication interface (110) may be a device including hardware and software necessary to transmit and receive signals, such as control signals or data signals, through wired or wireless connections with other network devices. This communication interface (110) may support various types of intelligent object communication (IoT (Internet of Things), IoE (Internet of Everything), IoST (Internet of Small Things), etc.) and may support M2M (machine to machine) communication, V2X (vehicle to everything communication), D2D (device to device) communication, etc. That is, the processor (140) may receive various data or information from an external device connected through the communication interface (110) and may also transmit various data or information to the external device. Furthermore, the communication interface (110) may include at least one of a WiFi module, a Bluetooth module, a wireless communication module, and an NFC module.

[0078] In some examples, the communication interface (110) performs the role of converting and transmitting all control data signals exchanged with user terminals, gym equipment, external cloud services, etc., in conjunction with the network (400) into packet form. The module optionally includes various wireless protocols such as 5G, LTE, Wi-Fi, Bluetooth LE, Zigbee, NFC, ANT+, and LoRa as well as wired Ethernet, thereby supporting real-time data exchange between heterogeneous devices such as mobile terminals, wearable sensors, smart exercise equipment, and IoT gateways.

[0079] In some examples, the processor (140) receives information such as member heart rate and motion data, trainer availability time, gym IoT equipment status, and payment results through the communication interface (110), and conversely, can transmit personalized recommendation results, reservation confirmation notifications, equipment maintenance instructions, etc. to an external device. For example, application layer protocols such as RESTful API, WebSocket, MQTT, and gRPC are also built-in, allowing for software-based support without additional modules for inter-server integration or third-party service expansion.

[0080] Additionally, the communication interface (110) can be designed to offload and process the authentication, encryption, and session management of the terminal to a hardware security module (HSM) or a security processor, and can be configured to support distributed processing with edge computing nodes or CDN-based streaming transmission as needed. Thanks to this scalability, the processor (140) can maintain stable service even under high-load conditions where multiple user terminals and gym sensors are connected simultaneously.

[0081] The user interface (120) is designed to dynamically load role-specific screens and components within a single application framework, thereby providing layouts and functions optimized for each user scenario of trainers, members, gym administrators, and system administrators. For example, when a member logs in, the personalized goal achievement rate calculated in real-time by AI and recommended session fees are displayed in card form on the first screen, and if access to location information is allowed, a list of gyms easily accessible from the current location can be automatically sorted at the top. Additionally, for example, when a trainer logs in, their daily schedule and empty time slots are visualized in a timeline form at the top of the dashboard, and new inquiry history, session reviews, and average rating change trends are displayed as chart widgets at the bottom, enabling immediate response.

[0082] In some examples, the user interface (120) supports multimodal input methods, so operation via touch, keyboard, mouse, as well as voice commands and camera gestures may be possible. For example, if a trainer uses a smartwatch during a session to say “2 more sets remaining,” it is immediately reflected on the real-time recording screen, and if a member assumes a squat position in front of a mirror-type smart display, a camera-based posture tracking module may be activated to display knee angle and balance indicators as an overlay at the top of the screen. These multimodal input data are centralized on an internal event bus and can be synchronized so that each device shares the same session context.

[0083] In some examples, the output interface can automatically switch its information hierarchy and color theme based on the situational context. A high-contrast mode that enhances readability may be applied in bright lighting environments, while a dark mode considering battery efficiency may be applied in night mode or on small smartwatch screens. For system administrator terminals, if the server load exceeds a specific threshold, the top of the dashboard switches to an alert color, and simultaneously, push and email notifications are sent in combination to prompt an immediate response. On gym administrator terminals, if an "abnormal vibration" log is received from an exercise equipment IoT sensor, the corresponding equipment card blinks and the obsolescence graph automatically expands, displaying the timing for replacing necessary parts along with the estimated cost.

[0084] In another embodiment, the user interface (120) may be equipped with progressive web app (PWA) technology to maintain core functions even in an offline environment. For example, a member may be able to check reservations, watch offline videos, and upload progressive data even when the network is unstable in a subway or underground gym. Data is temporarily stored locally and automatically synchronized with the server when the network is restored, and if a conflict occurs, it can be resolved according to a priority policy based on roles. This caching mechanism can ensure that session notes are safely reflected on the server without data loss, even if a trainer writes them on the spot.

[0085] In addition, in some examples, considering multilingual and multicultural environments, the user interface (120) may support both an internationalization module and a localization module. The default language is set to Korean, but it can be immediately switched to English, Japanese, Spanish, etc. by detecting the system language of the terminal, and the unit system can also be automatically converted to matrix or imperial units. In terms of accessibility, alternative text for screen reading may be added for the visually impaired, and haptic feedback and screen flash options for major notifications may be added for the hearing impaired.

[0086] In some examples, the user interface (120) can be implemented as a widget-based modular structure. When a new health data measuring device or a third-party nutrition coaching service is integrated, the service provider can distribute a custom widget through the SDK, and the user can place it on the home screen by dragging and dropping. Similarly, when a gym operator wants to add their own promotional banner or event reservation function, the frontend can be dynamically updated with just server settings. This eliminates the need for separate app market distribution for updates, thereby ensuring both management efficiency and user convenience.

[0087] As such, the user interface (120) can expand the scope of the service of the present invention by providing a consistent and personalized experience to members, trainers, gyms, and system administrators through an expandable design that encompasses multiple roles, multiple devices, and multiple input channels, ensuring core functions even in offline situations, and covering accessibility and internationalization requirements.

[0088] In some examples, the memory (130) may include a volatile or non-volatile recording medium capable of storing various information necessary for the control (operation) of the operation of the personal training intermediary device (100) and / or server (300) and storing control software.

[0089] The memory (130) is connected to one or more processors (140) via an electrical or internal communication interface and can store codes that cause the processor (140) to control the personal training intermediary device (100) when executed by the processor (140).

[0090] Here, the memory (130) may be a non-transient storage medium such as a magnetic storage medium or a flash storage medium, or a transient storage medium such as RAM, but the scope of the present invention is not limited thereto. Such memory (130) may include internal memory and / or external memory, and may include volatile memory such as DRAM, SRAM, or SDRAM, non-volatile memory such as OTPROM (one time programmable ROM), PROM, EPROM, EEPROM, mask ROM, flash ROM, NAND flash memory, or NOR flash memory, flash drives such as SSD, CF (compact flash) card, SD card, Micro-SD card, Mini-SD card, Xd card, or memory stick, or storage devices such as HDD.

[0091] Additionally, information related to an algorithm for performing learning according to the present disclosure may be stored in the memory (130). Furthermore, various information necessary within the scope of achieving the purpose of the present disclosure may be stored in the memory (130), and the information stored in the memory (130) may be updated as it is received from a server or external device or input by a user.

[0092] The processor (140) can control the overall operation of the personal training intermediary device (100). Specifically, the processor (140) is connected to the configuration of the personal training intermediary device (100) including a memory (130) and can control the overall operation of the personal training intermediary device (100) by executing at least one command stored in the memory (130).

[0093] Such a processor (140) can be implemented in various ways. For example, the processor (140) can be implemented as at least one of an Application Specific Integrated Circuit (ASIC), an embedded processor, a microprocessor, hardware control logic, a hardware finite state machine (FSM), or a digital signal processor (DSP).

[0094] The processor (140) is a type of central processing unit that can control the operation of the personal training intermediary device (100) by running control software loaded in the memory (130). The processor (140) may include all types of devices capable of processing data. Here, 'processor' may mean a data processing device embedded in hardware that has a physically structured circuit to perform a function expressed by code or instructions included in a program, for example.

[0095] FIG. 6 is a flowchart for explaining a personal training brokerage method according to one embodiment, and FIG. 7 is a trainer grade classification table according to one embodiment. Hereinafter, with reference to FIG. 6 and FIG. 7, the personal training brokerage method of the processor (140) will be described in detail.

[0096] Referring to FIG. 6, in step S10, the processor (140) receives information on the qualifications and experience of the trainer from the first user terminal (210), calculates the trainer's grade score according to the trainer grade calculation algorithm, determines the trainer's grade based on the grade score, and registers the trainer whose calculated grade score is greater than or equal to a preset standard in the platform database.

[0097] First, the processor (140) can first verify the original profile data, such as the trainer's certification, education, career, field of expertise, and past member ratings, received from the first user terminal (210). In the verification step, the processor (140) can verify the authenticity of the submitted certification by comparing it with a public database, an issuing agency API, or a blockchain-based authentication token, and can cross-verify the career details with affiliated gym records or session logs. The data, once authenticity is verified, can be converted into a fixed-length feature vector through preprocessing steps such as missing value handling, categorical encoding, and normalization.

[0098]

[0099] Subsequently, the processor (140) first determines whether the minimum qualification requirements are met according to the trainer grade calculation algorithm, and if the conditions are met, 2 points can be awarded as a base score. Then, points can be accumulated by referring to the weighting table for each category of the held certifications. For example, referring to Table 1, private certifications registered in the PQI are awarded differentially from 1 to 2 points, but only up to two can be recognized, and national certified Level 2 certifications are awarded 3 points per certification, but only up to two can be recognized. National certified Level 1 certifications are awarded 6 points, international intermediate certifications are awarded 8 points, and international advanced certifications and health exercise manager certifications are each awarded 12 points, but only up to two or one of each item can be recognized. 2 points can be added if proof of military service, enrollment / leave of absence from a physical education college, or a graduation certificate is submitted, and 1 point can be added if verifiable field experience is within 2 years. At this time, the maximum number of recognized items for each category is accumulated only within the limits listed in the table to prevent excessive score bias.

[0100] In some examples, the trainer rating algorithm may be configured to calculate a total score by summing a base score based on whether the trainer meets qualification requirements and a weighted score set based on the trainer's ability verification information.

[0101] The processor (140) can also reflect factors other than qualifications and experience. For example, by analyzing internal platform logs, if a trainer's activity level falls within the top 10 percent of the total, up to 3 points can be added as an activity weight, and if a member's rating is below a certain standard or claims are repeated, points can be deducted. Additionally, the trainer-member matching success probability or session satisfaction prediction value calculated by a machine learning model can be converted into a weight and added to or subtracted from the score. Supply imbalances can also be mitigated by dynamically adjusting the weights by considering external variables such as regional demand index, the number of affiliated gyms, and the suitability of equipment infrastructure.

[0102] In some examples, the rating algorithm may be configured to assign a base score based on whether the trainer meets minimum qualification requirements, and to sum predefined weighted scores for certifications held, experience, and areas of expertise. Additionally, the rating algorithm may be configured to score and add or subtract quality indicators based on past or externally collected member ratings, reviews, and claim counts for the trainer, or to apply weights to trainer-member matching success probability or satisfaction predictions predicted by a machine learning model trained on big data within the platform, or to add or subtract demand indices considering the trainer's available areas, number of affiliated gyms, and infrastructure suitability. The processor (140) may also apply additional weights for specific fields if the training field registered by the trainer corresponds to a rare area of ​​expertise designated by the platform, such as high-risk specialized exercises or rehabilitation exercises. The processor (140) analyzes platform operation logs to calculate activity indicators such as the number of sessions, member retention rate, and content upload frequency based on the last 90 days, and, for example, can add up to 3 points if the activity level falls within the top 10 percent or subtract up to 2 points if it falls within the bottom 10 percent. At the same time, the processor (140) calculates quality indicators by averaging external / internal ratings, and can subtract 2 points for a rating of 4.8 or higher and 2 points for a rating of less than 4.0.

[0103] In some examples, the processor (140) receives a trainer-member matching success probability or session satisfaction prediction value predicted by a machine learning model as a real number in the range of 0.0…1.0, converts it to a linear scale, and can add or subtract up to 3 points. The prediction model may be retrained periodically, including collaborative filtering embeddings and graph confidence indicators.

[0104] The processor (140) can look up the regional gym demand index and apply bonus points if there is a shortage of trainers in a specific region, and deduct points if it is saturated. If the number of affiliated gyms is above a certain standard, up to 2 points can be added as an infrastructure suitability score.

[0105] The processor (140) can calculate a final grade score by summing up all the basic score, bonus points, and deduction points accumulated above, and record the score in the trainer profile record.

[0106] In some examples, the processor (140) may first integrate trainer profile data and platform operation data into a single feature vector to execute a rating algorithm. During the integration process, text and categorical values ​​may be converted into one-hot encoding or word embeddings, and continuous values ​​may be scaled by min-max normalization or standardization to balance the influence between different units. This preprocessed feature vector may be loaded in the form of a matrix in which the row direction consists of individual trainers and the column direction consists of feature items.

[0107] In some examples, the processor (140) may generate a primary score vector by multiplying a predefined weight matrix by this characteristic matrix. The weight matrix may be generated by combining an initial weight set by a group of experts and an experience-based coefficient extracted from platform operation statistics, and the weight value of each column may represent the relative importance that the corresponding characteristic contributes to the grade score. For example, the weight of an international advanced certification item may be set to a higher real value than that of a domestic private certification item so that it has a greater influence even with the same unit score.

[0108] In some examples, the processor (140) can calculate a secondary score by adding a statistically based correction value to the primary score vector. The statistically based correction value can be calculated using percentile ranks, Gini coefficients, square root transformations, etc., to mitigate distribution distortion and can mitigate the phenomenon of scores being excessively concentrated in specific items. At the same time, activity bias or quality deviation can be reflected in real time by selecting the top 10 percent activity range or the lower rating range from the entire platform trainer distribution and setting addition and subtraction values.

[0109] In some examples, the processor (140) may input a secondary score vector into a machine learning correction module to estimate a final correction value. The machine learning correction module may be implemented as a gradient boosting model or a neural network ensemble model trained to predict target variables such as past matching success rates, session re-enrollment rates, and member satisfaction. The model may output a matching success probability using features such as individual trainers' secondary scores, regional demand indices, gym infrastructure suitability, and seasonal factors, and may linearly scale this probability to convert it into an additive and subtractive value within a range of ±3 points.

[0110] In some examples, the processor (140) may calculate a final grade score by summing the static weight score, the statistical-based correction value, and the machine learning-based correction value. In the calculation process, weight parameters such as λ1, λ2, and λ3 may be used to adjust the contribution rate of each correction step, and the optimal λ value may be periodically updated through cross-validation or Bayesian optimization. The calculated grade score is recorded in an encrypted storage before being transmitted to the grade determination module and can be used later as reference data for audit log analysis or model retraining.

[0111] Referring to FIG. 7, in some examples, the processor (140) can derive the trainer's final grade by comparing the score obtained from the grade calculation algorithm with a pre-set threshold interval. For example, referring to Table 2, if the score exceeds 3 points, it can be mapped to a pink dumbbell; if it exceeds 5 points, to a yellow dumbbell; if it exceeds 10 points, to a blue plate; and if it exceeds 18 points, to a white belt. If the score is 18 points or higher and satisfies fast pass conditions, such as possessing two or more international advanced difficulty certifications or having an elite athlete career, it can be directly mapped to the ET (Elite Trainer) grade. These mapping intervals can be periodically refreshed according to changes in platform operation indicators or amendments to regulations.

[0112]

[0113] In some examples, the processor (140) can look up the basic session unit price and service rate corresponding to the determined grade in a lookup table and automatically reflect them in the trainer profile. For example, a standard unit price of 30,000 won for pink dumbbells, 40,000 won for yellow dumbbells, 60,000 won for blue plates, 80,000 won for white belts, and 100,000 won or more for ETs can be applied, and the final payment amount can be calculated by multiplying this by additional factors such as regional premium, peak time surcharge, long-term membership discount, and promotional coupon. For example, when booking a weekend prime time session at a gym located in the metropolitan prime zone, the final rate can be calculated by sequentially multiplying the basic unit price by a regional factor of 1.2 times and a peak time factor of 1.1 times.

[0114] In some examples, the processor (140) may convert the tier-specific rates into the base currency and then convert them into the local currency by referring to a real-time exchange rate API to support a multilingual and multi-currency environment, and display the value with the applied currency commission rate on the member screen. Additionally, when a corporate affiliate member purchases a large number of sessions, a quotation may be automatically generated by applying a tier-specific discount separately.

[0115] In some examples, the determined grade and rate information can be stored as metadata in the trainer profile database, and the same information can be synchronized with the cache node and search index to increase the search response speed. The processor (140) can send a notification of promotion and new registration approval to the trainer terminal (210) at the time of grade registration, and can reflect the session unit price change history in real time to enable immediate booking.

[0116] In some examples, the processor (140) may increase exposure priority by applying a promotional rate to new trainers during the first 30 days after grade registration. In another embodiment, to provide additional incentives to trainers of a certain grade or higher during a specific campaign period, the processor (140) may be configured to automatically calculate a discount or additional allowance by temporarily inserting campaign parameters into the rate calculation formula.

[0117] In some examples, the processor (140) may store trainers whose rating score is less than the minimum registration standard of 3 points or whose required document verification has not been completed in a provisional registration state in the platform database and then send a notification requesting supplementation to withhold the granting of a formal rating. Conversely, if a high-scoring trainer additionally meets the rating requirements, the processor may automatically perform a promotion process and apply a new rate immediately upon promotion to display updated prices to the member.

[0118] In this way, the processor (140) determines the trainer's grade based on grade scores and threshold intervals, calculates the final session price by applying rate policies defined for each grade and various adjustment factors, stores the information in the platform database, and can reflect it in real time on the trainer terminal and member interface.

[0119] In some examples, the processor (140) can update the trainer's grade determined at the step of registering the trainer in the platform database periodically or in an event-triggered manner according to preset conditions.

[0120] At this time, the pre-set conditions may include at least one of the following: when the number of cumulative PT sessions of the trainer exceeds a predetermined threshold; when the trainer's expertise is enhanced by acquiring a new certification or degree; when the rating, re-registration rate, and satisfaction indicators collected from members are above or below a predetermined standard; when the trainer's platform contribution indicator is above or below a predetermined standard; when the risk indicator is above or below a predetermined standard; when the frequency of recent activity or the period of inactivity is above or below a predetermined standard; or when the deviation indicator between the matching suitability predicted by the AI ​​recommendation model and the actual feedback is above or below a predetermined standard.

[0121] In some examples, if Table 3 is referenced, the processor (140) can automatically promote and demote the trainer rank by periodically or event-triggering predefined update conditions to manage the trainer rank as a dynamic value rather than a fixed value. The processor (140) can first monitor the trainer's cumulative PT session count and set it to promote to a higher tier within the same series if the value exceeds 600, and to the next rank if it exceeds 900. When the promotion condition is met, the processor (140) checks the current rank lock period and determines whether to unlock it, and if unlocked, immediately recalculates the rank score and maps it to the new rank.

[0122]

[0123] Referring to Tables 4 and 5, when the processor (140) receives from the first user terminal (210) the fact that a trainer has improved their expertise by acquiring a new certification or degree, it can verify the evidence at source, update the certification weighting table, recalculate the grade score, and process the promotion immediately. Similarly, if Fast Pass requirements are met, such as acquiring an additional international intermediate or advanced certification, obtaining a health exercise manager certification, or proving elite athlete experience, the trainer can be updated to a higher grade immediately without waiting for a separate session.

[0124]

[0125]

[0126] In some examples, the processor (140) may analyze member ratings, re-registration rates, and session satisfaction indicators collected in real-time from the platform to promote promotion by granting bonus points when the rating is maintained at 4.8 or higher, or to lower the rating by applying deduction points when the rating is maintained at 4.0 or lowered for a long period or claims accumulate. Additionally, the rating may be re-evaluated at each point in time by applying additional bonus points if the trainer's platform contribution indicator is above a certain standard, and deduction points if the risk indicator (e.g., missed reservations, rule violations) is above a standard.

[0127] In some examples, the processor (140) can be configured to measure the deviation between the matching suitability predicted by the AI ​​recommendation model and the actual member feedback, and to deduct the rating score if the deviation indicator is above a preset threshold, and to grant a bonus if the deviation is low and excellent matching continues. Additionally, if the recent activity frequency is below a threshold or the dormancy period is prolonged, the rating may be switched to a temporary hold state and platform exposure may be restricted.

[0128] In some examples, when a rank changes as a result of performing update logic, the processor (140) may immediately reflect the new rate, symbol, and exposure priority in the trainer profile and send promotion and demotion notifications to the trainer terminal (210). The change history is recorded in the audit log and can be used as evidence in the event of a future dispute. The processor (140) may repeat this rank update procedure periodically or whenever an event occurs to keep the trainer rank system within the platform always up to date.

[0129] In some examples, the processor (140) can refine the rating maintenance policy by combining various update cycles and trigger methods. For example, at dawn on Monday, the cumulative number of sessions and activity metrics of all trainers can be aggregated in batch mode to identify candidates for promotion, and at midnight every day, event logs can be analyzed in real-time streams to immediately downgrade trainers with plummeting ratings and accumulated claims. In batch mode, the weighted moving average of the last year's data can be calculated to mitigate short-term volatility, and in stream mode, sensitivity to the last 24 hours' metrics can be increased to defend platform quality in real-time.

[0130] In some examples, the processor (140) can dynamically adjust thresholds to reflect seasonal demand and major events. For instance, in January and July when fitness demand surges, the promotion session threshold can be relaxed to 90%, and in March and September when demand decreases, the downgrade threshold can be tightened by 10% to prevent oversupply. During the major sports competition season, the weight of rehabilitation and conditioning specialist trainers can be temporarily increased to ensure that experts in the field are quickly displayed at the top.

[0131] In some examples, the processor (140) may build a risk profile for each trainer and comprehensively evaluate quantitative and qualitative indicators. Quantitative indicators may include the reservation no-show rate, refund rate, and number of terms and conditions violations, while qualitative indicators may include the sum of negative keyword frequencies extracted from sentiment analysis in session reviews. If the risk indicator exceeds a range threshold, a deduction factor may be applied to the rating score or the status may be switched to a temporary suspension, and the processor (140) may send a notification requesting correction to the trainer.

[0132] In some examples, the processor (140) may include the use of newly released AI assistance tools in the contribution indicator. If a trainer actively utilizes AI form checks, nutrition coaching bots, and personalized exercise prescription features to increase member satisfaction, the contribution score may increase, and promotion may be accelerated. Conversely, if an act of bypassing the platform's official workflow by sharing contacts externally and deviating from transactions is detected, a penalty period may be imposed along with a reduction in contribution.

[0133] In some examples, the processor (140) may provide a cushion period to minimize the impact of the update results on the member experience. For 72 hours after the promotion and demotion are confirmed, the algorithmically changed rates are displayed only in the back office, and trainers may submit objections. If no objections are raised, the rates may be automatically reflected on the public profile and payment screen at the end of the cushion period. During this process, audit logs and snapshots before and after the change are stored for a long time and can be used as evidence in the event of a dispute.

[0134] In some examples, the processor (140) may apply custom renewal rules at the request of a partner company that owns a multi-branch gym chain. In one embodiment, a trainer belonging to the chain may add the completion of separate internal training as a prerequisite for promotion to ensure systematic service quality, and the processor (140) may automatically verify eligibility by linking with a training completion API. In another embodiment, in a program dedicated to franchise gyms, the weighting for international yoga and Pilates certifications may be increased, for example, by approximately 1.5 times to increase the promotion speed of the relevant professional trainer.

[0135] In some examples, the processor (140) can automatically detect long-term inactive trainers and switch their rating to a pending state. For example, a trainer who has had no sessions for the past 90 days and has no login logs within the last 30 days is labeled inactive, and their rating can be restored only after they complete mandatory online refresher training upon applying for reactivation. At this time, for 14 days immediately after restoration, the sensitivity of rating and claim metrics can be increased to intensively monitor initial quality.

[0136] In some examples, the processor (140) can balance supply by reflecting the number of gyms where local trainers are active in the area in the score when local demand surges. For example, if the number of users near a new residential complex increases rapidly, a local correction score of 2 points is given to trainers affiliated with adjacent gyms to promote promotion, and conversely, in areas with an oversupply, a reduction of 1 point is applied to adjust exposure frequency.

[0137] In this way, the processor (140) can continuously maintain the trainer rating system within the platform in an up-to-date, fair, and demand-friendly state by combining predefined update conditions and real-time evaluation indicators to readjust the ratings periodically and on an event basis.

[0138] In step S20, the processor (140) receives a request from a service user member from the second user terminal (220) and obtains member profile information and at least one preference tag.

[0139] In some examples, a member's preference tags consist of keywords expressing exercise goals set by the member, preferred coaching styles, preferred trainer characteristics, or other requirements; the member may directly select and input one or more tags, or the member's preferences may be automatically inferred from the member's past behavior logs and assigned as tags.

[0140] When the processor (140) receives a service usage request from the second user terminal (220), it can retrieve profile information that has been entered during the sign-up stage or that has been updated by the member. For example, member profile information may include gender, age, body measurements, medical history, years of exercise experience, occupational characteristics, available weekly leisure time, preferred exercise location, available equipment, health goals to be achieved, payment history, social connection information, etc. These profile items can be entered directly by the member upon initial sign-up or modified at any time in the settings menu thereafter, and the modification history is version-controlled and can be immediately reflected as input values ​​for the recommendation model.

[0141] The processor (140) can then obtain at least one member’s preference tag and use it as a personalized matching parameter. The preference tag may consist of keywords such as body fat reduction, muscle strength increase, rehabilitation exercise, lower body focus, hybrid training, preference for group sessions, preference for non-verbal coaching, preference for female trainers, preference for low-noise environments, preference for outdoor running, plant-based diet, and request for wearable integration.

[0142] The processor (140) can collect preference tags via a direct input method. The processor (140) can expose a UI to allow the user to directly select or freely input one or more keywords from a list of tags, and can record the selected or entered values ​​in a database in real time. For example, if a member selects a desired keyword through a dropdown, toggle switch, or hashtag search bar during the app onboarding stage, a tag array can be generated in real time.

[0143] In addition, in one embodiment, a natural language sentence entered by a member in a free-form descriptive format can be analyzed by a natural language processing module to extract intent, sentiment, and keywords, and then semantically mapped with a predefined set of tags to automatically assign propensity tags.

[0144] The processor (140) can collect preference tags directly from the member. The processor (140) can update the tag array by receiving UI elements directly manipulated by the member in real time on the event bus.

[0145] The processor (140) can also obtain propensity tags through indirect inference.

[0146] The processor (140) can collect unstructured logs such as a member's past search terms, reservation patterns, exercise videos watched, click and scroll behavior logs, chatbot conversation history, and wearable sensor data, and apply natural language processing and association rule mining techniques to automatically infer preferences and assign them as propensity tags.

[0147] The processor (140) utilizes collaborative filtering embeddings to present tags commonly possessed by similar member groups as recommended tag candidates, and can confirm the corresponding tags when a member presses a confirmation button. For example, behavior logs such as the types of sessions a member has previously booked and paid for, video content watched, coaching tips clicked, wishlists, wishlist cancellation logs, and sensor-based heart rate fluctuation graphs can be analyzed for time-series patterns to predict preferences by comparing them with similar member clusters. In some examples, stress levels are periodically received from the member's smartwatch to classify users who frequently experience high-stress periods as candidates for stress relief tags.

[0148] In addition to direct input, the processor (140) can automatically infer tendency tags by analyzing the member's past behavior logs. If a pattern is observed where the member repeatedly schedules the 'Lower Body Focus Program' or only views running-related videos, the 'Part_Lower Body' or 'Exercise_Running' tag can be automatically assigned. Additionally, by analyzing heart rate and stress levels collected from the smartwatch, the 'Goal_Stress Relief' tag can be assigned to members who are frequently stressed, and the 'Time_Night' tag can be assigned to members who frequently schedule night sessions.

[0149] The processor (140) can feed chatbot conversations or descriptive inputs into a natural language processing model to extract intent and sentiment, and can supplement propensity tags by semantically mapping the extracted keywords to a predefined set of tags. The processor (140) can supplement propensity tags through a conversational chatbot. For example, the chatbot can ask a member a question such as, "When receiving coaching, do you prefer detailed explanations or demonstration-based coaching?" and classify the response in real time to update tags. The chatbot can call a photo analysis API to extract the calorie and nutrient ratios of a meal photo uploaded by a member and automatically assign diet-related tags.

[0150] In addition, in some examples, context variables such as location, weather, and gym density can be checked to automatically add the 'environment_hometraining' tag to members living in outlying areas, for example, to increase recommendation accuracy. The processor (140) can adjust the preference tag based on context.

[0151] In one embodiment, the home training equipment tag weight can be increased when a member's location moves to the outskirts of a gym-dense area. Additionally, in one embodiment, if a member repeatedly schedules a night session, a night-only tag can be added, and parameters such as lighting brightness and noise tolerance can be adjusted.

[0152] In some examples, the processor (140) may combine the collected and generated propensity tags with member profile embeddings and store them as input features for the recommendation AI engine, and maintain the up-to-date nature of the personalized matching by calling an incremental learning module to retrain the model weights whenever the tags change. The processor (140) may execute consistency verification logic at each tag input time to detect conflicting goals and constraints, and if a conflict is detected, it may adjust the recommendation ranking or display a confirmation popup to the user to guarantee tag quality.

[0153] The processor (140) can send a real-time feed message to the AI ​​recommendation model whenever a tag update event occurs, so that it can be immediately reflected in the next session matching.

[0154] In some examples, the processor (140) may define multilayer fields including physical indicators, age, gender, occupation, underlying diseases, past exercise experience, preferred exercise time zone, residential area, type of terminal used, payment history, etc. to construct member profile information and collect the corresponding values ​​from the second user terminal (220).

[0155] In some examples, the processor (140) can preprocess the collected profile fields according to standard specifications, store them in a JSON schema or relational table, and then convert them into future model input vectors.

[0156] In some examples, the processor (140) can subdivide the tendency tags into exercise goals, preferred coaching style, trainer preference characteristics, session environment requirements, desired auxiliary services, allergy diet restrictions, rest patterns, motivation type, social tendency, and keywords for sensitivity to the latest trends.

[0157] In some examples, the processor (140) can convert public interests obtained from a social login API, types of posture imbalances extracted from body shape image analysis results, real-time location-based weather preferences, etc., into additional features and extend them into dynamic preference tags.

[0158] In some examples, the processor (140) maps a tag dictionary into a BERT embedding space to allow multilingual input, and when a new language expression is detected, it can be associated with an existing tag based on similarity to manage it without duplication.

[0159] In step S30, the processor (140) filters trainer candidates that match the member's preference tags and profile conditions.

[0160] The processor (140) can first query a pool of trainers in the trainer profile database that meet basic constraints, such as physical distance accessible to the member, gym affiliation, and available time slots for booking. The retrieved records can be filtered in the first stage to include candidates without physical and temporal conflicts by comparing the member's real-time location information with the trainer's movement radius and whether online sessions are provided. The processor (140) can estimate the member's expected travel time by linking with a real-time traffic API and keep only trainers whose travel time is within the maximum allowed travel time set by the member as candidates. In the same process, the processor (140) can further exclude candidates with low accessibility by calculating weights for detailed mobility convenience indicators, such as the number of public transportation transfers, parking availability, and the presence of accessibility facilities for the disabled. When a request for an online session is detected, the processor (140) can ignore the physical distance filter and keep only trainers whose video communication equipment specifications and internet bandwidth test results meet minimum standards.

[0161] In some examples, the processor (140) may subsequently compare the tendency tags possessed by the member with the specialty tags possessed by the trainer using a mapping table to retain only candidates with an intersection. In this case, tag matching may be recognized as a valid match only when the similarity is above a variable threshold, after calculating similarity by referring to semantic embedding similarity, tag hierarchy, and thesaurus, in addition to simple string comparison. In the tag matching stage, the processor (140) calculates similarity using a BERT-based Korean and English multilingual embedding model, and when the member selects the "low-carb diet" tag, it recognizes "low-carb," "keto," and "ketogenic diet" as the same cluster and retains candidates above the matching threshold. The processor (140) is configured to allow matching of higher concepts by reflecting the tag hierarchy, so that when the member selects strength, trainers specializing in "powerlifting" and "Olympic lifts" may also be retained in the filtering stage.

[0162] In some examples, the processor (140) may additionally check the gender preference, instruction style preference, language preference, health constraints, etc., of the member profile to exclude candidates that do not match the corresponding attributes of the trainer profile. For example, for a member with a female instruction preference set, male trainers may be automatically excluded, or for a member with a history of knee injury, trainers who specialize in high-intensity jump programs may be excluded. In addition to gender and language preference, the processor (140) may consider religious and cultural requirements, and exclude trainers who do not support options, for example, if conditions include a request for a Halal-friendly training space or a restriction on specific music genres. For members with a history of herniated discs in their health constraints, the processor (140) may check whether they possess a spinal stability certification and filter out trainers who do not. Under the same conditions, the processor (140) may exclude candidates who do not possess a senior physical education instructor certification for senior members.

[0163] In some examples, the processor (140) can filter trainers whose recent rating is below a threshold, whose no-show rate is excessively high, or whose risk score is above a threshold by considering platform quality indicators. At the same time, the probability of booking failure can be reduced by checking the number of remaining booking slots per trainer and temporarily excluding trainers that are saturated. The processor (140) can temporarily exclude trainers whose negative sentiment rate in recent reviews has increased by utilizing sentiment analysis scores as well as the rating re-registration rate in quality indicator filtering. The processor (140) can raise the risk score in real time if a no-show rate spikes or a violation of terms is detected, and can remove a trainer from the candidate list if the threshold is exceeded. At the same time, the processor (140) can temporarily exclude a trainer from the candidate list if the number of available booking slots per trainer is less than 3, judging that there is a high probability of booking failure.

[0164] In some examples, the processor (140) may cross-reference the session type support information of the trainer profile with the gym equipment information to remove candidates that do not meet the conditions when a member's request is for a group session or includes whether specific equipment is available. If the request is for a pet-friendly session, the trainer may be filtered immediately if they do not support that option. When a member requests a group session, the processor (140) may check the trainer's maximum concurrent instruction setting value and keep only candidates that are at least one greater than or equal to the requested number. When the use of specific equipment, such as a climbing wall or a Pylates reformer, is required, the processor (140) may cross-reference with the affiliate gym equipment list to remove trainers who do not have such equipment. When a request for a dog-friendly session is made, the processor (140) may keep only trainers with the dog-friendly flag turned on.

[0165] In some examples, the processor (140) stores the result set of multi-condition filtering completed in a memory cache and, before passing it to the next stage of the fitness calculation module, may re-search by relaxing the tag similarity threshold or expanding the distance limit if the number of candidates is too small. The final set of candidates thus obtained can be used to generate a ranking and recommendation list in a subsequent stage.

[0166] In some examples, the processor (140) may gradually expand the search range by lowering the tag similarity threshold, for example, by 0.05, in the first re-search step and expanding the maximum travel distance limit, for example, by approximately 20%, in the second re-search step, if the filtered set of candidates is less than the minimum target quantity. If there are still insufficient candidates after re-search, the processor (140) may activate the time zone auto-suggestion function to present the member with an alternative, nearby time zone, and if the member accepts, re-filter based on that time zone. The processor (140) may temporarily store the final secured set of candidates in a memory cache and check for missing values ​​or duplicate records through a data consistency check before passing it to the suitability calculation module.

[0167] In step S40, the processor (140) applies a machine learning-based AI recommendation model to the filtered trainer candidates to calculate a suitability score between the member and each trainer candidate, and ranks two or more trainer candidates according to the suitability score and provides them as a recommendation list to the member's second user terminal (220).

[0168] In some examples, machine learning-based AI recommendation models are implemented with deep learning algorithms and are trained to include content-based matching that considers the similarity between a member's preference tag and a trainer profile tag, and collaborative filtering that considers past matching success data of multiple members, and can be trained to predict matching suitability scores by considering collective statistical data including the matching success rates and satisfaction ratings of multiple members together with the behavior logs of individual members.

[0169] In some examples, the machine learning-based AI recommendation model may be a hybrid recommendation algorithm that combines a content-based filtering module that performs cosine similarity or inner product operations between trainer profile tags and member propensity tags, and a collaborative filtering module that learns latent factors by decomposing a matching rating matrix of multiple users or predicts preferred trainers for similar member groups.

[0170] In some examples, machine learning-based AI recommendation models can be designed to calculate a fit score by fusing features learned through two main paths. The first path is content-based matching, which takes member preference tags and trainer profile tags embedded in pre-trained language models such as BERT and Word2Vec as input and extracts semantic proximity between tags by performing cosine similarity, scale jump attention, and cross dot product operations. In this process, in addition to tag embeddings, numerical features such as distance and time convenience, session cost, equipment compatibility, and available trainer slots can be projected into the same embedding space to learn multidimensional relationships together.

[0171] The second path is based on collaborative filtering learning. It takes hundreds of thousands of member-trainer matching rating matrices, session re-registration status, and satisfaction indicators as input to estimate latent factors using Factorization Machines or Graph Neural Networks (GNNs). Additionally, by defining nodes in the user item interaction graph as member trainers and weighting edges based on session success, preference probabilities can be learned using GraphSAGE or LightGCN. This learned collaborative path can complement information on collective preference patterns that are difficult for content paths to capture—specifically, "trainers actually preferred by similar members."

[0172] In the fusion layer that integrates the prediction logits of two paths, various strategies such as weighted averages, stacking ensembles, and band-it-based dynamic weight adjustment can be applied. For example, in a cold start situation, the weight of the content path can be increased, and as a stable history accumulates, the contribution rate of the collaboration path can be gradually increased to guide the model to adapt in a balanced manner across the data space.

[0173] The learning loss can be configured to minimize the absolute satisfaction prediction error while directly optimizing the matching rank using an objective function that combines rank-based lists and point-wise regression losses. Collective statistical metrics (matching success rate, average rating, etc.) are reflected as sample weights for each mini-batch to maintain overall platform performance, and individual behavior logs (click sequence, session duration, review sentiment score, etc.) can be processed by an attention-based time-series encoder to capture short-term preference changes.

[0174] In the inference phase, latency can be minimized by adopting a two-stage structure that first performs a primary cutoff by rapidly scoring candidate criterion fit using linear and tree-based lightweight models, and then applies a deep learning hybrid model only to the top candidates. Additionally, by incorporating meta-learning techniques, initial parameters for new trainers and members can be rapidly adapted using a small sample size, and gender, age, and regional diversity penalties are inserted as coefficients into the loss function to mitigate bias toward specific groups and ensure fairness.

[0175] In some examples, the model output may include tag matching results, collaboration preference metrics, and recent quality metrics as explainability metadata, along with goodness-of-fit probability values ​​for each candidate trainer. This metadata is used by clients to visualize "why this trainer was recommended" in the form of natural language or icons, which can increase member trust and reduce the cost of re-searching.

[0176] In some examples, the processor (140) can convert the set of trainer candidates obtained in step S30 and the member's propensity tags and profile vectors into input tensors for a machine learning-based AI recommendation model. The input tensors may include multidimensional characteristics for each member-trainer pair, such as tag embeddings, distance and time convenience indicators, real-time rates, recent ratings, session frequency, and platform contribution. The processor (140) can place these tensors in GPU memory or TPU memory and then perform a forward pass of the model to output a fitness score for each candidate trainer in the form of a vector.

[0177] In some examples, the processor (140) can combine multiple representations in a stacked form by generating one-hot and multi-hot tag embeddings, as well as Word2Vec and BERT-based contextual embeddings and graph embeddings, in parallel during the input tensor construction phase. These multiple embeddings can be passed through a convolution layer and a transformer encoder layer once to extract interaction characteristics. At this time, the distance and time convenience indicators can be scaled by applying inverse weights, the session unit price by log scaling, and the recent rating by exponentially weighted moving average.

[0178] That is, the processor (140) can pair the trainer candidates selected in step S30 with member information one by one to create row-unit data. This can include various numerical characteristics such as tag values ​​like the exercise goals and teaching style selected by the member, the actual distance and estimated time between the member and the trainer, the session price applied when booking now, the trainer's recent average rating, the number of sessions conducted in the last month, and the number of content or reviews contributed to the platform. After stacking the multiple rows created in this way to form one large input tensor and loading it onto GPU or TPU memory to perform forward delivery to the AI ​​recommendation model, a fit score indicating 'how well this trainer fits this member' can be obtained in the form of a vector for each row of the tensor.

[0179] In some examples, tag information may be represented as a one-hot vector containing only 0s and 1s, but it may also be converted into Word2Vec or BERT embeddings so that tags with similar meanings, such as 'strength training' and 'powerlifting', are mapped closely to each other, or graph embeddings may be generated by learning the relationships between tags as a graph structure. The processor (140) may stack the various types of embeddings created in this way layer by layer and combine them into one large input, and then pass them through a convolution layer and a transformer encoder layer to automatically learn the complex interactions between tags and between tags and numeric features.

[0180] In some examples, because the range of units and values ​​for each feature is significantly different, making it difficult for the model to learn, the processor (140) may be normalized by taking the reciprocal of the distance and time convenience indicators so that the value increases as it gets closer, changing the session unit price to a logarithmic scale to smooth out the difference in values, and applying an exponentially weighted moving average to the recent rating so that the value is reflected more heavily for the newer reviews. Since the input tensor, after being normalized in this way, has various features reflected in a balanced manner, the model can calculate the fit score for each trainer candidate stably and consistently.

[0181] In some examples, the processor (140) can implement the AI ​​recommendation model as a hybrid structure combining a content filtering module based on a multilayer perceptron and a collaborative filtering module based on matrix decomposition and a graph neural network. The content filtering module can calculate the association between tags through cosine similarity between member tag embeddings and trainer tag embeddings, inner product operations, attention weighted sums, etc. The collaborative filtering module can decompose hundreds of thousands of past member-trainer matching ratings into a low-dimensional latent space or predict the preferred trainer probability of similar member clusters by learning graph edge weights. The processor (140) can integrate the output probabilities of the two modules into a final fit score by weighted averaging them.

[0182] In some examples, the processor (140) can capture high-dimensional semantic relationships by fusing tag embeddings of 128 dimensions or more in the content filtering path of the hybrid model with cosine similarity and scale jump attention techniques. In the collaborative filtering path, 64 dimensions of latent factors based on matrix decomposition can be extracted, and in the graph neural network, 32 dimensions of connection strengths based on node2beck can be learned to output continuous values. The final fit probability can be calculated by multiplying the logits of the two paths by weighting factors λ1 and λ2 and then applying softmax regularization.

[0183] That is, the processor (140) can configure the AI ​​recommendation model like a tree with two roots. One root is content-based filtering, which creates numerical coordinates of 128 dimensions or more from the tag vector chosen by the member and the trainer profile tag vector, and then calculates how similar the two vectors are using cosine similarity and attention-weighted sum. In this way, a trainer who teaches "powerlifting" well to a member who wants "strength training" can naturally receive a high score.

[0184] Another approach is collaborative filtering, which can extract 64-dimensional latent factors by 'matrix decomposing' hundreds of thousands of past matching data, or learn 32-dimensional connection strengths using a Node-2Vec-based graph neural network by viewing members and trainers as points and lines on a graph. Through this, information on "trainers actually liked by similar members" can be predicted as probability values.

[0185] By adjusting the importance of the scores (logits) produced by the content path and collaboration path, respectively, by multiplying them by weighting factors λ1 and λ2, and then adding the two values ​​and applying softmax normalization, a final fit probability between 0 and 1 is obtained. This process balances the two perspectives of tag semantic similarity and actual usage patterns, enabling the identification of the trainer best suited for the member.

[0186] In some examples, the processor (140) can optimize the loss function by simultaneously using collective statistical indicators and individual behavior logs during the model training phase. Collective indicators may include matching success rates, session re-enrollment rates, and average satisfaction ratings, while individual logs may include click sequences, session retention times, and review keyword embeddings. The processor (140) can adopt a dual-path learning strategy to capture both long-term preference trends and short-term interest fluctuations. Additionally, it can reliably estimate initial fitness scores even in cold start situations where data is scarce, such as with new trainers or new members, by utilizing meta-learning techniques. The processor (140) can optimize performance by combining focal loss and ranking-based supervised loss during model training. Collective statistical indicators can be group-normalized by year, quarter, or month, and individual behavior logs can be input into a masked LSTM based on session length to reflect short-term time series characteristics. The meta-learning part can utilize a MAML framework to quickly adapt initial values ​​for cold start member trainers.

[0187] That is, when training the model, the processor (140) utilizes both the overall flow of the platform and detailed behavioral records left by individual members. The collective statistics showing the overall flow include values ​​such as the rate at which a match was actually made, the rate at which it was re-booked, and the average satisfaction, while the individual records include the order in which a button was pressed, how long a session was watched, and keywords included in reviews. By using these two types of data with different characteristics simultaneously, the model can identify not only long-term trends common to all members but also short-term interests that have just emerged.

[0188] In some examples, the approach of learning large and small flows separately and then combining them into one is called a dual-path learning strategy. The long-term path learns overall changes in preferences by steadily accumulating statistics. On the other hand, the short-term path uses masked recurrent neural networks to handle behavioral sequence data of varying lengths, sensitively capturing changes that have occurred over the past few days or weeks. Combining the information obtained from these two paths improves recommendation accuracy.

[0189] In some cases, there is little data available, such as with new members or new trainers, which is called a cold start. The processor (140) mitigates this situation by introducing meta-learning techniques. Simply put, this involves having the model practice solving small problems quickly multiple times in advance, so that it can adapt quickly with minimal training even when encountering unfamiliar data. A typical method is to use the Mammel framework, which prepares the initial values ​​themselves to be easy to adapt to.

[0190] In some examples, focal loss and ranking loss are used together as criteria to reduce errors during model training. Focal loss helps focus on rare but important cases when dealing with data where successes are infrequent, while ranking loss ensures that the order of recommendation results is correctly aligned. Collective statistics are grouped by time intervals, such as years or months, and standardized through averaging or standardization to ensure uniform size, while individual logs retain only the necessary parts based on session length and are fed into the time-series model to capture the latest changes. By combining these various mechanisms, the model can respond evenly to long-term and short-term changes as well as situations with limited data, thereby maintaining stable recommendation quality.

[0191] In some examples, the processor (140) may sort the fitness scores obtained after real-time inference in descending order and select the top N trainers as a recommendation list. The recommendation list may include fields such as trainer name, profile picture, specialty, session price, expected satisfaction, and expected arrival time. The processor (140) may add an explainability indicator to each item to visually highlight the basis for tag matching and the basis for collaborative filtering. The recommendation list may be serialized in the form of a JSON response and transmitted to the member's second user terminal (220).

[0192] In some cases, the processor (140) may present descriptive possibilities such as “This trainer has three or more successful rehabilitation exercises and is rated as having a similarity of 0.87 with member tag ‘knee pain’”, including a Korean natural language summary generated from the XAI module along with a recommendation list. For example, the list may be serialized into a GraphQL API and provided to both mobile apps and web clients with the same schema.

[0193] That is, the processor (140) sorts the suitability scores calculated by the AI ​​model in order of highest to lowest, and then selects a few trainers that are judged to be the most suitable to create a recommendation list. This list includes information such as each trainer's name and photo, main field, cost per session, predicted satisfaction value, and estimated arrival time when the member travels.

[0194] In some examples, to explain to the member why these results occurred, the processor (140) displays the reasoning in small print for each trainer item, such as 'matches knee rehabilitation tag' or 'a large number of similar members re-registered'. This allows the recommendation to be delivered in a form where the reasoning is visible, rather than as a simple black box result, thereby increasing credibility.

[0195] In some examples, the entire recommendation list is bundled in JSON format and sent to the member's smartphone or web browser. Since the platform supports both mobile apps and websites, the processor (140) serializes the data into the same GraphQL schema format and provides it identically to both environments.

[0196] In addition, in some examples, the Explainable Artificial Intelligence (XAI) module automatically generates a one-line Korean summary and displays detailed evidence such as, "This trainer has successfully performed rehabilitation exercises three or more times and has a similarity score of 0.87 with the 'knee pain' tag specified by the member." Thanks to this summary, members can easily see why each trainer is suitable for them.

[0197] In some examples, the processor (140) may adopt a two-stage ranker structure in which, to keep real-time inference latency at, for example, approximately 50ms or less, the latency-sensitive layer is separated into a latency-processing layer in on-device mode, and a sub-model lightweighted with a linear blending model is placed to calculate a primary score, and then a secondary deep model is performed on the server side. When selecting the top N trainers, the sorting criteria may be multiplied by a fairness correction factor to ensure diversity by gender, age, and region is greater than a set value.

[0198] That is, since the processor (140) must display recommendations very quickly, it processes the calculation process in two stages. When a member opens the screen, a lightweight mini-model runs on the mobile phone first to instantly assign approximate scores to candidate trainers. This mini-model has a simple calculation formula and is small in size, so it produces results in the blink of an eye. Only the top candidates selected in this first stage are sent to the server, where a more complex deep learning model spends sufficient time to recalculate precise scores. Because the process involves a fast first stage on the mobile phone and an accurate second stage on the server, the overall latency is short while the quality of the recommendation is maintained.

[0199] In some cases, when determining the recommendation order, a fairness factor is multiplied to slightly adjust the results to prevent an excessive concentration of trainers of a specific gender, age, or region. Thanks to this, the gender ratio and regional distribution are maintained evenly above the minimum diversity standards set by the platform, allowing members to receive balanced recommendations for trainers from various backgrounds.

[0200] In some examples, the processor (140) can store recommendation results in a cache server to minimize response delays for the same request over a short period. After the recommendation list is displayed in the member interface, if a member selects a specific trainer or displays scroll and click patterns, the processor (140) can collect the corresponding behavior logs into a feedback channel and reuse them as model input features. The processor (140) can update the model parameters by running offline batch training or online fine-tuning at regular intervals to reflect the latest feedback. For example, the processor (140) can store TOP-K results in a cache server with an LRU policy, asynchronously push behavior logs to a Redis stream, and then deliver them to an ETL cluster via a Kafka pipeline to merge real-time logs and batch logs. These logs can be collected into a Spark ML pipeline in a nightly batch and aggregated into a model retraining dataset.

[0201] That is, the processor (140) can store the recommendation result just calculated in a memory-based cache server so that if the same member requests it again within a short time, it can respond immediately without waiting. When the cache space is full, the oldest unused data can be emptied first to maintain the latest results.

[0202] In some examples, when a member touches a trainer card or scrolls up and down on the recommendation screen, the processor (140) can collect the click non-exposure information through a real-time feedback channel and reuse it for model training. The collected behavior logs can be transmitted asynchronously so as not to affect application performance.

[0203] In some examples, the processor (140) may temporarily store the collected logs in a memory queue and then send them to a message stream platform, and load them into an analysis-only repository through a data transformation pipeline. In this process, real-time logs and periodically accumulated logs can be combined to form a single consistent data set.

[0204] In some examples, the processor (140) can use a bulk processing engine to add new logs to the model training data when traffic is low, such as during late-night hours each day, and update the model parameters to the latest state through offline batching. When a sudden change in trend is detected, small-scale fine-tuning can be performed in real-time to prevent the recommendation quality from degrading.

[0205] In some examples, the processor (140) may optionally apply federated learning or differential privacy noise injection for privacy protection. Additionally, it may incorporate an A / B test framework to experimentally verify differences in click-through rates and booking rates between recommendation algorithm versions. Through these procedures, the processor (140) can reliably provide real-time personalized recommendations while continuously improving the quality of member-trainer matching. For privacy protection, the processor (140) may apply Differential Privacy SGD during federated learning to perform model updates within the noise ε = 1.0 range. Furthermore, for example, in the A / B test management console, recommendation algorithm versions v1, v2, and v3 can be distributed with traffic weights of 30-30-40, and click-through rates, booking rates, refund rates, and session completion rates can be monitored via a real-time dashboard. Through these procedures, the processor (140) can provide a highly reliable real-time recommendation service while simultaneously ensuring personalization and system fairness.

[0206] That is, the processor (140) can apply 'feeded learning' to improve the model without sending member data directly to the server. This method is structured such that the model is trained in small units on each user's smartphone or gym terminal, and then only the training results (changes in weights) are encrypted and collected and combined on the server. Thanks to this, the original data of individual members does not leave the terminal, thus protecting privacy. Since there may still be a risk of re-identification, the processor (140) also uses a 'differential privacy' technique that slightly mixes random noise into the updated weights. For example, if the ε value that controls the noise size is set to about 1.0, the possibility of personal identification can be effectively reduced without significantly compromising statistical accuracy.

[0207] In some examples, experiments are required to determine if the recommendation algorithm is actually better when refreshed. The processor (140) has an A / B testing framework built in to run multiple versions of the algorithm simultaneously and compare their performance in real time. For example, if there are three versions v1, v2, and v3, traffic is divided in a 30:30:40 ratio and applied to different members. Then, metrics such as click-through rate, booking completion rate, refund rate, and session completion rate are checked immediately on the dashboard to determine which version is the best. Through this process, low-performing versions can be quickly removed, and high-performing versions can be immediately expanded.

[0208] Ultimately, the processor (140) can provide more precise and fair real-time recommendations reliably by protecting personal information while constantly improving the model and providing the verified optimal version to the member.

[0209] In step S50, when the processor (140) receives a signal to select one of the recommended trainers and a signal to select a desired affiliated exercise space through the second user terminal (220), it transmits a matching request to the first user terminal (210) of the selected trainer to confirm the reservation of a personal training session between the trainer and the member.

[0210] In some examples, when a member selects both a trainer and an affiliated exercise space, the processor (140) can first check again in the trainer's real-time schedule whether the requested time slot is still available. If it is determined that there is no schedule conflict, the processor can temporarily secure a reservation slot by applying a transaction lock to prevent duplicate bookings from occurring in the same time slot. Immediately after securing the slot, a matching request notification can be sent to the trainer's terminal to wait for approval.

[0211] In some examples, the processor (140) can mark the room number and list of essential equipment to be used during the corresponding time period as locked by linking with the gym manager terminal simultaneously with the trainer approval. The gym manager can check the newly created lock request on the manager dashboard and automatically issue an equipment inspection checklist to inspect for any abnormalities in advance. If the inspection is passed, the 'Facility Ready' button is pressed on the manager terminal, and when that signal returns to the platform, the processor (140) can confirm the room lock and display a 'Ready' status badge on the schedule cards of both the trainer and the member.

[0212] In some examples, when the trainer returns approval, the processor (140) can simultaneously confirm the room or equipment for the corresponding time slot by linking with the facility reservation system of the exercise space selected by the member. Once the space reservation is complete, a reservation record with the same session ID can be created in three places—the trainer schedule, the facility calendar, and the member calendar—to maintain data consistency. Then, a payment window is provided to the member to inform them of the amount reflecting the session unit price, VAT, and platform fees, and upon receiving a payment completion signal, the reservation can be converted to a confirmed state.

[0213] The processor (140) can separate and settle the basic session unit price and platform fee during the payment stage, and calculate the gym usage fee as a separate item to provide real-time profit forecasting to the gym manager. Once payment is completed, the processor (140) can record the transaction history in a distributed ledger system to ensure transparency in refunds and settlements. At the same time, the projected profit and room usage schedule are updated on the gym manager's terminal, allowing daily and weekly rental rates to be viewed as graphs.

[0214] In some examples, the processor (140) can send push notifications and emails to both the trainer and the member at the time of reservation confirmation to provide information such as the date, time, location, required items, and cancellation deadline. At the same time, convenience can be enhanced by displaying a session card widget containing a calendar invitation link and a location sharing button on both user terminals. Reservation details are recorded simultaneously in a distributed cache and a persistent database so that they can be recovered without loss in the event of a failure.

[0215] In some examples, the processor (140) may register with the scheduler to automatically send reminder notifications one day and one hour before the session starts, even after the reservation is confirmed. If the trainer does not approve or the member does not complete the payment, the locked slot may be automatically unlocked after a certain period of time to allow another member to make the reservation. If a reservation cancellation occurs, automatic settlement may be processed according to the refund policy, and the changes may be reflected in real time for both users.

[0216] In the session preview notification, attire, arrival route, and parking instructions can be transmitted separately to the member terminal (220), session goals and member health precautions to the trainer terminal, and room setting time and disinfection cycle to the gym manager terminal. On the day of the session, QR codes or facial recognition information can be automatically issued by linking with the gym smart kiosk for entry authentication, and when entry is confirmed, the processor (140) can save the attendance record to the session log.

[0217] In some examples, when a session ends, the processor (140) may send an equipment closing inspection checklist to the gym manager, display a popup for writing a brief review on the trainer terminal (210), and display a satisfaction survey link on the member terminal (220). If the gym manager closes it as 'normal', it may create a next usage reservation and an automatic cleaning schedule, and if an equipment malfunction is reported, it may issue a maintenance ticket to notify the maintenance team.

[0218] In some examples, when a reservation is cancelled, the payment amount may be partially refunded or a penalty may be imposed by combining the time of cancellation and the gym's individual refund policy. If a penalty is incurred, the processor (140) can automatically recalculate and distribute the trainer allowance, platform fee, and gym usage fee ratios, and can make the empty slot available again to the gym manager to switch it to a re-bookable state.

[0219] In some examples, the processor (140) records transaction logs and push notification status in dual storage for data consistency so that they can be retransmitted even in the event of a communication failure. If a schedule change or cancellation occurs after midnight, an automatic rescheduling batch job can be run at dawn the next day to update all calendar invitations, facility calendars, trainer calendars, and member calendars in bulk. By doing so, trainers, members, and gym managers can all smoothly conduct personal training sessions while accurately sharing schedules in real time.

[0220] In step S60, the processor (140) collects and stores feedback data about the trainer from the member after the personal training session.

[0221] In some examples, feedback data and member-trainer matching logs can be incorporated as training data for machine learning-based AI recommendation models to improve the accuracy of future trainer recommendations.

[0222] In some examples, the processor (140) may display an automatic popup on the member terminal (220) immediately after the session ends to prompt the member to enter a rating and a descriptive comment. The rating is divided into detailed categories such as five stars, effort, expertise, and kindness, and autocomplete keywords are provided in the comment input window so that the member can quickly leave an opinion. If the member selects voice feedback, the voice-to-text conversion module within the terminal can be configured to immediately convert it into text and send it.

[0223] In some examples, the processor (140) can display a star rating interface with a question, "How was this trainer session?" immediately at the bottom of the member's terminal when the session ends. In addition to the five-star rating, detailed score sliders such as effort, expertise, and friendliness are arranged side by side so that the member can adjust them with a single finger. In the comment input field, keywords such as "good at motivating" and "excellent posture correction" are pre-presented in chip form so that they can be pressed quickly, and additional comments can be entered as text. If a "voice comment" button is provided to account for situations where the mobile phone cannot be held, the voice-to-text conversion engine within the terminal can recognize both Korean and English and generate a text string with minimal delay to send to the server.

[0224] In some examples, the processor (140) can bundle the transmitted rating, descriptive sentence, keyword selection history, voice source file hash, session duration, and re-registration intent check into a single feedback record and record it in a server log storage via an encrypted channel. At the time of storage, the member's real name and sensitive data can be removed, leaving only the hashed session ID, trainer ID, and time information, thereby minimizing the exposure of personal information.

[0225] In some examples, the processor (140) can bundle the collected information—star rating value, detail score, free input sentence, selected keyword ID, SHA-256 hash of voice file, actual session duration, and "willingness to rebook / not" toggle result—into a single JSON record and load it into a log store via a TLS encrypted channel. At this time, the member ID can be hashed and the terminal identifier can be tokenized so as not to directly store the original personal information.

[0226] In some examples, the stored feedback records can be converted into structured features, such as positive / negative scores, key keyword embeddings, and sentence lengths, through natural language processing and sentiment analysis in a preprocessing pipeline. The processor (140) can combine raw star ratings and session logs with these converted features and periodically aggregate them into an offline training dataset. Sentiment scores can be adjusted, for example, by weighting the latest 30 days to emphasize short-term satisfaction changes, and star ratings can be adjusted by applying an exponential moving average to reflect long-term quality trends.

[0227] In some examples, the server-side preprocessing pipeline inputs free sentences into a KoBERT-based sentiment analysis model to extract positive and negative scores, and converts noun and verb keywords into Word2Vec vectors. Small but strong signaling features such as review length, frequency of emoticon usage, and number of exclamation marks can also be added. The processor (140) combines the newly extracted structured features, raw scores, and session meta-information to create a 'feedback feature table', which can then be merged into an offline training dataset on a daily basis.

[0228] In some examples, the processor (140) can perform batch learning at regular intervals to incorporate new feedback into the model parameters. In this case, the recent intention to re-register or whether to actually re-book can be added as a target variable to the ranking loss function to learn whether satisfaction led to actual behavior. When an urgent change in quality is detected, weights can be immediately updated through online fine-tuning, and trainers whose downward ratings have surged can be set to have their recommendation scores temporarily lowered.

[0229] In some examples, batch learning gives more weight to sentiment scores from the past month to emphasize recent changes in satisfaction. For star ratings, an exponential moving average is applied to retain only long-term quality trends, allowing the influence of older ratings to naturally diminish. The ranking loss function can be designed to include a target variable indicating whether the same trainer was rebooked within a week of the session ending, allowing the model to directly learn whether high ratings led to actual behavior.

[0230] In some examples, if the average rating of a particular trainer drops sharply within a short period or the frequency of negative sentiment keywords exceeds a warning threshold, the processor (140) can call an online fine-tuning mode to immediately lower the recommendation weight of the trainer. At the same time, it can send an automatic notification to the quality control team to request an investigation into the cause and corrective measures.

[0231] In some examples, after the model is newly trained, the processor (140) can gradually distribute it to an A / B test group to verify whether the click-through rate and booking rate improve, and then scale it to the entire traffic if performance improvement is confirmed. During this process, all feedback records are version-tagged to track which model learned which data, and rollbacks can be performed quickly if necessary. The processor (140) can repeat this procedure to continuously improve member-trainer matching accuracy and maintain the up-to-dateness and reliability of the recommendation system.

[0232] In some examples, the model updated with new parameters is first distributed to an A / B experiment group that applies to only some of the users. The processor (140) monitors click-through rates, booking confirmation rates, session completion rates, and refund rates on a real-time dashboard and can expand the application to the entire traffic only when the new model is statistically significantly superior to the existing model. The feedback records used in each experiment are stored with model version tags so that later, which data trained which model can be tracked and a specific version can be rolled back immediately if necessary.

[0233] Through this cyclic process, the processor (140) can absorb all fresh feedback generated in each session as training data, respond in real-time to rapid quality changes, and continuously strengthen reliable personalized recommendations.

[0234] FIGS. 8 to 13 are exemplary diagrams illustrating interface screens of a personal training brokerage platform according to one embodiment, and FIG. 14 is an application page configuration table of a personal training brokerage platform according to one embodiment.

[0235] FIG. 8 schematically illustrates the configuration of a mobile application main screen according to one embodiment. Referring to FIG. 8, it is designed so that key information to check and major navigation paths can be viewed at a glance upon the user's first entry.

[0236] The top area is a header where the app logo and status bar are located, and icons for the time, battery, and communication status can be placed there as well. In the banner area directly below, a dynamic marketing message is displayed in large text, and a date indicating the statistical reference date can be inserted in the bottom right corner. On the central 'Recruitment' card, the total revenue amount is aggregated in real-time and displayed as a large number for visual emphasis.

[0237] The function shortcut area in the middle consists of two icon buttons. The left button, labeled 'Find a Trainer' along with a person icon, leads to the personal training recommendation screen, while the right button, featuring a map icon and the 'Find a Gym' label, connects to the affiliate facility search screen. Each button adjusts responsively based on device resolution, and the label font can be enlarged if accessibility mode is enabled.

[0238] The My Menu area below lists frequently used self-service items in the form of single-line cards. The first card contains the title "How do I get PT?" and a short explanatory text, and links to a beginner user guide screen. The second card (150) is the "Today's News" item, which connects to a news feed where platform announcements and fitness trend content are provided in a summary list format.

[0239] A tab bar navigation is located at the very bottom of the screen, allowing you to quickly switch between key sections of the app via 'Home', 'Messages', 'Wishlist', and 'Profile' icons. The icon border and label color of each tab change depending on the selection state, intuitively guiding you to your current location.

[0240] FIG. 9 illustrates an exemplary UI configuration of a trainer search screen. The overall internal layout of the smartphone is divided into a top header area, a list view area, and a bottom tab bar navigation (600), and key interaction elements for the user to search for and select trainers are arranged in stages.

[0241] The 'Find Trainer' page title is displayed on the left side of the header area, and a heart icon leading to the favorites page is placed on the right to enhance immediate accessibility. A search bar is located below the title to support keyword-based search, followed by a horizontally arranged sorting tab bar. The sorting tab consists of five options: 'All', 'Recommended', 'Popular', 'Reviews', and 'Map', and the list can be updated in real-time based on the corresponding criteria upon tapping.

[0242] Trainer cards can be arranged continuously in a vertical scrolling format within the list view area. Each card includes a space for a profile picture, the trainer's name, a summary of their specialty, their activity area, and desired session rates. A heart favorite button is located at the bottom right, allowing users to instantly save individual trainers. Text within the card automatically wraps to fit the width, and a default icon may be displayed as a substitute if the profile picture is not loaded. In the second list section below the area separator, additional cards are seamlessly connected via an infinite loading method when scrolling, ensuring that even large amounts of data are displayed without lag.

[0243] In some examples, the bottom tab bar navigation consists of four icon buttons: 'Home', 'Messages', 'Wishlist', and 'Profile'. The 'Wishlist' tab, corresponding to the current page, is highlighted with an accent color and underline to intuitively guide the user to their current location, and selecting another tab allows navigation to that section with a natural fade transition effect.

[0244] As such, this drawing visually demonstrates that user convenience and data exploration efficiency can be simultaneously secured by providing a UI example in which search, sort, list, save, and navigation functions are harmoniously arranged on a single screen.

[0245] Figure 10 shows an example UI configuration of a trainer profile detail screen. Inside the smartphone exterior, a top status bar, a header and profile area, an action button area, an information tab area, a self-introduction area, and bottom tab bar navigation can be arranged in succession.

[0246] On the left side of the header and profile area (120), a trainer profile picture and a trainer name (130) may be displayed in the center, and a heart icon for favorites may be placed at the top right so that the trainer can be saved immediately as a favorite trainer.

[0247] In the action button area (150), a ‘Pay’ button and a ‘Consult’ button are placed side by side so that a member can immediately pay for a trainer session or connect to a 1:1 message consultation. The two buttons can dynamically change color or activation status depending on the real-time availability of the trainer.

[0248] At the top of the information tab area (170), 'Basic Information' and 'Reviews' tabs are displayed, and when the Basic Information tab is selected, fields such as gender, MBTI, major qualifications, and career may appear in a list format. When switching to the Reviews tab, a member evaluation card and an average rating graph may be loaded.

[0249] In the self-introduction area (180), an introductory message entered directly by the trainer is displayed as a speech bubble card, allowing personal information such as personality, goals, and training philosophy to be emphasized. Below the introduction text, a video link or a certificate PDF badge can be additionally inserted to increase credibility.

[0250] The tab bar navigation at the bottom of the screen consists of 'Home', 'Messages', 'Wishlist', and 'Profile' icons, and when you enter the current profile screen, the 'Wishlist' tab label is displayed in an accent color to intuitively guide you to your current location.

[0251] Figure 11 illustrates an exemplary screen configuration for a member to input a preference tag. A display area is placed inside a frame representing the exterior of a smartphone to visually provide a member interface. A preference input card with rounded corners is placed in the center of the display area and may include survey items and slider controls.

[0252] At the top of the card, the title 'Select Preference' is displayed, allowing users to recognize that the current stage is for inputting preferences. The first question is 'Which trainer do you prefer?', with 'Flexible' labels placed on the far left and 'High Intensity' on the far right, and intensity preference can be continuously adjusted by dragging the central slider handle. The second question is 'What style of communication do you want?', with 'Relaxed' labels on the left and 'Strict' on the right, allowing users to select their preferred communication tone. The third question is 'What type of instruction do you want?', with the 'Repetitive' label displayed as the default on the left and an empty selection point on the right, allowing users to choose a specific instruction method.

[0253] A check-shaped indicator is displayed at the center point of each slider to visually highlight the current selected value. When the user moves the handle, the processor (140) can read the slider position in real time, convert it into a continuous propensity score, and store it in temporary memory. The outline of the propensity input card is marked with a line that contrasts with the background to ensure visibility, and the overall screen layout is a responsive design so that the aspect ratio can be maintained at various resolutions.

[0254] This diagram can visually explain that a member can set preference tags in detail and quickly by intuitively entering consecutive values ​​for three questions through a slider-based interface.

[0255] Figure 12 illustrates an exemplary UI configuration of a trainer recommendation screen that highlights and presents recommendation results. Within a frame representing the exterior of a smartphone, the display area can be arranged in a continuous sequence of a top status bar and header, a recommendation highlight card, an information guide banner, a list of nearby trainers, and a bottom tab bar navigation.

[0256] The top status bar displays the time, battery, and signal icons, allowing users to check the basic device status. In the header directly below the status bar, the title "Find a Trainer" is centered alongside the app logo, enabling users to clearly identify the current feature section. Personalized text, such as "We found a trainer that is perfect for you," is displayed as a subtitle below the title to enhance the credibility of the recommendation results.

[0257] The recommended highlight card, displayed after the header, can concisely arrange key information such as rank badges, trainer illustrations, names, activity locations, specialty tags, and weekly session stats. The card can be designed so that pressing the entire touch area links to a detailed profile. A separate 'View Details' button is placed at the bottom of the card, allowing users to clearly identify the action.

[0258] An information banner can be placed below the recommendation card. The banner can provide a new member guide by displaying a lightbulb icon along with text such as "Learn tips for finding a good trainer." A cache flag can be set to prevent the banner from being re-exposed during the same session if it is closed by pressing the X icon in the upper right corner.

[0259] At the middle and bottom of the screen, a list of 'Neighborhood Trainers' is displayed in card form, providing additional distance-based recommendation information. Each list card displays a thumbnail, name, one-line description, availability, and an active gym icon, and can be immediately added to a wishlist by pressing the favorites button on the right.

[0260] The bottom tab bar navigation consists of 'Home', 'Messages', 'Wishlist', and 'Profile' icons. When the recommendations screen is active, the 'Home' tab is highlighted in an accent color to intuitively guide users to its location. Smooth fade animations are applied to transitions between tabs to minimize visual discontinuity.

[0261] Figure 13 illustrates the step-by-step UI flow of a trainer submitting qualification materials and being assigned a grade. On the first screen, the trainer's current visualization under the title "Average Grade of Trainers in Korea" allows for an intuitive check of their relative position. By pressing the "Check My Grade" button at the bottom of the screen, the user can proceed to the document submission stage.

[0262] The second screen is the 'Register / Edit Profile' stage, where trainers can upload supporting files for each category, such as Myeonghwi Classification, fitness documents, and motivational stories, and attach additional files if necessary. Clicking the X icon next to each file immediately deletes it, allowing for easy management of the document list, and the submission can be completed by pressing the 'Save and Next' button at the bottom.

[0263] The third screen is an animated display showing the progress of the rating review. A circular progress bar is displayed in the center to visualize the review progress in real time, and the text "Review in Progress" is positioned in the center, allowing the status to be checked without screen transitions.

[0264] The fourth screen is the rating results screen, displaying the final rating name along with the date the review was completed. Below the rating name, an evaluation summary and areas for improvement are provided as text blocks, allowing trainers to understand the direction for future preparation. At the bottom, the 'Rating Guide' and 'Check' buttons are placed side by side, allowing users to view an explanation of the rating system or download the detailed evaluation report.

[0265] Figure 14 summarizes the functional flow by user type in the mobile application of the present invention in a table format. In the left column, usage scenarios such as 'Main', 'Sign Up', 'Find Trainer', 'Find Gym', 'Chat', 'Payment', and 'My Page' are listed step by step, and detailed user roles such as consumer, trainer, and gym manager can be distinguished under each scenario. In the right column, the screen transition sequence and sub-function buttons for each step are listed as detailed items, allowing the entire service flow to be grasped at a glance.

[0266] For example, the main scenario can define the core navigation paths of the home screen by including '1.1 Burner Slide', '1.2 Find Trainer Link Button', '1.3 Find Gym Link Button', and '1.4 PT Guide Link Button'. In the sign-up scenario, role-specific registration procedures such as '1.9 Consumer Sign-up' and '1.11 Trainer Sign-up' are linked with the 'Service Guide and Main Page Navigation Button' to explain the onboarding flow step by step.

[0267] In the trainer search scenario, list details and TIP buttons are arranged sequentially, such as '1.15 Recommended Trainers', '1.16 Basic Information', and '1.17 Gym List', allowing the search -> details -> inquiry stages to flow naturally. In the chat scenario, 1:1 chat paths between consumers, trainers, and gyms are separated into distinct items, enabling clear definition of possible combinations within the messaging interface. In the payment stage, '1.21 Chat -> Payment' is listed, allowing for the implementation of an intuitive UX that transitions immediately to the payment screen after a real-time consultation.

[0268] In the My Page section, in addition to common menus, dedicated sub-menus for consumers, trainers, and gym managers may be displayed, and role-specific functions such as rating reassessment and settlement history may be listed in detail. This table-based flowchart is synchronized with screen design documents during the development phase to prevent feature omissions and can be utilized as a reference document for test cases during the QA phase. Through a hierarchical organization method similar to drawings, the entire service of the present invention can be systematically designed and managed.

[0269] That is, the present invention can unify the entire process from personal training matching, reservation, payment, and feedback by integrating trainer terminals, member terminals, gym manager terminals, and system manager terminals into a cloud-based platform. The processor (140) can simultaneously achieve advanced recommendation and operational automation by collecting and preprocessing data generated from each terminal in real time and structuring it in a database.

[0270] In some examples, machine learning-based AI recommendation models can improve matching accuracy by representing member preference tags and trainer profile tags as multiple embeddings, thereby simultaneously considering content similarity and collective preference patterns. A hybrid structure combining deep learning, collaborative filtering, and meta-learning can stably estimate initial fitness scores even in cold start situations. A dual-path learning strategy can enhance the level of real-time personalization by reflecting both long-term preference trends and short-term interest fluctuations.

[0271] In some cases, the trainer rating module can implement a fair and transparent compensation system by dynamically calculating scores based on qualifications, experience, reputation, and activity level, and automatically applying rates by grade. Rating updates incorporate the latest feedback through a combination of batch learning and online fine-tuning, enabling immediate risk mitigation in the event of quality degradation.

[0272] The session booking engine can streamline the user experience by handling trainer approval, gym space locking, payments, notifications, and reminders in a one-stop manner. Transaction locking and a dual storage structure can ensure data consistency and disaster resilience. Equipment inspection and settlement logic integrated with the gym manager interface can enhance facility operational efficiency.

[0273] In some cases, federated learning and differential privacy techniques can be applied to continuously improve model performance while protecting personal information. A / B testing frameworks can experimentally validate recommendation algorithms to systematically optimize key metrics such as click-through rates, booking rates, and refund rates. Fairness coefficients and diversity constraints can mitigate recommendation bias by maintaining balance in gender, age, and region.

[0274] As such, the present invention can simultaneously achieve high satisfaction and operational efficiency for both trainers and members by providing an integrated platform that encompasses data integration, AI personalized recommendations, rating management, reservation automation, personal information protection, and system fairness.

[0275] The embodiments according to the present disclosure described above may be implemented in the form of a computer program that can be executed through various components on a computer, and such a computer program may be recorded on a computer-readable medium. In this case, the medium may include a magnetic medium such as a hard disk, a floppy disk, and a magnetic tape, an optical recording medium such as a CD-ROM and a DVD, a magneto-optical medium such as a floptical disk, and a hardware device specifically configured to store and execute program instructions, such as a ROM, RAM, or flash memory.

[0276] Meanwhile, the above computer program may be one specifically designed and configured for the present disclosure, or one known and available to a person skilled in the art of computer software. Examples of computer programs may include machine code, such as that produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

[0277] In the specification of this disclosure (particularly in the claims), the use of the term "above" and similar descriptive terms may be in both singular and plural. Furthermore, where a range is described in this disclosure, it is to include an invention to which individual values ​​belonging to said range are applied (unless otherwise stated), as is equivalent to describing each individual value constituting said range in the detailed description of the invention.

[0278] Unless explicitly stated otherwise, the steps constituting the method according to the present disclosure may be performed in a suitable order. The present disclosure is not necessarily limited by the order in which the steps are described. The use of any examples or exemplary terms (e.g., etc.) in the present disclosure is merely for the purpose of describing the present disclosure in detail and, unless limited by the claims, the scope of the present disclosure is not limited by such examples or exemplary terms. Furthermore, a person skilled in the art will understand that various modifications, combinations, and changes may be made according to design conditions and factors within the scope of the claims or equivalents to which they are added.

[0279] Accordingly, the scope of the present disclosure is not limited to the embodiments described above, and all scopes equivalent to or equivalently modified from the claims set forth below, as well as the claims set forth below, shall be considered to fall within the scope of the scope of the present disclosure. Explanation of the symbols

[0280] 1 : Personal Training Brokerage System 100 : Personal training intermediary device 110: Communication interface 120 : User Interface 130 : Memory 140 : Processor 200 : User terminal 300 : Server 400 : Network

Claims

Claim 1 A big data-based trainer rating calculation and propensity tag-based AI recommendation matching method, wherein each step is performed by a processor, comprising the steps of: calculating an initial rating score of a trainer based on the trainer's qualification and career information according to a trainer rating calculation algorithm; linearly scaling a trainer-member matching success probability predicted by a machine learning model to convert it into an additive or subtractive value and applying it to the initial rating score to calculate a final rating score; determining the trainer's rating based on the final rating score; and registering the trainer whose calculated final rating score is above a preset standard in a platform database. A method for big data-based trainer rating and preference tag-based AI recommendation matching, comprising: receiving a request from a service user member to obtain member profile information and at least one preference tag; filtering trainer candidates that meet the member's preference tag and profile conditions; applying a machine learning-based AI recommendation model to the filtered trainer candidates to calculate a suitability score between the member and each trainer candidate, and ranking two or more trainer candidates according to the suitability score and providing them as a recommendation list to the member's user terminal; upon receiving a signal from the member to select one of the recommended trainers and a signal to select a desired affiliated exercise space, transmitting a matching request to the selected trainer to confirm the reservation of a personal training session between the trainer and the member; and collecting and storing feedback data regarding the trainer from the member after the personal training session. Claim 2 In claim 1, the big data-based trainer grade calculation and propensity tag-based AI recommendation matching method, wherein in the step of registering the trainer in the platform database, the trainer grade calculation algorithm is configured to calculate an initial grade score by summing a basic score assigned based on whether the trainer meets qualification requirements and a weighted score set based on the trainer's ability verification information. Claim 3 In claim 2, the trainer rating algorithm is configured to assign a basic score based on whether the trainer meets minimum qualification requirements, and to sum predefined weighted scores for each category of held certifications, experience, and area of ​​expertise, and additionally, to score quality indicators based on past or externally collectible member ratings, reviews, and claim counts for the trainer, and to add or subtract from them, thereby providing a big data-based trainer rating and propensity tag-based AI recommendation matching method. Claim 4 In claim 1, the step of registering the trainer in the platform database includes the step of setting and storing service rate information and PT session prices corresponding to the determined grade based on a pre-set grade-specific service rate and PT session pricing policy, a big data-based trainer grade calculation and propensity tag-based AI recommendation matching method. Claim 5 A big data-based trainer grade calculation and propensity tag-based AI recommendation matching method, wherein the step of registering the trainer in a platform database further includes the step of updating the trainer grade determined therein in a periodic or event-triggered manner according to a preset condition, and the preset condition includes at least one of the following: when the trainer's cumulative PT session count exceeds a preset threshold; when the trainer's expertise is enhanced by acquiring a new certification or degree; when the rating, re-registration rate, and satisfaction indicators collected from members are above or below a preset standard; when the trainer's platform contribution indicator is above or below a preset standard; when the risk indicator is above or below a preset standard; when the recent activity frequency or dormancy period is above or below a preset standard; and when the deviation indicator between the matching suitability predicted by the AI ​​recommendation model and the actual feedback is above or below a preset standard. Claim 6 A big data-based trainer rating and propensity tag-based AI recommendation matching method, wherein the member’s propensity tag obtained in the step of obtaining the propensity tag is composed of keywords expressing exercise goals, preferred coaching styles, preferred trainer characteristics, or other requirements set by the member, and the member directly selects and inputs one or more tags, or the member’s preference is automatically inferred from the member’s past behavior log and assigned as a tag. Claim 7 In claim 1, the step of providing a recommendation list to the user terminal of the member, wherein the machine learning-based AI recommendation model is implemented as a deep learning algorithm, is trained to include content-based matching that considers the similarity between the member's propensity tag and the trainer profile tag, and collaborative filtering that considers past matching success data of multiple members, and is trained to predict a matching suitability score by considering collective statistical data including the matching success rate and satisfaction rating of multiple members together with the behavior log of an individual member. Claim 8 In claim 7, the method for big data-based trainer rating calculation and propensity tag-based AI recommendation matching, wherein in the step of providing a recommendation list to the user terminal of the member, the machine learning-based AI recommendation model is a hybrid recommendation algorithm combining a content-based filtering module that performs cosine similarity or inner product operations between a trainer profile tag and a member propensity tag, and a collaborative filtering module that learns latent factors by decomposing a matching rating matrix of multiple users or predicts preferred trainers for similar member groups. Claim 9 A big data-based trainer rating calculation and propensity tag-based AI recommendation matching method according to claim 1, further comprising the step of reflecting the feedback data and member-trainer matching logs as training data for the machine learning-based AI recommendation model to update it in order to improve the accuracy of future trainer recommendations. Claim 10 A big data-based trainer rating calculation and propensity tag-based AI recommendation matching device comprises: a memory; and at least one processor connected to the memory and configured to execute computer-readable commands included in the memory, wherein the at least one processor calculates an initial rating score of a trainer based on the trainer's qualification and career information according to a trainer rating calculation algorithm, calculates a final rating score by applying a linearly scaled trainer-member matching success probability predicted by a machine learning model to an additive or subtractive value, and determines the trainer's rating based on the final rating score, and registers the trainer whose calculated final rating score is above a preset standard in a platform database. A big data-based trainer rating and preference tag-based AI recommendation matching device configured to receive a request from a service user member, obtain member profile information and at least one preference tag, filter trainer candidates that match the member's preference tag and profile conditions, apply a machine learning-based AI recommendation model to the filtered trainer candidates to calculate a suitability score between the member and each trainer candidate, rank two or more trainer candidates according to the suitability score and provide them as a recommendation list to the member's user terminal, and upon receiving a signal from the member to select one of the recommended trainers and a signal to select a desired affiliated exercise space, transmit a matching request to the selected trainer to confirm the reservation of a personal training session between the trainer and the member, and collect and store feedback data regarding the trainer from the member after the personal training session.

Images