Property Selection Support System

Abstract

This system provides a property selection support system that is suitable for obtaining property information that matches the user's potential and dynamically changing preferences. [Solution] The property selection support server 100 generates property vector data for each of the multiple properties based on the property information of that property. It also generates initial user preference vector data based on preference information extracted through interaction between the user and the AI ​​agent. It searches for property candidates based on the similarity between the property vector data and the user preference vector data and provides them to the user terminal 200. Then, it dynamically updates the user preference vector data based on operation information regarding the user's swipe operation on the provided property candidates.

Description

Technical Field

[0001] The present invention relates to a system for assisting in the selection of properties, and particularly to a property selection support system suitable for obtaining property information that conforms to the potential and dynamically changing preferences of users.

Background Art

[0002] Conventionally, as technologies for providing property information related to real estate, for example, the technologies described in Patent Documents 1 and 2 are known.

[0003] The technologies described in Patent Documents 1 and 2 perform matching processing using the desired conditions and attribute information of users (demanders), property information, and the attribute information of suppliers, and determine the possibility of a transaction being concluded. In Patent Documents 1 and 2, the attribute information of users and the attribute information of suppliers are each vectorized, and the cosine similarity between the vectorized attribute information of users and the vectorized attribute information of suppliers is considered to improve the matching accuracy.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] The technologies described in Patent Documents 1 and 2 had the following two problems. Firstly, in property selection, latent preferences that users cannot clearly articulate, such as design preferences and the atmosphere of the property, play a crucial role. However, the technologies described in Patent Documents 1 and 2 perform matching based on static information such as "desired conditions" and "attribute information" that users explicitly verbalize and input numerically, which makes it difficult to obtain property information that matches the user's latent preferences.

[0006] Secondly, a user's preferences can change dynamically as they browse various properties. However, the technologies described in Patent Documents 1 and 2 perform matching based on static information, which makes it difficult to obtain property information that matches the user's dynamic preferences. When a user's preferences change, they have to re-enter their desired conditions, which takes time to find the optimal property.

[0007] Therefore, the present invention has been made in view of the unresolved problems of the conventional technology, and aims to provide a property selection support system suitable for obtaining property information that matches the user's potential and dynamically changing preferences. [Means for solving the problem]

[0008] [Invention 1] To achieve the above objective, the property selection support system of Invention 1 includes: property selection means for selecting a property that suits the user's preferences from among a plurality of properties based on the degree of relevance of property feature data in a property feature data storage means that stores property feature data indicating the characteristics of each of a plurality of properties relating to real estate, and user preference feature data in a user preference feature data storage means that stores user preference feature data indicating the characteristics of the user's preferences; property information provision means for providing property information relating to the property selected by the property selection means to the user's terminal; operation information acquisition means for acquiring operation information relating to evaluation operations performed on the user's terminal for the property selected by the property selection means; and user preference feature data update means for updating the user preference feature data in the user preference feature data storage means based on the operation information acquired by the operation information acquisition means.

[0009] In this configuration, the property selection means selects a property that matches the user's preferences from among multiple properties based on the correlation between property feature data and user preference feature data, and the property information provision means provides property information about the selected property to the user's terminal. Then, the operation information acquisition means acquires operation information regarding the evaluation operation for the selected property, and the user preference feature data update means updates the user preference feature data based on the acquired operation information.

[0010] Here, property feature data includes any form of data capable of representing the characteristics of a property. Feature data can be structured as, for example, vector data (embedding representations) generated by machine learning models, feature maps, or parameters of statistical models. The same applies to user preference feature data.

[0011] Furthermore, as a measure of the correlation between property feature data and user preference feature data, for example, if the feature data is vector data, a similarity measure such as cosine similarity or dot product, or a distance measure such as Euclidean distance or Manhattan distance can be used.

[0012] Furthermore, the evaluation process includes any actions that allow users to intuitively input positive or negative evaluations of a property. Examples include swiping, tapping, flicking, or pressing buttons.

[0013] Furthermore, the property feature data storage means stores property feature data by any means and at any time. The property feature data may be stored in advance, or it may not be stored in advance, but rather stored in response to external input or other means during the operation of the system. The same applies to the user preference feature data storage means.

[0014] Furthermore, this system may be implemented as a single device, apparatus, terminal, or other device, or as a network system (e.g., a cloud system) in which multiple devices, apparatus, terminals, or other devices are connected in a communicative manner. In the latter case, each component may belong to any of the multiple devices, as long as they are connected in a communicative manner.

[0015] [Invention 2] Furthermore, the property selection support system of Invention 2 is the property selection support system of Invention 1, comprising: property feature data generation means for generating property feature data as vector data based on property information relating to the property; property feature data registration means for registering the property feature data generated by the property feature data generation means in the property feature data storage means; user preference feature data generation means for generating initial user preference feature data as vector data that can be mapped to the same vector space as the property feature data based on preference information relating to the user's preferences; and user preference feature data registration means for registering the user preference feature data generated by the user preference feature data generation means in the user preference feature data storage means, wherein the property selection means selects a property based on the similarity between the property feature data and the user preference feature data.

[0016] In this configuration, the property feature data generation means generates property feature data based on property information, and the property feature data registration means registers the generated property feature data in the property feature data storage means. Additionally, the user preference feature data generation means generates initial user preference feature data based on preference information, and the user preference feature data registration means registers the generated user preference feature data in the user preference feature data storage means. Finally, the property selection means selects a property based on the similarity between the property feature data and the user preference feature data.

[0017] [Invention 3] Furthermore, in the property selection support system of Invention 3, the property selection support system of Invention 2 is configured such that the user preference feature data generation means generates the initial user preference feature data based on the preference information extracted through dialogue between the user and the AI ​​agent.

[0018] With this configuration, the user preference feature data generation means generates initial user preference feature data based on preference information extracted through interaction between the user and the AI ​​agent.

[0019] [Invention 4] Furthermore, the property selection support system of Invention 4 is a property selection support system of any one of Inventions 1 to 3, wherein the user preference feature data updating means updates the user preference feature data so that when the evaluation related to the operation information is a positive evaluation, the user preference feature data approaches the features related to the property feature data of the property that was the subject of the operation related to the operation information, and updates the user preference feature data so that when the evaluation related to the operation information is a negative evaluation, the user preference feature data moves away from the features related to the property feature data of the property that was the subject of the operation related to the operation information.

[0020] With such a configuration, when the evaluation related to the operation information is a positive evaluation, the user preference feature data updating means updates the user preference feature data so as to approach the features related to the property feature data of the property that is the target of the operation related to the operation information. On the other hand, when the evaluation related to the operation information is a negative evaluation, the user preference feature data updating means updates the user preference feature data so as to move away from the features related to the property feature data of the property that is the target of the operation related to the operation information.

[0021] 〔Invention 5〕Furthermore, in the property selection support system of Invention 5, in the property selection support system of any one of Inventions 1 to 3, the operation is a swipe operation for switching the display of the property.

[0022] With such a configuration, the operation information acquisition means acquires operation information related to the swipe operation for the selected property.

Advantages of the Invention

[0023] As described above, according to the property selection support system of Invention 1, since the user preference feature data is updated based on the operation information related to the evaluation operation for the property, it is possible to obtain property information that conforms to the potential and dynamically changing preferences of the user as compared with the prior art.

[0024] Furthermore, according to the property selection support system of Invention 2, since the property is selected based on the similarity between the property feature data and the user preference feature data, which are vector data that can be mapped to the same vector space, it is possible to obtain property information that is more suitable for the potential and dynamically changing preferences of the user.

[0025] Furthermore, according to the property selection support system of Invention 3, since the initial user preference feature data can be generated by the dialogue between the user and the AI agent, it is possible to obtain property information that conforms to the user's preferences even at the initial stage.

[0026] Furthermore, according to the property selection support system of Invention 4, user preference feature data is updated to move closer to or further away from the features related to property feature data based on the evaluation of operation information, thereby enabling the acquisition of property information that better suits the user's latent and dynamically changing preferences.

[0027] Furthermore, according to the property selection support system of Invention 5, it is possible to grasp the user's potential and dynamically changing preferences through intuitive operation. [Brief explanation of the drawing]

[0028] [Figure 1] This is a block diagram showing the configuration of the network system according to this embodiment. [Figure 2] This diagram shows the hardware configuration of the property selection support server 100. [Figure 3] This diagram shows the data structures of the property information table 400, user information table 402, operation information table 404, property vector data table 406, and user preference vector data table 408. [Figure 4] This is a flowchart showing the process for generating property vector data. [Figure 5] Property search process flowchart. [Figure 6] This is a screenshot of the property listings displayed on user terminal 200. [Modes for carrying out the invention]

[0029] The embodiments of the present invention will be described below. Figures 1 to 6 show these embodiments. This embodiment will be explained using as an example a case in which a user who wishes to rent or buy a property searches for their desired property from among several properties related to real estate.

[0030] 〔composition〕 First, the configuration of this embodiment will be described. Figure 1 is a block diagram showing the configuration of the network system according to this embodiment.

[0031] As shown in Figure 1, the Internet 199 is connected to a property selection support server 100, which functions as a property selection support system, and a base station 210. The property selection support server 100 executes the main application logic, data management, and AI processing (vectorization, similarity search, etc.) of this system. The base station 210 is wirelessly connected to multiple user terminals 200 used by users and relays communication between the user terminals 200 and the Internet 199.

[0032] [Property Selection Support Server] Next, we will explain the configuration of the property selection support server 100. Figure 2 shows the hardware configuration of the property selection support server 100.

[0033] As shown in Figure 2, the property selection support server 100 consists of a CPU (Central Processing Unit) 30 that controls calculations and the entire system based on a control program, a ROM (Read Only Memory) 32 that stores the control program for the CPU 30 in a predetermined area, a RAM (Random Access Memory) 34 for storing data read from the ROM 32 and other memory, as well as calculation results necessary for the calculation process of the CPU 30, and an I / F (Interface) 38 that mediates data input and output to external devices. These components are connected to each other and enable data exchange via a bus 39, which is a signal line for data transfer.

[0034] I / F38 is connected to an external device consisting of an input device 40, such as a keyboard and mouse, which can input data as a human interface; a storage device 42, which stores data and tables as files; a display device 44, which displays a screen based on an image signal; and a communication device (not shown) for connecting to the Internet 199.

[0035] The property selection support server 100 can be configured, for example, as a cloud server (e.g., Google Cloud).

[0036] The CPU 30 implements various functions in this embodiment by executing programs stored in the storage device 42. This includes AI (Artificial Intelligence) processing functions. The AI ​​processing functions use a predetermined machine learning model to perform vectorization processing of property information and user preference information, as well as vector similarity search processing (for example, Vertex AI Matching Engine function). Furthermore, some processing (for example, the preference learning processing described later) can be configured to run on an event-driven execution environment (serverless environment) such as Cloud Functions.

[0037] The storage device 42 may consist of, for example, a hard disk drive or an SSD (Solid State Drive). Alternatively, the storage device 42 may be configured as a database (for example, a NoSQL database such as Firestore) or an index for vector searching (for example, an index for the Vertex AI Matching Engine).

[0038] [Data structure] Next, we will explain the data structure of the storage device 42. Figure 3 shows the data structures of the property information table 400, user information table 402, operation information table 404, property vector data table 406, and user preference vector data table 408.

[0039] The property information table 400 is a table for registering property information. As shown in Figure 3(a), the property information table 400 consists of a column for registering a property ID that uniquely identifies the property, a column for registering the name of the property, a column for registering basic property information (e.g., rent, floor plan, area, facilities, location), a column for registering descriptive text (e.g., a description of the property), and a column for registering image information (e.g., image data of the property or its URL). The property ID is the primary key.

[0040] User information table 402 is a table for registering user information. As shown in Figure 3(b), user information table 402 is composed of a column for registering a user ID that uniquely identifies the user, a column for registering user attribute information (e.g., age group, language), and a column for registering user preference information. The user ID is the primary key.

[0041] The operation information table 404 is a table for registering operation information. As shown in Figure 3(c), the operation information table 404 is composed of a column for registering an operation log ID that uniquely identifies the operation, a column for registering the user ID of the user who performed the operation, a column for registering the property ID of the property that was the target of the operation, a column for registering the operation type (for example, "LIKE (positive evaluation)" or "SKIP (negative evaluation)"), and a column for registering the date and time of the operation. The operation log ID is the primary key.

[0042] The property vector data table 406 is a table for registering property vector data. As shown in Figure 3(d), the property vector data table 406 is configured to have a column for registering the property ID and a column for registering vector data (e.g., a high-dimensional embedding representation) that indicates the characteristics of the property. The property ID is the primary key.

[0043] The user preference vector data table 408 is a table for registering user preference vector data. As shown in Figure 3(e), the user preference vector data table 408 is configured to have a column for registering the user ID and a column for registering vector data indicating the user's preferences. The user ID is the primary key.

[0044] [User terminal 200] Next, we will describe the configuration of user terminal 200. The user terminal 200 consists of portable devices such as smartphones and tablets. The hardware configuration consists of a CPU, ROM, RAM, and I / F connected via a bus. The I / F is connected to a touch panel (functioning as both an operation and display unit), storage device, camera, GPS (Global Positioning System), accelerometer, gyroscope, and wireless communication device. The user terminal 200 has a property search application (for example, a web application developed with Flutter Web, etc.) installed, and property information is displayed and swiped through this application.

[0045] [Operation] Next, the operation of this embodiment will be described. [Property vector data generation process] First, let's explain how to generate property vector data.

[0046] Figure 4 is a flowchart showing the property vector data generation process. The CPU 30 consists of an MPU (Micro-Processing Unit) and the like, and starts a predetermined program stored in a predetermined area of ​​the ROM 32, and executes the property vector data generation process shown in the flowchart of Figure 4 according to that program. The property vector data generation process is executed when property vector data is initially generated in bulk, or when property information is added or updated and property vector data is generated (for example, event-driven processing by Cloud Functions), and when it is executed on the CPU 30, it first proceeds to step S100, as shown in Figure 4.

[0047] In step S100, for the property from which property vector data is generated (hereinafter referred to as the "target property"), basic information, descriptive text, image information, and other information corresponding to the property ID are obtained from the property information table 400 as property information.

[0048] Next, the process moves to step S102, where the acquired multimodal property information is input into a machine learning model (for example, an embedding model such as Vertex AI Embedding) to generate high-dimensional property vector data that comprehensively represents the characteristics related to the basic information of the property, as well as latent characteristics including the design and atmosphere of the property or the atmosphere of the surrounding environment.

[0049] Next, the process moves to step S104, where the generated property vector data is associated with the property ID of the target property and registered or updated in the property vector data table 406, ending the series of processes and returning to the original process.

[0050] If there are multiple target properties, steps S100 to S104 are executed for each target property.

[0051] [Property search process] Next, we will explain how a user searches for properties. Figure 5 is a flowchart of the property search process.

[0052] Figure 6 shows the screen displaying property candidates on user terminal 200. The CPU 30 starts a predetermined program stored in a predetermined area of ​​the ROM 32 and executes the property search process shown in the flowchart of Figure 5 according to that program. The property search process is executed in response to a search request from the user terminal 200, and when it is executed by the CPU 30, it first proceeds to step S200, as shown in Figure 5.

[0053] In step S200, an AI agent is used to conduct a natural language chat with a user (hereinafter referred to as "target user") who is trying to search for properties, and the conversation history between the user and the AI ​​agent is obtained. The AI ​​agent can be implemented using, for example, a large language model (LLM; for example, OpenAI's GPT (Generative Pre-trained Transformer) or ChatGPT API).

[0054] Next, the process moves to step S202, where the acquired dialogue history is analyzed to extract preference information (e.g., keywords and summaries) related to potential preferences regarding the property's design and atmosphere or the surrounding environment, in addition to desired conditions such as rent, floor plan, area, facilities, and location. The extracted preference information is then associated with the target user's user ID and registered in the user information table 402.

[0055] Next, the process moves to step S204, where the extracted preference information is input into the same machine learning model as described above to generate high-dimensional vector data that comprehensively represents the target user's explicit preferences, such as desired property conditions, and the characteristics of the target user's latent preferences, including the design and atmosphere of the property or the atmosphere of the surrounding environment, and user preference vector data that can be mapped to the same vector space as the property vector data. This generates initial user preference vector data for the target user.

[0056] Next, the process moves to step S206, where the generated user preference vector data is associated with the user ID of the target user and registered in the user preference vector data table 408.

[0057] Next, the process moves to step S208, where, for each property vector data in the property vector data table 406, the similarity between the user preference vector and the property vector is calculated based on the target user's user preference vector data and that property vector data. A predetermined number of property information (for example, 10) corresponding to the property ID of that property are then retrieved from the property information table 400 in order of the calculated similarity. Here, for the first search, the initial user preference vector data generated in step S204 is used, while for subsequent searches, the user preference vector data updated in steps S216 and S220 is used.

[0058] Next, the process moves to step S210, where, based on the retrieved property information, property information in card format that can be swiped is generated on the user terminal 200, and the generated property information is sent to the user terminal 200. Upon receiving the property information, the user terminal 200 displays the property candidates in card format based on the received property information, as shown in Figure 6. The user can select or reject properties by swiping the property candidates. For example, swiping to the right on a property candidate results in a positive evaluation, while swiping to the left results in a negative evaluation.

[0059] Next, the process moves to step S212, where operation information regarding the user's swipe operation on the property candidates displayed on the user terminal 200 is received from the user terminal 200. Then, the received operation information is associated with the user ID of the target user and registered in the operation information table 404. The processing in step S212 is performed in real time, for example, via the Firebase / HTTPS API.

[0060] Next, the process moves to step S214, where it is determined whether the user's evaluation is positive based on the received operation information. If it is determined to be positive (YES), the process moves to step S216.

[0061] In step S216, the user preference vector data is updated to more closely resemble the features of the property vector data, based on the user preference vector data of the target user and the property vector data of the property candidate being manipulated. For example, the update is performed by calculating a weighted average of the user preference vector and the property vector.

[0062] Step S218 is used to determine whether a property has been found (for example, whether the user has performed the search completion operation). If a property has been found (YES), the series of processes is terminated and the process returns to the original state.

[0063] On the other hand, if it is determined in step S218 that no property is found (NO), the process proceeds to step S208. The process returns to step S208, and subsequent searches are performed.

[0064] On the other hand, if step S214 determines that the target user's evaluation is not positive (i.e., negative) (NO), the process proceeds to step S220, where the user preference vector data is updated based on the target user's user preference vector data and the property vector data of the property candidate being manipulated, so as to move away from the features related to that property vector data.

[0065] Once the processing in step S220 is complete, the process proceeds to step S218. 〔effect〕 Next, the effects of this embodiment will be described.

[0066] In this embodiment, the property selection support server 100 searches for property candidates based on the similarity between property vector data and user preference vector data, and provides them to the user terminal 200. Then, it dynamically updates the user preference vector data based on operation information related to the user's swipe operation on the provided property candidates.

[0067] This allows the system to track changes in user preferences in real time. Furthermore, users only need to perform intuitive operations, eliminating the need for cumbersome re-entry of conditions as in the past. Therefore, compared to conventional methods, it becomes possible to obtain property information that better matches the user's latent and dynamically changing preferences.

[0068] Furthermore, in this embodiment, the property selection support server 100 generates property vector data based on property information and generates initial user preference vector data that can be mapped to the same vector space as the property vector data, based on preference information. Then, it searches for properties based on the similarity between the two vectors.

[0069] This makes it possible to compare properties with latent preferences that are difficult to articulate using a unified scale (similarity in a vector space), and to obtain property information that better suits the user's latent and dynamically changing preferences. In particular, in this embodiment, property vector data is generated based not only on basic information but also on multimodal property information such as descriptive text and image information, so it is possible to accurately capture latent characteristics including the design and atmosphere of the property or the atmosphere of the surrounding environment.

[0070] Furthermore, in this embodiment, the property selection support server 100 generates initial user preference vector data based on preference information extracted through interaction between the user and the AI ​​agent.

[0071] This allows for the generation of initial user preference vector data through interaction between the user and the AI ​​agent, enabling the acquisition of property information that matches the user's preferences even in the initial stages.

[0072] Furthermore, in this embodiment, if the user's evaluation is positive, the property selection support server 100 updates the user preference vector data to approximate the features related to the property vector data, based on the user preference vector data of the target user and the property vector data of the property candidate being manipulated.

[0073] This makes it easier to select properties that are similar to those that have received positive ratings, thus allowing users to obtain property information that better suits their latent and dynamically changing preferences.

[0074] Furthermore, in this embodiment, if the user's evaluation is negative, the property selection support server 100 updates the user preference vector data to move away from the characteristics related to the property vector data, based on the user preference vector data of the target user and the property vector data of the property candidate being manipulated.

[0075] This makes it less likely that properties similar to those that have received negative ratings will be selected, allowing users to obtain property information that better suits their latent and dynamically changing preferences.

[0076] Furthermore, in this embodiment, a swipe operation is used as the user's operation to switch between displaying properties.

[0077] This allows users to understand their latent and dynamically changing preferences through intuitive operation, without having to perform complicated procedures.

[0078] In this embodiment, steps S100 and S102 correspond to the property feature data generation means of Invention 2, step S104 corresponds to the property feature data registration means of Invention 2, steps S200 to S204 correspond to the user preference feature data generation means of Invention 2 or 3. Furthermore, step S206 corresponds to the user preference feature data registration means of Invention 2, step S208 corresponds to the property selection means of Invention 1 or 2, step S210 corresponds to the property information provision means of Invention 1, and step S212 corresponds to the operation information acquisition means of Invention 1.

[0079] Furthermore, in this embodiment, steps S214, S216, and S220 correspond to the user preference feature data updating means of Invention 1 or 4, and the storage device 42 corresponds to the property feature data storage means of Invention 1 or 2, or the user preference feature data storage means of Invention 1 or 2. Also, the property vector data corresponds to the property feature data of Invention 1, 2, or 4, and the user preference vector data corresponds to the user preference feature data of Inventions 1 to 4.

[0080] [Variation] In the above embodiment and its modifications, a binary swipe operation of "LIKE" or "SKIP" was used as the evaluation operation. However, the system is not limited to this, and operations such as inputting three or more levels of evaluation (for example, 1 star to 5 stars) or operations with different weights, such as "Super Like" (especially high evaluation), can be used. In this case, the amount of update (distance moved) of the user preference vector data can be made variable according to the evaluation level and weight.

[0081] Furthermore, in the above embodiment and its modifications, user preference vector data was updated based on explicit user operation information. However, the system is not limited to this; "implicit feedback" such as the time spent on the property details screen, the number of times images were enlarged, the operation of transitioning to the map application, or the amount scrolled on the property details page can also be considered as positive evaluations, and the user preference vector data can be updated accordingly.

[0082] Furthermore, in the above embodiment and its modifications, operation information was uniformly used to update the user preference vector data regardless of when it was acquired. However, the invention is not limited to this, and a decay coefficient (forgetting coefficient) corresponding to the passage of time can be applied to past operation information to weight it so that the most recent operation is more strongly reflected in the user preference vector data.

[0083] Furthermore, in the above embodiment and its modifications, the user preference vector data was updated on a property-by-property basis. However, the method is not limited to this; it is also possible to focus on specific features of a property (for example, an image vector of the kitchen area or a text vector of "south-facing") and update the user preference vector data by weighting these specific feature vectors.

[0084] Furthermore, in the above embodiment and its variations, user preference vector data was updated based on nonverbal feedback obtained through swipe operations. However, the system is not limited to this. It is also possible to analyze the user's voice utterances (for example, "This room is too dark," or "The kitchen is good, but the bathroom is small") using speech recognition and natural language processing, determine whether the evaluation is positive or negative based on the analysis results, and selectively update specific attribute vectors (for example, brightness vectors or equipment vectors).

[0085] Furthermore, not limited to the above embodiments and their variations, a configuration can be adopted in which, in cooperation with a map application, the user updates user preference vector data by tapping, swiping, or performing other operations on property pins displayed on the map, and based on the update results, the display / hide, color-coding, or other aspects of other property pins displayed on the map are changed in real time (for example, filtering out areas or properties that do not match the user's preferences from the map).

[0086] Furthermore, in the above embodiments and their variations, user preference vector data was updated according to predetermined rules based on user evaluations. However, the user preference vector data can also be updated using reinforcement learning based on user evaluations. For example, a configuration can be adopted in which user actions (LIKE / SKIP / inquiry) are defined as "rewards," and the recommendation policy is learned and updated to maximize long-term rewards (conversion rates) using algorithms such as the Multi-Armed Bandit problem or Deep Reinforcement Learning.

[0087] Furthermore, while initial user preference vector data was generated through dialogue with an AI agent in the above embodiment and its modifications, the system is not limited to this. Initial user preference vector data can also be generated based on external data such as information on properties the user has previously contracted, content of SNS (Social Networking Service) posts and images that the user has "liked," or activity history from other linked lifestyle apps. Additionally, initial user preference vector data can be generated based on the user's answers to a predetermined number of questions (for example, a housing personality diagnosis consisting of six simple questions). Moreover, initial user preference vector data can also be generated by combining these elements (dialogue with an AI agent, external data, and answers to questions, etc.).

[0088] Furthermore, in the above embodiments and their modifications, basic information, descriptive text, image information, and other information were used to generate property vector data. However, the method is not limited to these, and multimodal information such as 360-degree panoramic images of the property, 3D model data (VR data), property introduction videos, or ambient sound data (noise level, etc.) of the area surrounding the property can also be used.

[0089] Furthermore, in the above embodiment and its modifications, the search was performed in real time while the user was operating the application. However, the system is not limited to this, and can also be configured to operate in the background, periodically or in an event-driven manner, calculate the similarity between the user preference vector and the property vector based on stored user preference vector data and property vector data of newly registered properties, and send a push notification to the user terminal 200 when the similarity exceeds a predetermined threshold.

[0090] Furthermore, not limited to the above embodiments and their variations, it is possible to adopt a configuration that incorporates an advertising delivery model that, for specific properties (sponsored properties) for which advertising fees have been paid, provides preferential treatment in the recommendation list or highlights them when the similarity between the property vector and the user preference vector exceeds a predetermined level.

[0091] Furthermore, while the above embodiment and its modifications provided recommendations for a single user, the system is not limited to this. When multiple users, such as a couple or friends sharing an apartment, jointly search for a property, the preference vector data of each user can be integrated (for example, by averaging or combining using AND conditions) to recommend a property that suits the preferences of the entire group.

[0092] Furthermore, while the above embodiment and its modifications employ a content-based filtering method that recommends properties based solely on the user preference vector data of the target user, it is also possible to employ a collaborative filtering method that identifies other user groups (clusters) with similar preference vector data to the target user and recommends properties highly rated by those other user groups to the target user. Alternatively, a composite method combining the content-based filtering method and the collaborative filtering method can be employed.

[0093] Furthermore, while the above embodiment and its modifications display properties one by one in card format, the system is not limited to this. A pairwise approach can be adopted, where two properties are displayed side-by-side and the user is asked to select which one they prefer. This configuration makes it easier to extract relative preferences even when absolute evaluation is difficult, and accelerates the convergence of user preference vector data.

[0094] Furthermore, while the above embodiment and its modifications selected properties solely based on vector similarity, the system is not limited to this. It can employ a more complex search method that involves pre-filtering properties based on user-specified requirements (e.g., maximum rent, pet-friendly status, or specific area) and then performing a search based on vector similarity on the extracted property group.

[0095] Furthermore, although the above embodiment and its variations did not provide reasons for recommending a property, it is possible to adopt a configuration in which the reasons why a particular property was recommended (for example, "You tend to prefer open kitchens and exposed concrete walls") are generated through component analysis of vector data or verbalization using LLM (Large Language Model) and presented to the user along with the property information.

[0096] Furthermore, while the above embodiment and its variations involve a single service provider offering services to the user, the system is not limited to this. A B2B (Business to Business) configuration can be adopted in which the functions of the property selection support server 100 are provided as an API (Application Programming Interface) to external real estate portal sites and management companies, enabling the use of a common user preference vector across diverse platforms.

[0097] Furthermore, not limited to the above embodiments and their variations, the learned user preference vector data can be linked to services in other industries (alliance partners) and used as input data for cross-selling (recommending related products) furniture, home appliances, or moving plans that match the user's preferences (for example, preferred interior style or lifestyle).

[0098] Furthermore, in the above embodiment and its modified form, the property selection support server 100 is configured as a single server, but it is not limited to this and can be configured as multiple servers or other devices. Also, it is not limited to a configuration that uses the internal storage device 42, but can adopt a configuration that uses an external storage device such as a database server. In addition, it is possible to adopt a configuration in which part of the vector calculation is performed on the user terminal 200 side using edge computing technology.

[0099] Furthermore, while the above embodiments and their modifications have described their application to a network system consisting of the Internet 199, the invention is not limited to this, and may also be applied to, for example, a so-called intranet that communicates using the same method as the Internet 199. Of course, it is not limited to networks that communicate using the same method as the Internet 199, but can be applied to any network using any communication method.

[0100] Furthermore, in the above embodiment and its modifications, the learning process was centrally performed on the property selection support server 100. However, the system is not limited to this, and it is possible to employ a configuration in which a preference model is updated within the terminal and only the model parameters are synchronized, by using technologies such as federated learning, which performs part of the learning process on the user terminal 200, without sending private information such as detailed personal operation history to the server.

[0101] Furthermore, in the above embodiments and their modifications, the process shown in the flowcharts of Figures 4 and 5 was described in the case of executing a program that is pre-stored in ROM 32. However, the invention is not limited to this, and a program describing these procedures may be read into RAM 34 from a storage medium in which the program is stored and then executed.

[0102] Furthermore, the above embodiments and their modified forms (including their respective constituent technologies) are mutually applicable.

[0103] Furthermore, while the above embodiments and their modifications applied the present invention to the case where a user searches for a desired property, the invention is not limited to this and can be applied to other cases as follows.

[0104] (1) This can be applied when matching any type of space, such as private lodging facilities, hotels and other accommodations, co-working spaces, rental conference rooms, or monthly parking lots.

[0105] (2) This can be applied to other fields where user sensibilities and latent preferences are important, such as talent placement (matching job seekers with job information), product recommendations on e-commerce sites (fashion, interior design, etc.), or matching used car sales. [Explanation of symbols]

[0106] 100…Property selection support server, 30…CPU, 32…ROM, 34…RAM, 38…I / F, 39…Bus, 40…Input device, 42…Storage device, 44…Display device, 199…Internet, 200…User terminal, 210…Base station, 400…Property information table, 402…User information table, 404…Operation information table, 406…Property vector data table, 408…User preference vector data table

Claims

1. Property feature data storage means for storing property feature data that indicates the characteristics of each of multiple properties relating to real estate, as vector data generated based on multimodal property information including basic information, descriptive text, and image information of the property, A user preference feature data generation means generates initial user preference feature data that indicates the characteristics of the user's preferences as vector data that can be mapped to the same vector space as the property feature data, based on the user's preference information extracted through the interaction between the user and the AI ​​agent, information on the property contracted by the user, the user's behavioral history, or the user's answers to predetermined questions. A user preference feature data storage means for storing the user preference feature data generated by the user preference feature data generation means, A property selection means that selects a property from among the multiple properties that matches the user's preferences based on the degree of relevance between the property feature data in the property feature data storage means and the user preference feature data in the user preference feature data storage means. A property information provision means that provides property information relating to the property selected by the property selection means to the user's terminal, An operation information acquisition means for acquiring operation information related to evaluation operations performed on the user's terminal for the property selected by the property selection means, The system includes a user preference feature data updating means that updates the user preference feature data in the user preference feature data storage means based on the operation information acquired by the operation information acquisition means, The property selection means is characterized by selecting a property that suits the user's preferences from among the multiple properties based on the degree of relevance between the property characteristic data and the user preference characteristic data updated by the user preference characteristic data updating means.

2. In claim 1, A property selection support system characterized in that, when the evaluation of the operation information is positive, the user preference feature data is updated so that the user preference feature data approaches the features of the property feature data of the property that was the subject of the operation related to the operation information, and when the evaluation of the operation information is negative, the user preference feature data is updated so that the user preference feature data moves away from the features of the property feature data of the property that was the subject of the operation related to the operation information.

3. In either claim 1 or 2, The property selection support system is characterized in that the operation is a swipe operation to switch the display of the property.

Images