An online property-based value assessment system

Abstract

The application provides an online value evaluation system based on real estate, and belongs to the technical field of online evaluation, and comprises: an online data acquisition module, which is used for acquiring a basic feature vector and multi-modal evaluation data of a target real estate online; a feature extraction and mining module, which is used for spatial mapping of explicit features and defect features, potential correlation mining of implicit features, and obtaining high-correlation implicit descriptions in combination with the basic feature vector; an evaluation processing module, which is used for identifying real evaluation, false evaluation and fuzzy evaluation, correcting the fuzzy evaluation after preliminary analysis according to the spatial mapping result and the high-correlation implicit description in combination with social platform theme discussion data; and a value evaluation module, which is used for obtaining a comprehensive value according to a basic value baseline, real evaluation and corrected evaluation, and outputting a value evaluation report. The problems of difficult identification of real and false evaluation and difficult solution of fuzzy evaluation are effectively solved, and the evaluation result is comprehensive and accurate.

Description

Technical Field

[0001] This invention relates to the field of online valuation technology, and in particular to an online valuation system for real estate. Background Technology

[0002] With the development of the real estate market, the demand for property valuation is increasing. Currently, existing property valuation methods mostly focus on the basic attributes of a property, such as its location, building area, unit type, and year of construction, calculating its basic value based on these attributes. However, in actual transactions and market perception, property value is also influenced by additional factors such as user evaluations, including assessments of the surrounding environment, property services, and transportation convenience.

[0003] While some online property valuation platforms collect user reviews, they suffer from the following shortcomings: First, they cannot effectively identify and filter fake reviews, leading to an impact on the accuracy of the valuation results. Second, they struggle to effectively process vague reviews (such as vague descriptions like "the surrounding environment is nice" or "transportation is convenient"), failing to transform them into quantifiable value factors. Third, they do not systematically integrate genuine, fake, and vague reviews as added value into the property valuation system, resulting in incomplete and inaccurate valuation results.

[0004] Therefore, this invention proposes an online property valuation system. Summary of the Invention

[0005] This invention provides an online property valuation system to solve the aforementioned technical problems.

[0006] This invention proposes an online property valuation system, comprising:

[0007] The online data acquisition module is used to acquire the basic feature vector and multimodal evaluation data of the target property online through a distributed crawler cluster and open API interface. The multimodal evaluation data includes: text evaluation, image evaluation and social platform topic discussion data.

[0008] The feature extraction and mining module is used to extract explicit and implicit features from the text evaluation, extract defect features from the image evaluation, perform spatial mapping between the explicit features and defect features, mine potential associations of the implicit features, and obtain a highly correlated implicit description by combining the basic feature vector.

[0009] The evaluation processing module is used to identify real, false, and fuzzy evaluations of the text and image evaluations based on a pre-built evaluation classification model, eliminate the false evaluations, perform preliminary analysis on the fuzzy evaluations, and correct the fuzzy evaluations after preliminary analysis based on spatial mapping results, highly correlated implicit descriptions, and social platform topic discussion data.

[0010] The value assessment module is used to obtain gridded transaction data of the same area as the target property to obtain a basic value baseline. At the same time, it obtains weighted evaluation factors for different value types based on a deep belief network, and obtains a comprehensive value based on the basic value baseline, the actual evaluation and the corrected evaluation, and outputs a value assessment report.

[0011] Preferably, the basic feature vector includes: location entropy value, spatial topological features, physical attenuation coefficient, and corresponding efficiency index.

[0012] Preferably, the evaluation processing module includes:

[0013] The cross-validation unit is used to cross-validate the identified fuzzy evaluations with the user identity information stored on the blockchain. If the user has not completed real-name authentication or the authentication information has no geographical association with the property, a false and irrelevant label is assigned to the fuzzy evaluation; otherwise, a potentially false label is assigned to the fuzzy evaluation.

[0014] The unit may be used to quantify the uncertainty of the fuzzy evaluation based on the correlation between historical fuzzy evaluation and actual transaction price fluctuations, convert it into expectation, entropy and hyperentropy, and analyze the expectation, entropy and hyperentropy based on a preset analysis mechanism, and assign possible reference labels or no reference labels.

[0015] The retention unit is used to retain the fuzzy evaluation only when the fuzzy evaluation is simultaneously assigned a possible false label and a possible reference label;

[0016] Otherwise, the fuzzy evaluations are discarded, and the fuzzy evaluations after preliminary analysis are the retained fuzzy evaluations.

[0017] Preferably, the evaluation processing module further includes:

[0018] A primary correction unit is used to map the fuzzy evaluation after the preliminary analysis to obtain a quantized ternary fuzzy array by fuzzy element mapping, and to correct the ternary fuzzy array by combining the fuzzy membership constraints of the basic feature vector. The fuzzy elements include: object elements, feature elements and quantization elements.

[0019] The contribution determination unit is used to extract the comparison relationship between the core value description and spatial mapping results of the real evaluation and the highly correlated implicit description, and combine it with the social platform topic discussion data to obtain the value factor of the core value description, and combine it with the exponential decay weight of the evaluation release time to determine the first contribution value.

[0020] The element labeling unit is used to assign labels to the quantified fuzzy elements based on the matching relationship between the core value description and the element evaluation description of the fuzzy evaluation after preliminary analysis, and based on the corresponding first contribution value. The labels are strong correlation labels, weak correlation labels, and no correlation labels.

[0021] The secondary correction unit is used to perform secondary correction on the ternary fuzzy array based on the label results;

[0022] The correction evaluation unit is used to obtain a correction evaluation based on the results of the first correction and the second correction.

[0023] Preferably, the feature extraction and mining module includes:

[0024] A function construction unit is used to construct an expectation function for implicit features, and to determine the depth level of potential association mining based on the expectation function, and to mine from a preset knowledge graph based on the depth level.

[0025] The associated feature extraction unit is used to extract associated features of the mining results at each level based on the basic feature vector, and to perform descriptive transformation on the associated features at each level to obtain a highly associated implicit description.

[0026] Preferably, the feature extraction and mining module further includes:

[0027] The lateral mapping unit is used to laterally map the explicit features to the defect features to determine the horizontal matching degree. ,in, To display the horizontal offset between the three-dimensional coordinate points of the feature and the three-dimensional coordinate points of the defect feature; The width of the wall corresponding to the explicit feature;

[0028] The vertical mapping unit is used to vertically map the explicit features to the defect features to determine the vertical matching degree. ,in, This represents the vertical offset between the three-dimensional coordinates of the explicit feature and the three-dimensional coordinates of the defect feature. The wall height corresponding to the explicit feature;

[0029] The diagonal mapping unit is used to diagonally map the explicit features to the defect features and determine the diagonal matching degree. , where d is the absolute distance difference between the three-dimensional coordinates of the explicit feature and the three-dimensional coordinates of the defect feature to the diagonal; This represents the standard deviation of the wall dimensions. The length of the diagonal of the space in which the explicit feature is located;

[0030] The temporal mapping unit is used to perform temporal mapping between the explicit features and the defect features to determine the temporal matching degree. ,in, The maximum allowable time difference is set to 365 days. This corresponds to the time the evaluation was published; For the present time; This refers to the actual time when the defect began to appear; The vertical height of the wall is based on the current time. The vertical height of the wall based on the evaluation release time; The difference between the current time and the start time of the event corresponding to the explicit feature; For error weights; The attenuation coefficient;

[0031] The horizontal matching degree Mx, the vertical matching degree My, the diagonal matching degree Md, and the temporal matching degree Ms are used as the spatial mapping results.

[0032] Preferably, the contribution determination unit includes:

[0033] The positioning subunit is used to locate attribute words and feature words in the actual evaluation and construct a structured description. ,in, The semantic vector representing the true evaluation;

[0034] Associated subunits are used to determine the structured description. Attribute correlation with the basic feature vector And retain descriptions with attribute relevance greater than the preset relevance to form a core value description set. ,in, This represents the j-th core value description; m represents the number of core value descriptions.

[0035] The graph construction subunit is used to construct a three-layer heterogeneous graph G=(V,E) based on the correspondence between the core value description, spatial mapping results, and highly correlated implicit description. The first layer consists of core value description nodes, the second layer consists of spatial mapping result nodes, and the third layer consists of highly correlated implicit description nodes. V represents the nodes involved in the three layers, and E represents the edges corresponding to the nodes.

[0036] The three-domain extraction subunit is used to extract three-domain features associated with the core value description from the topic discussion data of the social platform. The three-domain features include: the sentiment domain, the dissemination domain, and the spatiotemporal domain.

[0037] The factor determination subunit is used to obtain the value factors describing the core value based on the three-layer heterogeneous graph and the three-domain features.

[0038] Preferably, the value assessment module includes:

[0039] The benchmark correction unit is used to correct the basic value benchmark line based on the spatiotemporal decay coefficient.

[0040] The factor analysis unit is used to assign initial weights to the weight evaluation factors output by the deep belief network according to the feature category, and at the same time, capture the temporal correlation between the weight evaluation factors.

[0041] The comprehensive acquisition unit is used to obtain the comprehensive value based on the corrected baseline, initial weights and time series correlation, the value contribution quantification value of the true evaluation, and the value contribution quantification value of the corrected evaluation.

[0042] Compared with the prior art, the beneficial effects of this application are as follows:

[0043] By integrating multimodal data (text, images, and topic discussions) with basic features, and combining spatial mapping to verify the authenticity of the evaluation, a deep learning model is used to extract key influencing factors, dynamically correct fuzzy evaluations, and finally generate a comprehensive value assessment report. This effectively solves the problems of difficulty in distinguishing between real and fuzzy real estate evaluations, and makes the evaluation results comprehensive and accurate.

[0044] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0045] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0046] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0047] Figure 1 This is a structural diagram of an online real estate valuation system according to an embodiment of the present invention;

[0048] Figure 2 This is a structural diagram of the heterogeneous diagram in an embodiment of the present invention. Detailed Implementation

[0049] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0050] This invention proposes an online property valuation system, such as... Figure 1 As shown, it includes:

[0051] The online data acquisition module is used to acquire the basic feature vector and multimodal evaluation data of the target property online through a distributed crawler cluster and open API interface. The multimodal evaluation data includes: text evaluation, image evaluation and social platform topic discussion data.

[0052] The feature extraction and mining module is used to extract explicit and implicit features from the text evaluation, extract defect features from the image evaluation, perform spatial mapping between the explicit features and defect features, mine potential associations of the implicit features, and obtain a highly correlated implicit description by combining the basic feature vector.

[0053] The evaluation processing module is used to identify real, false, and fuzzy evaluations of the text and image evaluations based on a pre-built evaluation classification model, eliminate the false evaluations, perform preliminary analysis on the fuzzy evaluations, and correct the fuzzy evaluations after preliminary analysis based on spatial mapping results, highly correlated implicit descriptions, and social platform topic discussion data.

[0054] The value assessment module is used to obtain gridded transaction data of the same area as the target property to obtain a basic value baseline. At the same time, it obtains weighted evaluation factors for different value types based on a deep belief network, and obtains a comprehensive value based on the basic value baseline, the actual evaluation and the corrected evaluation, and outputs a value assessment report.

[0055] Preferably, the basic feature vector includes: location entropy value, spatial topological features, physical attenuation coefficient, and corresponding efficiency index.

[0056] In this embodiment, the distributed crawler cluster is a distributed data acquisition system composed of multiple servers. It improves efficiency through parallel crawling and avoids being restricted by a single IP. For example, 10 servers are deployed, each running 5 crawler processes to crawl housing information from platforms such as Lianjia and Beike, while also crawling discussions about the target community on Weibo and Zhihu. Specifically, it is built based on the Scrapy-Redis framework, uses an IP proxy pool (such as Abu Cloud Proxy) to dynamically switch IPs, and sets a crawling frequency threshold (such as ≤30 times per minute for a single domain name) to avoid triggering anti-crawling measures.

[0057] In this embodiment, the open API interface is a standardized data interface provided by a third-party platform, used to legally obtain authorized data. For example, calling the Gaode Map API to obtain supporting facilities (schools, hospitals, subway entrances) within 3 kilometers of the target property; calling the open API of the housing and construction department to obtain the year of construction and building area of ​​the property registration. Specifically, it involves: authenticating with the API key, requesting according to the interface document specifications (such as RESTful format), and setting a timeout retry mechanism (maximum 3 times).

[0058] In this embodiment, the target property is the specific house to be evaluated, which has a unique identifier (such as property certificate number, community name + unit number + room number), for example, Room 501, Unit 2, Building 3, XX Community, Chaoyang District, Beijing.

[0059] In this embodiment, the basic feature vector is a set of quantitative data describing the physical and locational attributes of a property. It is formed into a vector by dimension, such as [location entropy value 0.85, spatial topology feature 0.72, physical attenuation coefficient 0.3, supporting efficiency index 0.9] (each value is the result of normalization processing). Specifically, it is extracted through structured fields (such as building area obtained from filing information) and calculated by quantitative models (such as location entropy value = number of business districts / sum of squared distances).

[0060] In this embodiment, the multimodal evaluation data is various types of user evaluation data, covering text, images, topic discussions, and other forms.

[0061] Text reviews: User-submitted text descriptions, such as "The property management in our community is terrible; the garbage hasn't been collected for a week."

[0062] Image evaluation: Photos / videos of houses uploaded by users, such as close-ups of cracks in the living room wall and videos of the community's greenery.

[0063] Social platforms: Discussions about the target community on social media platforms (Weibo, Douyin), such as the popularity and sentiment of comments on the topic of "owners' rights protection in XX community".

[0064] In this embodiment, explicit features are the property attributes clearly described in the text evaluation, which can be directly quantified or located, such as "building area 120㎡", "master bedroom facing south" and "built in 2010". Specifically, it involves extracting numbers and directional words based on regular expressions and matching property attribute words with knowledge graphs (such as associating "facing south" with the "lighting" feature).

[0065] In this embodiment, implicit features are potential information that is not directly stated in the text evaluation but can be obtained through semantic reasoning. For example, "living comfortably" implies good lighting and sound insulation; "convenient to go to work" implies convenient transportation. Specifically, the semantic similarity between the text and preset feature words (such as lighting and transportation) is calculated using the BERT model. If the similarity is ≥0.7, the corresponding feature is associated.

[0066] In this embodiment, the defect features are physical defects of the building identified from the image evaluation, such as structural damage and aging facilities. For example, the wall crack is 1.2m long, the broken window glass area is 0.05㎡, and there are water seepage marks in the bathroom. Specifically, the YOLOv8 model is used to detect the defect area in the image, and the defect contour is segmented by MaskR-CNN. The defect size (length and area) is calculated and normalized (e.g., crack length / wall height).

[0067] In this embodiment, spatial mapping involves matching the spatial location described in the text with the location of defects in the image to verify the authenticity of the description. For example, the text "There is a crack 1.5m above the ground on the east wall of the master bedroom" matches the crack feature at the same coordinate in the image (based on the 3D model of the house), and the matching degree is calculated to be 85%. Specifically, a 3D grid coordinate system of the house is constructed, the text location is resolved as (room ID, wall orientation, height), and the image defects are generated into 3D coordinates using SfM (Structure of Motion) technology. If the Euclidean distance error is ≤0.3m, a match is determined.

[0068] In this embodiment, latent association mining is carried out through knowledge graph reasoning to uncover the potential associations between implicit features and basic features.

[0069] In this embodiment, the highly correlated implicit description is an implicit feature description that is highly correlated with the basic feature vector and has practical reference value. For example, the implicit description "convenient to buy groceries" has a correlation of 0.82 with the basic feature "supporting efficiency index (supermarket density)" and is marked as highly correlated.

[0070] In this embodiment, the evaluation classification model is a machine learning model used to distinguish evaluation types (real, fake, and ambiguous). It integrates multi-dimensional features, such as a Bi-LSTM+GAT model, which takes text semantics, user behavior, and image defect features as input and outputs classification results (real evaluation probability 0.92, fake 0.05, and ambiguous 0.03). Specifically, it is trained with historical labeled data (100,000 real / fake / ambiguous evaluations), and features include sentiment entropy, IP concentration, and defect matching degree. An F1-score ≥ 0.85 is used as the model's pass threshold.

[0071] In this embodiment, the real evaluation is an evaluation published by a real homeowner that is consistent with the actual situation of the house. For example, a homeowner (real-name verified) may publish "The property management of the community is on duty 24 hours a day. Last year, the water leaked in my house and they arrived in 10 minutes", with a video of the water leak being repaired on site. After verification, it is consistent with the facts.

[0072] In this embodiment, false reviews are misleading reviews posted by non-real users, such as positive reviews from intermediaries or malicious negative reviews from competitors. For example, the same IP address may post five identical "excellent" reviews of the community within one hour, even if the user is not real-name authenticated and the distance between the IP address and the community is greater than 50km.

[0073] In this embodiment, fuzzy evaluation is an evaluation that is vague and difficult to quantify directly, such as "okay," "so-so," or "so-so." For example, the community environment is passable, the house is not old, and the price is about the same, without clearly specifying specific characteristics.

[0074] In this embodiment, the preliminary analysis is a preprocessing step for verifying the authenticity and filtering the reference value of fuzzy evaluations. For example, a fuzzy evaluation is based on blockchain verification that the user is the owner and has a geographical correlation of 0.8, but its correlation with historical transaction price fluctuations is only 0.3. It is judged to be retained but with low weight. Specifically, by combining blockchain identity verification (whether it is real-name authentication, distance between IP and property) and gray relational analysis (correlation with transaction data), fuzzy evaluations that simultaneously meet the requirements of identity credibility + correlation ≥ 0.5 are selected.

[0075] In this embodiment, the revised evaluation is the result of quantitatively correcting the fuzzy evaluation retained in the initial analysis by combining spatial mapping results, highly correlated implicit descriptions, and topic discussion data. For example, the fuzzy evaluation that the community is a bit noisy is corrected to moderate noise impact with an average nighttime decibel level of 58dB by combining spatial mapping (near the main road), implicit description of poor sleep at night, and topic discussion data (top 10% of surrounding noise complaints) with spatial mapping (near the main road), implicit description of poor sleep at night, and topic discussion data (top 10% of surrounding noise complaints). Specifically, the fuzzy description is converted into a quantitative score (e.g., a bit noisy corresponds to 55-65dB) using a cloud model, and the score range (58-68dB) is adjusted by combining the sentiment of topic discussion (60% negative).

[0076] In this embodiment, the gridded transaction data for the same area is obtained by dividing the area where the target property is located into grids (e.g., 500m×500m) and statistically analyzing the property transaction records (price, area, unit type, etc.) within each grid over the past 6 months. For example, if the target property is located in grid A of Chaoyang District, Beijing, there are 20 transactions in this grid over the past 6 months, with an average price of 65,000 yuan / ㎡ and 70% of the transactions being for units with an area of ​​100-120㎡. Specifically, the grids are divided using GeoPandas, transaction data is extracted from the databases of Lianjia and the Housing and Construction Bureau, aggregated by grid ID, and outliers (e.g., prices deviating from the mean by ±3σ) are removed.

[0077] In this embodiment, the baseline value is a property valuation calculated based on gridded transaction data within the same region, combined with basic feature vectors. For example, given the target property's basic features (120㎡, built in 2010, supporting facility efficiency index 0.8), and referring to transaction data within the same grid, a baseline value of 7.8 million yuan is calculated. Specifically, an XGBoost model is used to train the relationship between transaction prices and basic features. The target property features are input, and the baseline value is output. .

[0078] In this embodiment, a deep belief network (DBN) is a deep neural network used to extract high-weight value influencing factors from evaluation data. For example, by analyzing 100,000 evaluation data through DBN, it is found that "property level", "transportation convenience" and "school district quality" are the top 3 evaluation factors with the highest weights. Specifically, a 3-layer DBN is constructed (128-dimensional input layer, 64-dimensional hidden layer and 10-dimensional output layer), trained with evaluation data (quantified real evaluation and corrected evaluation), and outputs the weights of each factor (the sum of the weights is 1).

[0079] In this embodiment, the weighted evaluation factors are the evaluation dimensions that have a significant impact on the property value. The higher the weight, the greater the impact of the dimension on the valuation. For example, the weighted evaluation factors are "property quality (0.3), transportation convenience (0.25), school district quality (0.2), environmental noise (0.15), and green coverage rate (0.1)".

[0080] In this embodiment, the comprehensive value is the final valuation that integrates the basic value benchmark and evaluation factors (real evaluation and revised evaluation), reflecting the comprehensive market value of the property.

[0081] In this embodiment, the valuation report is a structured report that includes basic property information, valuation results, influencing factor analysis, and data sources. For example, the report includes the target property address, basic characteristics, comprehensive value of RMB 7.95 million (±5%), weighted evaluation factors (property, transportation), and data sources (transaction data, 120 real evaluations, and 30 corrected evaluations).

[0082] Location entropy is used to quantify the locational advantages of a property's location. It reflects the region's economic vitality and convenience of life by calculating the spatial agglomeration and radiation intensity of surrounding core resources (such as business districts, transportation hubs, and educational and medical facilities). The higher the value, the more superior the location (after normalization, ∈ [0,1]). For example, the location entropy of a school district property in Haidian District, Beijing, is calculated as follows: within 3 kilometers, there are 5 large business districts (weight 0.3), 3 subway stations (weight 0.2), 2 key primary schools (weight 0.3), and 1 tertiary hospital (weight 0.2), resulting in an entropy of 0.85 (high locational advantage). In contrast, a property in the suburbs has only one small supermarket and no subway access, resulting in a location entropy of 0.23 (low locational advantage). Specifically, distributed crawlers are used to collect resource POI data within a 1 / 3 / 5-kilometer radius (obtained through Gaode / Baidu Maps API).

[0083] Weights are assigned based on resource type (Education > Transportation > Commerce > Healthcare, with a total weight of 1).

[0084] The location entropy value is calculated using the location entropy formula. Wherein, resource density = resource quantity / area (square kilometers), and is normalized (divided by the maximum entropy value of the same city), where n1 represents the number of regions.

[0085] Spatial topology features describe the rationality and functionality of the internal spatial structure of a property. They are quantified through indicators such as the connectivity, squareness, and functional zoning of the apartment layout, reflecting the actual utilization efficiency of the house. The higher the value, the better the spatial structure (after normalization, ∈ [0,1]). For example, the spatial topology features of a three-bedroom apartment are calculated as follows: squareness (90%, wall turning angle close to 90°), connectivity of functional areas (living room and dining room are directly connected, distance between bedroom and bathroom is <5m), and space utilization (shared area accounts for 15%), with a comprehensive score of 0.82 (high-quality apartment). On the other hand, the spatial topology features of a "knife handle" shaped one-bedroom apartment (squareness 60%, chaotic circulation) are 0.35 (poor apartment). Specifically, the apartment structure data is extracted from the property CAD drawings or panoramic images (wall and door / window positions are identified through OpenCV edge detection).

[0086] Calculate core metrics:

[0087] Squareness = Area of ​​the smallest circumscribed rectangle of the apartment / Actual usable floor area;

[0088] Connectivity = Number of direct connections between functional areas (living room, bedroom, kitchen and bathroom) / Total number of possible connections;

[0089] Utilization rate = usable floor area / building area;

[0090] The analytic hierarchy process (AHP) was used to integrate the indicators, which were then weighted, summed, and normalized (weights: squareness 0.4, connectivity 0.3, utilization 0.3).

[0091] The physical attenuation coefficient measures the degree of physical wear and tear on a property due to its construction time, building material aging, and maintenance conditions. It reflects the durability of the house and its potential maintenance costs. The higher the value, the more severe the wear and tear (after normalization, ∈ [0,1], where 0 represents brand new and 1 represents severe aging). For example, a brick-concrete structure house built in 2000 has the following characteristics: 23 years of construction (weight 0.5), 30% of the exterior wall paint peeling area (weight 0.2), and 3 records of water pipe corrosion causing leaks (weight 0.3). The physical attenuation coefficient is calculated to be 0.68 (relatively high wear and tear). In contrast, a steel-concrete structure house built in 2020 (no maintenance records, building material aging rate 5%) has a physical attenuation coefficient of 0.12 (slight wear and tear). Specifically, the following basic data is collected: construction time (obtained from the housing and construction department's API), building material type (brick-concrete / steel-concrete / frame), maintenance records (property system crawler), and defect features in the image evaluation (such as cracks and leaks, identified through YOLOv8).

[0092] Calculate core metrics:

[0093] Aging rate = Years since construction / Building design life (50 years for reinforced concrete structures);

[0094] Defect percentage = Defect area in image / Total building area;

[0095] Repair frequency = Number of repair records in the past 5 years / 5 (annual average);

[0096] Weighted summation: Physical attenuation coefficient = aging rate × 0.5 + defect ratio × 0.3 + maintenance frequency × 0.2, normalized (divided by the maximum attenuation value of similar houses).

[0097] The supporting facilities efficiency index quantifies the comprehensive efficiency of the "accessibility + quality" of supporting facilities around a property, covering core facilities such as education, medical care, commerce, and transportation. It reflects the convenience of residents' daily lives, with higher values ​​indicating better quality facilities (normalized to ∈ [0,1]). For example, the supporting facilities efficiency index of a certain community is calculated as follows:

[0098] Education: 10-minute walk to a key municipal primary school (quality score 0.9, accessibility 0.9, weight 0.3).

[0099] Healthcare: 5 minutes by bike to the community hospital, 15 minutes by car to a top-tier hospital (quality score 0.7, accessibility 0.8, weight 0.2).

[0100] Commercial: 5-minute walk to supermarket, 10-minute walk to shopping center (quality rating 0.8, accessibility 0.9, weight 0.2).

[0101] Transportation: 3-minute walk to the bus stop, 10-minute walk to the subway station (Quality rating 0.85, Accessibility 0.95, Weight 0.3).

[0102] Comprehensive calculation: (0.9×0.9×0.3)+(0.7×0.8×0.2)+(0.8×0.9×0.2)+(0.85×0.95×0.3)=0.83 (high-efficiency matching).

[0103] Specifically, data is collected using a three-tiered indicator system: type (education / medical care, etc.), quality (star rating), and accessibility (time / distance). Quality rating: education is based on college entrance examination pass rate, medical care on top-tier hospital qualifications, and commerce on the richness of business formats (quantified from 1 to 5 stars). Accessibility: walking / cycling / driving time (obtained through the route planning function of the map API), normalized to [0,1] (the shorter the time, the higher the value). For each type of facility, a score of "quality × accessibility" is calculated, summed according to weights (education 0.3, transportation 0.3, commerce 0.2, medical care 0.2), and then normalized (divided by the city's highest facility efficiency value).

[0104] The beneficial effects of the above technical solution are: by integrating multimodal data (text, images, topic discussions) with basic features, combining spatial mapping to verify the authenticity of the evaluation, using deep learning models to extract key influencing factors, dynamically correcting fuzzy evaluations, and finally generating a comprehensive value assessment report, the solution effectively solves the problems of difficulty in distinguishing between real and fuzzy real estate evaluations, and makes the evaluation results comprehensive and accurate.

[0105] This invention proposes an online property valuation system, wherein the valuation processing module includes:

[0106] The cross-validation unit is used to cross-validate the identified fuzzy evaluations with the user identity information stored on the blockchain. If the user has not completed real-name authentication or the authentication information has no geographical association with the property, a false and irrelevant label is assigned to the fuzzy evaluation; otherwise, a potentially false label is assigned to the fuzzy evaluation.

[0107] The unit may be used to quantify the uncertainty of the fuzzy evaluation based on the correlation between historical fuzzy evaluation and actual transaction price fluctuations, convert it into expectation, entropy and hyperentropy, and analyze the expectation, entropy and hyperentropy based on a preset analysis mechanism, and assign possible reference labels or no reference labels.

[0108] The retention unit is used to retain the fuzzy evaluation only when the fuzzy evaluation is simultaneously assigned a possible false label and a possible reference label;

[0109] Otherwise, the fuzzy evaluations are discarded, and the fuzzy evaluations after preliminary analysis are the retained fuzzy evaluations.

[0110] In this embodiment, fuzzy evaluations are user-posted evaluations that are semantically vague and lack clear quantitative standards. They cannot directly determine the direction or degree of their impact on property value. For example, descriptions without specific indicators such as "the community environment is okay", "the house is not too old", and "the price is about the same".

[0111] The user identity information stored on the blockchain is real-name authentication data (immutable and traceable) stored on the blockchain, including name, ID number, geographical location, etc., used to verify the authenticity of the review publisher. For example, after a user completes real-name authentication on the review platform, the identity information is stored in a consortium blockchain (such as a dedicated consortium blockchain for real estate reviews) through hash encryption. Each piece of information is associated with a unique blockchain address and can be queried and verified through smart contracts.

[0112] Cross-validation verifies the correlation between a user's identity and their review by comparing data from multiple sources, ensuring that the reviewer and the property have a reasonable connection. For example, if a user reviews "the soundproofing in XX community is average", cross-validation shows that their permanent address on the blockchain is within 3 kilometers of the community, and there are matching delivery addresses in the past 6 months, thus determining that the association is valid.

[0113] Real-name authentication is the identity verification completed by users through official channels (such as the public security system) to prove that the account belongs to a real natural person. For example, when a user uploads a photo of their ID card and completes facial recognition on a rating platform, the information is verified by the public security interface and marked as "real-name authenticated".

[0114] Geographic association refers to a reasonable spatial relationship (usually within a short distance) between a user's identity information (such as permanent address, IP address) and the physical location of the target property. For example, if the target property is located in XX Community, Haidian District, Beijing, and the user's real-name authentication address is XX Street, Haidian District (within 3 kilometers), or the IP address of the review post is located in Haidian District, it is considered to have a geographic association. If the user's address is in Guangzhou City, Guangdong Province, there is no geographic association. Specifically, the straight-line distance between the user's address and the property is calculated using the Gaode Map API, a threshold is set (such as within 5 kilometers for a valid association), and the IP address (accurate to the city / district level) is used for auxiliary verification.

[0115] The "false and irrelevant" label is used to mark vague reviews that are "unreliable and unrelated to the property." This indicates that the review is highly likely to be false information and is unrelated to the actual situation of the property. For example, a post by an unverified user stating "XX community is terrible" and whose IP address is 100 kilometers away from the property would be labeled as "false and irrelevant."

[0116] The "potentially false" label is used to mark vague reviews that are "credible in identity but have potential for being fake." It indicates that the reviewer's identity is real, but the authenticity of the content needs further verification. For example, a user who has been verified by real name may post "The property management in my community is terrible" but without a specific event description, and the user has no consumption record in the community in the past year, thus being given the "potentially false" label.

[0117] Historical fuzzy evaluation data is a set of fuzzy evaluations of the same area and type as the target property within a certain period of time (e.g., 1 year). It is used to analyze the correlation between fuzzy evaluations and actual transactions. For example, 1,000 fuzzy evaluations of Haidian District, Beijing in the past year, such as "the house price is okay" and "the environment is average", can be correlated with the corresponding second-hand housing transaction data.

[0118] Actual transaction price fluctuations refer to the changes in the actual transaction prices of second-hand homes in the area where the target property is located over time, reflecting the market's true feedback on the property's value. For example, in a certain community, the average transaction price rose from 5 million yuan per unit to 5.3 million yuan per unit from January to June 2023, a fluctuation of 6%; from July to December, it dropped to 5.1 million yuan per unit, a fluctuation of -3.8%.

[0119] The correlation refers to the degree of association between historical fuzzy evaluation content and the fluctuation of transaction prices during the same period. The higher the correlation, the greater the reference value of the fuzzy evaluation for prices. For example, statistics show that when the number of fuzzy evaluations of "good environment" increases, the average housing price in the same area rises by 2%, with a correlation coefficient of 0.6 (moderate correlation); when the number of evaluations of "poor property" increases, the average housing price falls by 1.5%, with a correlation coefficient of 0.5. Specifically, the correlation between the frequency (or sentiment) of fuzzy evaluations and the fluctuation of transaction prices is calculated using the Pearson correlation coefficient. An absolute value of the coefficient ≥ 0.4 is considered to indicate a valid correlation.

[0120] Uncertainty quantification transforms the semantic uncertainty of fuzzy evaluations into calculable numerical features, using "expectation, entropy, and hyperentropy" to describe the degree of fuzziness. For example, the fuzzy evaluation "the lighting is okay" is quantified as follows: expectation = 0.6 (equivalent to "medium lighting"), entropy = 0.2 (degree of uncertainty), and hyperentropy = 0.05 (range of entropy fluctuation), indicating that the fuzziness of the evaluation is low. Among these features, expectation reflects the central trend of the evaluation, entropy reflects the degree of fuzziness, and hyperentropy reflects the stability of entropy.

[0121] The expected value is the average quantitative value of the fuzzy evaluation, reflecting the central trend of the evaluation (such as the average score corresponding to "okay"). For example, 100 fuzzy evaluations of "the community facilities are okay" correspond to an average score of 0.6 (out of 1) after quantification, so the expected value is 0.6.

[0122] Entropy describes the degree of uncertainty in fuzzy evaluations. The higher the value, the more fuzzy the semantics of the evaluation and the greater the disagreement. For example, the entropy of "the community environment is okay" is 0.1 (small disagreement), while the entropy of "the community environment is average" is 0.3 (some people think "average" is preferred, while others think it is bad, resulting in a large disagreement).

[0123] Hyperentropy describes the degree of entropy fluctuation and reflects the stability of uncertainty in fuzzy evaluation. The higher the value, the more unstable the fuzziness of the evaluation. For example, the hyperentropy of "the house price is about the same" is 0.03 (small entropy fluctuation, stable fuzziness); the hyperentropy of "the apartment layout is acceptable" is 0.08 (large differences in entropy values ​​of evaluations published at different times, unstable fuzziness).

[0124] The preset analysis mechanism is based on rules set according to historical data. It is used to judge the reference value of fuzzy evaluation. It filters effective evaluations by thresholds of expectation, entropy, and hyperentropy. For example, the rule is set as "expectation ∈ [0.3, 0.7] (avoiding extreme values), entropy < 0.3 (low fuzziness), hyperentropy < 0.1 (fuzziness is stable)". If the condition is met, a possible reference label is assigned; otherwise, a no-reference label is assigned.

[0125] The potential reference label is a fuzzy evaluation marked as "low uncertainty and related to transaction price fluctuations". This indicates that the evaluation has potential reference significance for property valuation. For example, the fuzzy evaluation "transportation is relatively convenient" has an expected value of 0.6, entropy of 0.2, and hyperentropy of 0.04, which meets the preset rules and is assigned a potential reference label.

[0126] The "no reference" label is a vague evaluation marked as "high uncertainty or irrelevant to transaction price fluctuations". It indicates that the evaluation has no practical reference value for valuation. For example, the entropy of the vague evaluation "that's how houses are" is 0.4 (exceeding the threshold) and its correlation with housing price fluctuations is only 0.2, so it is given the "no reference" label.

[0127] The fuzzy evaluations after preliminary analysis are the fuzzy evaluations retained after identity verification and reference value screening. They must simultaneously meet the criteria of potentially false labels (credible identity) and potentially reference labels (valuable reference).

[0128] The beneficial effects of the above technical solution are: filtering out fuzzy evaluations with false associations through blockchain identity verification, quantifying the reference value of the evaluation by combining historical data with the correlation between transaction prices, and ultimately retaining fuzzy evaluations with credible identities and practical reference value. This not only eliminates invalid noise but also provides a high-quality data foundation for the accurate correction of subsequent fuzzy evaluations, effectively improving the reliability of evaluation data in online real estate appraisals.

[0129] This invention proposes an online property valuation system, wherein the valuation processing module further includes:

[0130] A primary correction unit is used to map the fuzzy evaluation after the preliminary analysis to obtain a quantized ternary fuzzy array by fuzzy element mapping, and to correct the ternary fuzzy array by combining the fuzzy membership constraints of the basic feature vector. The fuzzy elements include: object elements, feature elements and quantization elements.

[0131] The contribution determination unit is used to extract the comparison relationship between the core value description and spatial mapping results of the real evaluation and the highly correlated implicit description, and combine it with the social platform topic discussion data to obtain the value factor of the core value description, and combine it with the exponential decay weight of the evaluation release time to determine the first contribution value.

[0132] The element labeling unit is used to assign labels to the quantified fuzzy elements based on the matching relationship between the core value description and the element evaluation description of the fuzzy evaluation after preliminary analysis, and based on the corresponding first contribution value. The labels are strong correlation labels, weak correlation labels, and no correlation labels.

[0133] The secondary correction unit is used to perform secondary correction on the ternary fuzzy array based on the label results;

[0134] The correction evaluation unit is used to obtain a correction evaluation based on the results of the first correction and the second correction.

[0135] In this embodiment, fuzzy element mapping decomposes fuzzy evaluation into three core elements: "thing-feature-quantification," and converts them into a structured fuzzy mathematical expression (mapped to a ternary fuzzy array). The thing element represents the object of evaluation (a real estate-related entity), such as "wall sound insulation," "property services," and "surrounding traffic." The feature element describes the attributes or dimensions of the thing, such as "sound insulation effect," "response speed," and "congestion level." The quantification element provides a fuzzy quantitative description of the feature (without explicit numerical values), such as "good," "average," "high," and "not too bad."

[0136] The quantized ternary fuzzy array is obtained by mapping fuzzy elements and describes the uncertainty of fuzzy evaluation through three quantitative indicators: expectation (Ex), entropy (En), and hyperentropy (He). It forms an array [Ex, En, He]. For example, the ternary fuzzy array for the fuzzy evaluation "the sound insulation effect of the wall is okay" is [0.6, 0.15, 0.04], where Ex=0.6 (the average corresponds to "above average"), En=0.15 (low fuzziness) and He=0.04 (high stability).

[0137] In this embodiment, the fuzzy membership constraint of the basic feature vector is a fuzzy membership rule set based on the basic feature vector (location entropy value, spatial topological features, etc.) to constrain the quantification range of the fuzzy evaluation so that it matches the actual attributes of the property. For example, if the target property's "supporting facilities efficiency index = 0.9" (dense surrounding supermarkets), then a fuzzy membership constraint is set for the fuzzy evaluation related to "shopping convenience" (such as "easy to buy things"): Ex ≥ 0.6 (to avoid underestimation); if the "physical attenuation coefficient = 0.7" (the house is relatively old), then a constraint is set for the fuzzy evaluation of "wall quality" (such as "the wall is okay"): Ex ≤ 0.5 (to avoid overestimation). Specifically, for each dimension of the basic feature vector (such as location entropy value, spatial topological features), a fuzzy membership function (such as an S-shaped function) is trained through historical data to clarify the reasonable range of Ex, En, and He for the corresponding fuzzy evaluation. If the range is exceeded, a forced correction is made (such as setting the upper limit of Ex to the basic feature quantification value + 0.1).

[0138] In this embodiment, the first correction is the initial adjustment of the ternary fuzzy array based on the fuzzy membership constraints of the basic feature vector, to ensure that the quantitative result of the fuzzy evaluation is consistent with the actual attributes of the property. For example, the original ternary array of the fuzzy evaluation "the community facilities are okay" is [0.5, 0.2, 0.05], but the "supporting facilities efficiency index = 0.9" of the basic feature vector, according to the fuzzy membership constraints (Ex≥0.6), the array after the first correction is [0.6, 0.18, 0.05] (increasing the expectation and reducing the fuzziness).

[0139] In this embodiment, the core value description of the real evaluation is a key description that has a significant impact on the property value and is extracted from the real evaluation. It has clear attributes and characteristics. For example, the core value description of the real evaluation "There is a subway line 3 300 meters away from the entrance of the community, and there is no congestion during the morning and evening rush hours" is "high transportation convenience (close to the subway and less congestion)".

[0140] The spatial mapping result is the spatial matching result between explicit features and defect features (related to the spatial mapping process mentioned above), including horizontal / vertical / diagonal / temporal matching degree. For example, the spatial mapping result between the actual evaluation of "water seepage in the east wall of the master bedroom" and the image defect is "horizontal matching degree 0.92, vertical matching degree 0.88" (height matching).

[0141] The comparison relationship is the matching or association between the core value description and the spatial mapping result or the highly correlated implicit description. It is used to verify the authenticity of the core value description. For example, the core value description "convenient transportation" is compared with the spatial mapping result "200 meters away from the subway station (match degree 0.9)" and the highly correlated implicit description "fast commute to work" to verify the authenticity of "convenient transportation".

[0142] The value factor is a quantitative indicator that measures the degree of influence of the core value description on the property value. The higher the value, the greater the impact of the description on the valuation. For example, the core value description "convenient transportation" combined with topic discussion data (positive sentiment accounts for 80%) and spatial mapping results (match degree 0.9) calculates a value factor of 0.35 (ranked first among all factors). Specifically, the value factor is obtained by weighted summation of "match degree of comparison relationship (0.4), sentiment tendency of topic discussion (0.3), and weight of associated implicit description (0.3)" through the analytic hierarchy process (AHP), and the sum after normalization is 1.

[0143] The exponential decay weighting of evaluation publication time is assigned different weights based on the interval between the evaluation publication time and the current evaluation time (the more recent the time, the higher the weight), reflecting the timeliness of the evaluation. For example, an exponential decay formula can be set. (t is the number of days): The weight of the evaluation within 30 days is 0.97, within 90 days is 0.91, within 180 days is 0.83, and after 365 days it is 0.6 (the weight decays over time).

[0144] The first contribution value is the specific contribution of the core value description to the property value, which is determined by the value factor and the index decay weight. For example, if the value factor of the core value description "convenient transportation" is 0.35 and the release time is 30 days (weight 0.97), then the first contribution value = 0.35 × 0.97 ≈ 0.34.

[0145] The element evaluation description is a structured description of fuzzy elements based on fuzzy elements after preliminary analysis (i.e., a combination of "thing element + feature element + quantitative element"). For example, the element evaluation description of the fuzzy evaluation "the green coverage rate of the community is relatively high" after preliminary analysis is "thing = community greening, feature = coverage rate, quantification = relatively high".

[0146] Matching relationship is the degree of semantic similarity or association between core value description and element evaluation description. For example, the semantic similarity between the core value description "convenient transportation (near the subway)" and the element evaluation description "the surrounding transportation is reasonably convenient" is 0.82 (high match); and the similarity with "the community greening is okay" is 0.15 (low match).

[0147] Strongly correlated, weakly correlated, and uncorrelated labels are association strength labels assigned to quantized fuzzy elements based on matching relationships and first contribution values, and are used for subsequent correction.

[0148] Strongly correlated tags: Matching degree ≥ 0.7 and first contribution value ≥ 0.3, indicating a high correlation between the fuzzy element and the core value description. For example, the element evaluation description "transportation convenience is acceptable" has a matching degree of 0.82 and a first contribution value of 0.34 with the core value description "transportation convenience," thus being assigned a strongly correlated tag. Weakly correlated tags: 0.3 ≤ matching degree < 0.7 or 0.1 ≤ first contribution value < 0.3, indicating a certain correlation but limited strength. For example, the element evaluation description "surrounding facilities are average" has a matching degree of 0.55 and a first contribution value of 0.2 with the core value description "commercial convenience," thus being assigned a weakly correlated tag. Uncorrelated tags: Matching degree < 0.3 or first contribution value < 0.1, indicating no substantial correlation. For example, the element evaluation description "wall sound insulation is acceptable" has a matching degree of 0.1 with the core value description "transportation convenience," thus being assigned an uncorrelated tag.

[0149] Secondary correction adjusts the ternary fuzzy array after primary correction based on the label results (strong / weak / no correlation). It strengthens the influence of highly correlated elements and weakens uncorrelated elements. For example, the ternary array [0.6, 0.18, 0.05] after primary correction has strong correlation labels on the fuzzy elements. During secondary correction, the expected value is increased (Ex=0.65) and the entropy is decreased (En=0.15), resulting in the final array [0.65, 0.15, 0.04] (enhanced determinism). If the labels are uncorrelated, the expected value is decreased (Ex=0.5) and the entropy is increased (En=0.25).

[0150] The revised evaluation is the final result of integrating the first revision (basic feature constraints) and the second revision (label association constraints), transforming the fuzzy evaluation into a more deterministic quantitative description. For example, the fuzzy evaluation "the greening of the community is okay" becomes the revised evaluation "the greening coverage of the community is above average (quantitative score 0.65±0.15)" after the first revision (Ex=0.6) and the second revision (the strong association label is improved to Ex=0.65).

[0151] The beneficial effects of the above technical solution are as follows: by decomposing fuzzy elements, quantifying and mapping, and making double corrections, the semantically ambiguous evaluation is transformed into a more certain quantitative result: the first correction combines the basic characteristics of the property to ensure that the evaluation matches the actual attributes, and the second correction strengthens the relevance of the evaluation by associating core value descriptions and topic discussion data. The final corrected evaluation not only retains the effective information of the user's subjective experience, but also reduces the fuzziness through mathematical modeling, providing accurate evaluation dimensions to support the comprehensive value assessment of the property.

[0152] This invention proposes an online property valuation system, wherein the feature extraction and mining module includes:

[0153] A function construction unit is used to construct an expectation function for implicit features, and to determine the depth level of potential association mining based on the expectation function, and to mine from a preset knowledge graph based on the depth level.

[0154] The associated feature extraction unit is used to extract associated features of the mining results at each level based on the basic feature vector, and to perform descriptive transformation on the associated features at each level to obtain a highly associated implicit description.

[0155] In this embodiment, the expectation function is used as a mathematical model to measure the benefits of potential association mining. It integrates factors such as association degree, mining cost, and information gain to determine the optimal mining depth (level), balancing the value and cost of the mining results. For example, an expectation function is constructed for the implicit feature "convenience in daily life". Where R(k) is the correlation degree of the k-th level (the higher the value, the greater the benefit), C(k) is the mining cost (the higher the level, the higher the cost), and γ1 and λ1 are weight coefficients (e.g., γ1=0.6, λ1=0.4). Calculated, E(1)=0.5, E(2)=0.7, E(3)=0.4, so the optimal depth level is 2. Specifically, the function parameters are trained through historical mining data, R(k) is calculated using grey correlation degree (the degree of correlation with basic features), and C(k) is quantified as the calculation time (seconds). The k with the maximum function value is the optimal depth level (usually 1-3 levels).

[0156] Latent association mining starts with implicit features and uses a knowledge graph to reason layer by layer to uncover deep-seated associations with property attributes, ratings, and amenities. For example, it can perform latent association mining on the implicit feature "convenient for children to go to school."

[0157] First-level association: "Near school";

[0158] Two-layer association: "Near school → School is a key primary school";

[0159] Three-tiered relationship: "Near key primary schools → Stable school district boundaries → Less pressure to get into high school".

[0160] The depth level is the number of reasoning steps in potential association mining (the higher the level, the deeper the mining), determined by the expected function calculation result, to avoid over-mining (high cost, more noise) or under-mining (incomplete information). For example: Level 1: direct association (e.g., "comfortable" → "good property"); Level 2: indirect association (e.g., "comfortable" → "good property" → "24-hour security"); Level 3: cross-domain association (e.g., "comfortable" → "good property" → "24-hour security" → "good surrounding security"). For example, a level threshold (1-3 levels) can be set, and the benefit value of each level can be calculated by the expected function. The level with the highest benefit is selected as the mining depth (e.g., if level 2 has the highest benefit, then mining is done up to level 2).

[0161] The pre-built knowledge graph is a semantic network containing entities (such as houses, amenities, and evaluations) and relationships (such as "located in," "affects," and "contains") in the real estate field. It is used to support hierarchical reasoning of potential associations. For example, the knowledge graph contains entities such as "property management," "security," and "public security," and relationships such as "property management includes security," "security affects public security," and "public security affects the living experience." This provides support for reasoning such as "comfortable → good property management → strict security → good public security." Specifically, the knowledge graph is built using Neo4j, and the entities cover real estate attributes (location, apartment type, etc.), amenities (schools, hospitals, etc.), and evaluation dimensions (property management, environment, etc.). Relationships are extracted through manual annotation and machine learning (such as the TransE model) to ensure the accuracy of the reasoning logic.

[0162] In this embodiment, the mining result is a set of information associated with implicit features mined by a preset knowledge graph according to the depth level. For example, the implicit feature "convenient commuting" is mined at a depth of 2 layers, and the result is "1st layer: near the subway; 2nd layer: subway line 3 (not congested during morning and evening rush hours)".

[0163] Associated features are those extracted from the mining results that have a significant correlation with the basic feature vector (correlation degree higher than a preset threshold). This ensures that the mined information matches the actual attributes of the property. For example, the correlation degree between the mining result "near Metro Line 3" and the basic feature vector's "location entropy value 0.85" (dense commercial areas usually have well-developed subways) is 0.75 (higher than the threshold of 0.6), so it is an associated feature. If the correlation degree between the mining result "near mountainous area" and "location entropy value 0.85" is 0.3 (lower than the threshold), then it is not an associated feature. The Pearson correlation coefficient is used to calculate the correlation degree between the mining results and each dimension of the basic feature vector. A threshold is set (such as 0.6), and features with coefficients greater than or equal to the threshold are selected as associated features.

[0164] Description transformation converts associated features from structured data (such as "Metro Line 3, 500 meters away") into natural language descriptions, making them more intuitive in reflecting the correlation of property value. For example, the associated feature "Metro Line 3 (500 meters away, congestion index during morning and evening rush hours 0.3)" is transformed into the description "Near Metro Line 3, 5-minute walk away, smooth traffic during morning and evening rush hours". Specifically, natural language is generated based on a template library (such as "near {transportation facilities}, {distance description}, {traffic status}"), and the templates are filled with quantitative data of basic features (such as distance and congestion index) to ensure that the description is accurate and easy to understand.

[0165] The beneficial effects of the above technical solution are: by optimizing the logic of mining depth through expectation function, hierarchical reasoning of knowledge graph, and basic feature screening, the fuzzy implicit features are transformed into accurate and highly correlated descriptions. This avoids the incomplete information of shallow mining and prevents noise interference from deep mining, ensuring that the mining results are strongly correlated with the actual attributes of the property, and improving the depth and accuracy of online property evaluation.

[0166] This invention proposes an online property valuation system, wherein the feature extraction and mining module further includes:

[0167] The lateral mapping unit is used to laterally map the explicit features to the defect features to determine the horizontal matching degree. ,in, To display the horizontal offset between the three-dimensional coordinate points of the feature and the three-dimensional coordinate points of the defect feature; The width of the wall corresponding to the explicit feature;

[0168] The vertical mapping unit is used to vertically map the explicit features to the defect features to determine the vertical matching degree. ,in, This represents the vertical offset between the three-dimensional coordinates of the explicit feature and the three-dimensional coordinates of the defect feature. The wall height corresponding to the explicit feature;

[0169] The diagonal mapping unit is used to diagonally map the explicit features to the defect features and determine the diagonal matching degree. , where d is the absolute distance difference between the three-dimensional coordinates of the explicit feature and the three-dimensional coordinates of the defect feature to the diagonal; This represents the standard deviation of the wall dimensions. The length of the diagonal of the space in which the explicit feature is located;

[0170] The temporal mapping unit is used to perform temporal mapping between the explicit features and the defect features to determine the temporal matching degree. ,in, The maximum allowable time difference is set to 365 days. This corresponds to the time the evaluation was published; For the present time; This refers to the actual time when the defect began to appear; The vertical height of the wall is based on the current time. The vertical height of the wall based on the evaluation release time; The difference between the current time and the start time of the event corresponding to the explicit feature; For error weights; The attenuation coefficient;

[0171] The horizontal matching degree Mx, the vertical matching degree My, the diagonal matching degree Md, and the temporal matching degree Ms are used as the spatial mapping results.

[0172] In this embodiment, explicit features (text descriptions) and defect features (images) are mapped to a three-dimensional coordinate system of the property (X: horizontal direction, Y: vertical direction, Z: spatial dimension, auxiliary diagonal calculation).

[0173] Horizontal offset The difference between the horizontal coordinates of the explicit feature description (e.g., "1m to the left of the east wall") and the actual horizontal coordinates of the defect (absolute value, unit: meters).

[0174] Wall width The horizontal length of the wall containing the explicit feature (e.g., the width of the east wall is 5m, unit: meters) is extracted from the basic feature vector.

[0175] Vertical offset The difference between the vertical coordinates of the explicit feature description (e.g., "1.2m above the ground") and the actual vertical coordinates of the defect (absolute value, unit: meters).

[0176] Wall height : The vertical height of the wall containing the explicit feature (e.g., floor height 3m, unit: meters), extracted from the basic feature vector.

[0177] Absolute distance difference d: The difference in distance from the three-dimensional point of the explicit feature and the defect feature to the wall diagonal (e.g., the upper left-lower right diagonal of the east wall) (absolute value, unit: meters).

[0178] Standard deviation of wall dimensions : Statistical standard deviation of wall dimensions (width / height) for the same apartment type (unit: meters), reflecting the fluctuation of wall dimensions (e.g., σ=0.1m, indicating that the wall width fluctuates by ±0.1m).

[0179] diagonal length : The diagonal length of the space (such as a room) where the explicit feature is located (unit: meters, calculated by the Pythagorean theorem, such as if the room is 4m long and 3m wide, then dmax=5m).

[0180] Maximum allowable time difference The set time threshold (365 days) will significantly reduce the time sequence matching degree if it is exceeded, reflecting the timeliness of the evaluation.

[0181] Evaluation release time : The timestamp of the user's posting of the review (e.g., 2024-06-01 10:00:00).

[0182] Current time : The current time at the time of the assessment (e.g., 2025-07-11 15:00:00).

[0183] Time when the defect actually begins to appear The defect start time is determined through historical maintenance records and inspection reports (e.g., 2024-01-01, obtained from property management work orders or professional inspections).

[0184] Time difference The difference between the current time and the start time of the event corresponding to the explicit feature (e.g., the explicit feature "water seepage" starts recording from 2024-06-01). =395 days).

[0185] Error weights The influence coefficient of time recording error ( ,like The larger the error, the stronger the correction.

[0186] Attenuation coefficient : Rate of time decay ( ,like =0.01, the longer the time, the faster the matching degree decays.

[0187] The beneficial effects of the above technical solution are: through four-dimensional modeling of horizontal, vertical, and diagonal (three-dimensional space) + temporal (time dimension), it can achieve cross-modal accurate correlation, scene adaptation and multi-dimensional integration, thereby indirectly improving the comprehensiveness of real estate value assessment.

[0188] This invention proposes an online property valuation system, wherein the contribution determination unit includes:

[0189] The positioning subunit is used to locate attribute words and feature words in the actual evaluation and construct a structured description. ,in, The semantic vector representing the true evaluation;

[0190] Associated subunits are used to determine the structured description. Attribute correlation with the basic feature vector And retain descriptions with attribute relevance greater than the preset relevance to form a core value description set. ,in, This represents the j-th core value description; m represents the number of core value descriptions.

[0191] The graph construction subunit is used to construct a three-layer heterogeneous graph G=(V,E) based on the correspondence between the core value description, spatial mapping results, and highly correlated implicit description. The first layer consists of core value description nodes, the second layer consists of spatial mapping result nodes, and the third layer consists of highly correlated implicit description nodes. V represents the nodes involved in the three layers, and E represents the edges corresponding to the nodes.

[0192] The three-domain extraction subunit is used to extract three-domain features associated with the core value description from the topic discussion data of the social platform. The three-domain features include: the sentiment domain, the dissemination domain, and the spatiotemporal domain.

[0193] The factor determination subunit is used to obtain the value factors describing the core value based on the three-layer heterogeneous graph and the three-domain features.

[0194] In this embodiment, attribute words refer to the real estate objects described in the evaluation (such as "property", "greenery", "transportation"), and feature words refer to the specific characteristics of the attributes (such as "24-hour patrol", "80% coverage", "300 meters from subway"). Structured description: Associating attribute words with feature words to form "attribute-feature" pairs (such as (property, 24-hour patrol)). Semantic Vectors: By fine-tuning a BERT model in the real estate domain (such as BERT-Property), structured descriptions are converted into 768-dimensional vectors (capturing semantic associations, such as the implicit relationship between "property patrol" and "security"). For example, evaluating "the green coverage rate of the community reaches 80%" → locating the attribute word "green" and the feature word "coverage rate reaches 80%" → structured description (green, high coverage rate) → obtained by BERT encoding to obtain a semantic vector v=[0.12,-0.35,...,0.21] (768 dimensions). Specifically, NLP tools (such as HanLP) are used for part-of-speech tagging to extract nouns (attribute words) and adjectives / quantifiers (feature words); a pre-trained BERT model (fine-tuned based on real estate corpus) is loaded to generate semantic vectors.

[0195] In this embodiment, attribute correlation is the cosine similarity between the semantic vector of the structured description and the dimension of the basic feature vector (measuring the degree of correlation between the description and the property attribute).

[0196] Preset relevance: Set a threshold (e.g., 0.6) to filter descriptions with high relevance.

[0197] Core value description set: Highly relevant structured descriptions (such as {(greenery, high coverage), (transportation, proximity to subway)}) are retained. For example, the cosine similarity between the structured description (greenery, high coverage) and the basic feature "supporting facilities efficiency index" (related to the greenery dimension) is 0.75 (>0.6), so it is included in the set; the description (property, poor attitude) has a correlation of 0.4 (<0.6) with the "supporting facilities efficiency index", so it is removed.

[0198] Specifically, for each dimension of the basic feature vector (such as "supporting efficiency index" corresponding to attributes such as "greening" and "commercial"), construct an attribute-dimension mapping table; calculate the cosine similarity between the structured description vector and the basic feature dimension vector, and filter descriptions that are ≥ the threshold.

[0199] The core value description node is each "attribute-feature" pair in the core value description set (e.g., node N1 = (greening, high coverage)).

[0200] The spatial mapping result node represents the matching degree between explicit features and defective features (e.g., N2=(Mx=0.92, high level matching degree), associated with the parameters Mx, My, etc. of the spatial mapping mentioned above).

[0201] Highly correlated implicit description nodes are implicit descriptions that are highly correlated with basic features (such as N3 = "Convenience in life", which is correlated with the implicit feature mining results mentioned above).

[0202] Edge E: The association between nodes, with the weight being the degree of association (e.g., the edge weight between N1 and N2 is 0.8, indicating the association between the greening description and the horizontal matching degree).

[0203] For example: heterogeneous graph structures, such as Figure 2 As shown:

[0204] First layer (core value): N1 = (greenery, high coverage), N2 = (transportation, proximity to subway);

[0205] Second layer (spatial mapping): N3=(Mx=0.92), N4=(My=0.85);

[0206] The third layer (implicit description): N5 = "Convenient living", N6 = "Efficient commuting";

[0207] Edges: N1-N3 weight 0.8 (related to greening and horizontal matching), N2-N5 weight 0.7 (related to transportation and living convenience).

[0208] Specifically: nodes store description content / matching degree, and edge weights are calculated using cosine similarity or grey relational degree.

[0209] Emotional domain characteristics: the emotional tendency (positive / negative / neutral) and intensity of topic discussions (e.g., "good greening" has a positive emotional share of 85%).

[0210] Characteristics of the dissemination domain: the scale of the topic discussion (number of reads, number of reposts, such as 500,000 reads) and speed (rate of increase in popularity, such as trending on social media within 24 hours).

[0211] Spatiotemporal characteristics: the time when the topic discussion occurred (e.g., 2025-07-01) and the geographical distribution (the IP addresses of the discussion users are concentrated within 5 kilometers of the property).

[0212] For example, extract topic discussion data for (greenery, high coverage):

[0213] Emotional domain: 82% positive, emotional intensity 0.8 (after normalization);

[0214] Dissemination scope: 300,000 views, dissemination speed 0.8 (daily increase in popularity of 50%);

[0215] Spatiotemporal domain: The discussion time is concentrated in the past month, and the average distance between IP address and property is 2 kilometers (0.9 after normalization).

[0216] Specifically, this involves: calling platform APIs (such as Weibo API) to obtain topic streams; using sentiment analysis tools (such as SnowNLP) to calculate sentiment trends; using time series models (ARIMA) to analyze the spread speed; and using geofencing (Gaode API) to resolve IP addresses.

[0217] The value factor is a quantitative indicator that measures the degree to which the core value description affects the property value (integrating graph correlation and topic discussion characteristics; the higher the value, the greater the impact).

[0218] Computational logic: Graph Attention Network (GAT) learns the association weights of nodes in a heterogeneous graph (e.g., the edge weight contribution of N1 and N3 is 0.4).

[0219] Weighted fusion of three domain features (emotion × 0.3 + propagation × 0.4 + spatiotemporal × 0.3, such as 0.38 after fusion);

[0220] The final value factor = graph weight × 0.6 + topic discussion fusion value × 0.4 (e.g., 0.4 × 0.6 + 0.38 × 0.4 = 0.392). For example, the value factor of the core value description (greening, high coverage) is 0.35 (the highest proportion among all descriptions, reflecting the strong impact of greening on value). Specifically, a graph attention model is constructed (PyTorch-Geometric implements GAT), the heterogeneous graph and topic discussion features are input, the model is trained to learn the weight allocation rules of the value factor, and the normalized value factor (the sum is 1) is output.

[0221] The beneficial effects of the above technical solution are as follows: Through the entire process of structured analysis → association filtering → graph modeling → thematic discussion integration, the fragmented description of real evaluations is transformed into accurately quantified value factors. It not only relies on basic features to ensure that the description is strongly bound to the property attributes, but also mines multi-dimensional relationships through heterogeneous graphs, and incorporates the emotional, dissemination, and spatiotemporal characteristics of social thematic discussions. The final output value factors accurately depict the impact weight of real evaluations on property value, providing a key basis for subsequent comprehensive value calculation, and significantly improving the accuracy and depth of the application of evaluation data in property valuation.

[0222] This invention proposes an online property valuation system, wherein the valuation module includes:

[0223] The benchmark correction unit is used to correct the basic value benchmark line based on the spatiotemporal decay coefficient.

[0224] The factor analysis unit is used to assign initial weights to the weight evaluation factors output by the deep belief network according to the feature category, and at the same time, capture the temporal correlation between the weight evaluation factors.

[0225] The comprehensive acquisition unit is used to obtain the comprehensive value based on the corrected baseline, initial weights and time series correlation, the value contribution quantification value of the true evaluation, and the value contribution quantification value of the corrected evaluation.

[0226] In this embodiment, the spatiotemporal attenuation coefficient is a coefficient (normalized to [0,1]) that attenuates and corrects the baseline value from the time dimension (the timeliness of transaction data, the closer the data, the more reliable it is) and the spatial dimension (the distance between the transaction property and the target property, the closer the data, the more comparable the data).

[0227] Time decay: Given that the average transaction price in the target property's area was 5 million six months ago, and setting a monthly decay rate u1=0.01, then the time decay coefficient... (The reference value dropped to 94.1% after 6 months).

[0228] Spatial attenuation: If a property for sale is 5 kilometers away from the target property, and the attenuation coefficient per kilometer is set to 0.1, then the spatial attenuation coefficient is... (The reference accuracy drops to 66.7% at a distance of 5 kilometers).

[0229] Overall adjustment: Baseline value Vbase = 5 million, adjusted Vbase′ = 5 million × 0.941 × 0.667 ≈ 3.14 million.

[0230] Feature categories are used to classify weighting factors (such as physical features "apartment type, decoration", and service features "property management, security"), making it easier to assign initial weights by category and reflect the differences in importance between categories.

[0231] For example: Service characteristics: include property management (0.3) and security (if they exist, the weight is assigned to this category);

[0232] Traffic characteristics: include traffic (0.25) and commuting convenience (if present);

[0233] Educational characteristics: Includes school district (0.2);

[0234] Environmental characteristics: Includes greenery (0.15);

[0235] Physical characteristics: including apartment type (0.1) and decoration standard (if applicable).

[0236] In this embodiment, the initial weights are the basic weight percentages allocated according to the feature categories (reflecting the overall priority of the category's impact on property value). For example, the initial weights for service features are 30% (corresponding to 0.3 for property), transportation features are 25% (corresponding to 0.25 for transportation), education is 20%, environment is 15%, physics is 10%, and the total is 100%.

[0237] In this embodiment, the temporal correlation refers to the dynamic relationship of the weight factors over time (e.g., the weight of the "transportation" factor will increase after the subway opens). For example, if the initial weight of the transportation factor is 0.25, after the subway opens (time t=0), the LSTM model learns the evolution of the weight over time. (When t=1 month, the weight ≈0.295, reflecting the continuous impact of traffic improvement). Specifically, the time series data of factor weights are trained using the LSTM model (sliding window is 3 months) to capture the dynamic trend of the weights; trigger conditions are set (such as policy release, facility opening) and the model is called in real time to update the time series correlation (such as when the subway opens, the LSTM is forced to be retrained).

[0238] The modified baseline Vbase′ is the basic value baseline after time and space decay correction (such as 3.14 million in the example), representing the benchmark reference of the regional market.

[0239] The quantified value of the real evaluation, Vreal, is the increase or decrease in property value after the real evaluation is quantified by value factors (e.g., "24-hour property patrol" corresponds to a value factor of 0.3, contributing +200,000). For example, the conversion between value factors and amount: through the historical transaction regression model, it is fitted that "each 0.1 weight of the property factor corresponds to a value of +50,000", then a property factor of 0.3 contributes 0.3 / 0.1×5=+150,000.

[0240] Multiple factors combined: Traffic factor contributes 0.25 + 100,000, total Vreal = 15 + 10 = +250,000.

[0241] The value contribution quantification value Vcorr of the revised evaluation is the increase or decrease in property value after the fuzzy evaluation is revised (e.g., greening generally corresponds to a value factor of -0.1, and a contribution of -50,000). For example, a revised evaluation of greening coverage of 60% (medium) corresponds to a value factor of -0.1, and the contribution obtained through the regression model is -0.1 / 0.1×5=-50,000 (negative impact).

[0242] The comprehensive value calculation is a final valuation that integrates the modified baseline (regional baseline), the evaluation contribution (user experience), and combines the initial weight (category priority) and the temporal correlation (dynamic changes).

[0243] Baseline percentage: If the adjusted baseline percentage is set at 60%, then 314 × 0.6 = 1,884,000;

[0244] Evaluation contribution percentage: set at 40%, then (25−5)×0.4=80,000;

[0245] Time series adjustment: Due to the opening of the subway, the weight of the transportation factor is increased by an additional 5%, contributing 500 × 0.05 × 0.4 = 100,000 (simplified calculation, actual calculation needs to be adjusted according to the baseline).

[0246] Total value: 1,884,000 + 80,000 + 100,000 = 2,064,000 (This example is for simplified logic; a more complex dynamic weighted model is required in practice).

[0247] Construct a dynamic weighted model Vtotal=wbase Vbase'+weval (Vreal+Vcorr), where wbase+weval=1, and weval dynamically increases with the amount of evaluation data (e.g., when there are ≥100 evaluations, weval increases from 40% to 50%).

[0248] Introducing a time-series correction term: Adjusting the factor contribution (e.g., Vreal′=Vreal) by changing the weights Δw(t) of the LSTM output. (1+Δw(t))).

[0249] The beneficial effects of the above technical solution are: by using a spatiotemporal attenuation calibration benchmark, hierarchical weighted characterization factor, time-series correlation to capture dynamics, and multi-source fusion to output value, it breaks through the limitations of traditional static benchmark + single evaluation and improves the accuracy of valuation in response to market changes and real estate characteristics.

[0250] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An online property valuation system, characterized in that, include: The online data acquisition module is used to acquire the basic feature vector and multimodal evaluation data of the target property online through a distributed crawler cluster and open API interface. The multimodal evaluation data includes: text evaluation, image evaluation and social platform topic discussion data. The feature extraction and mining module is used to extract explicit and implicit features from the text evaluation, extract defect features from the image evaluation, perform spatial mapping between the explicit features and defect features, mine potential associations of the implicit features, and obtain a highly correlated implicit description by combining the basic feature vector. The evaluation processing module is used to identify real, false, and fuzzy evaluations of the text and image evaluations based on a pre-built evaluation classification model, eliminate the false evaluations, perform preliminary analysis on the fuzzy evaluations, and correct the fuzzy evaluations after preliminary analysis based on spatial mapping results, highly correlated implicit descriptions, and social platform topic discussion data. The value assessment module is used to acquire gridded transaction data of the same area as the target property to obtain a basic value baseline. Simultaneously, it acquires weighted evaluation factors for different value types based on a deep belief network, and obtains a comprehensive value based on the basic value baseline, the actual evaluation, and the corrected evaluation, outputting a value assessment report. The feature extraction and mining module further includes: The lateral mapping unit is used to laterally map the explicit features to the defect features to determine the horizontal matching degree. ,in, To display the horizontal offset between the three-dimensional coordinate points of the feature and the three-dimensional coordinate points of the defect feature; The width of the wall corresponding to the explicit feature; The vertical mapping unit is used to vertically map the explicit features to the defect features to determine the vertical matching degree. ,in, This represents the vertical offset between the three-dimensional coordinates of the explicit feature and the three-dimensional coordinates of the defect feature. The wall height corresponding to the explicit feature; The diagonal mapping unit is used to diagonally map the explicit features to the defect features and determine the diagonal matching degree. , where d is the absolute distance difference between the three-dimensional coordinates of the explicit feature and the three-dimensional coordinates of the defect feature to the diagonal; This represents the standard deviation of the wall dimensions. The length of the diagonal of the space in which the explicit feature is located; The temporal mapping unit is used to perform temporal mapping between the explicit features and the defect features to determine the temporal matching degree. ,in, The maximum allowable time difference is set at 365 days. This corresponds to the time the evaluation was published; For the present time; This refers to the actual time when the defect began to appear. The vertical height of the wall is based on the current time. The vertical height of the wall based on the evaluation release time; The difference between the current time and the start time of the event corresponding to the explicit feature; For error weights; The attenuation coefficient; The horizontal matching degree Mx, the vertical matching degree My, the diagonal matching degree Md, and the temporal matching degree Ms are used as the spatial mapping results.

2. The online real estate valuation system according to claim 1, characterized in that, The basic feature vector includes: location entropy value, spatial topological features, physical attenuation coefficient, and corresponding efficiency index.

3. The online real estate valuation system according to claim 1, characterized in that, The evaluation processing module includes: The cross-validation unit is used to cross-validate the identified fuzzy evaluations with the user identity information stored on the blockchain. If the user has not completed real-name authentication or the authentication information has no geographical association with the property, a false and irrelevant label is assigned to the fuzzy evaluation; otherwise, a potentially false label is assigned to the fuzzy evaluation. The unit may be used to quantify the uncertainty of the fuzzy evaluation based on the correlation between historical fuzzy evaluation and actual transaction price fluctuations, convert it into expectation, entropy and hyperentropy, and analyze the expectation, entropy and hyperentropy based on a preset analysis mechanism, and assign possible reference labels or no reference labels. The retention unit is used to retain the fuzzy evaluation only when the fuzzy evaluation is simultaneously assigned a possible false label and a possible reference label; Otherwise, the fuzzy evaluations are discarded, and the fuzzy evaluations after preliminary analysis are the retained fuzzy evaluations.

4. The online real estate valuation system according to claim 3, characterized in that, The evaluation processing module further includes: A primary correction unit is used to map the fuzzy evaluation after the preliminary analysis to obtain a quantized ternary fuzzy array by fuzzy element mapping, and to correct the ternary fuzzy array by combining the fuzzy membership constraints of the basic feature vector. The fuzzy elements include: object elements, feature elements and quantization elements. The contribution determination unit is used to extract the comparison relationship between the core value description and spatial mapping results of the real evaluation and the highly correlated implicit description, and combine it with the social platform topic discussion data to obtain the value factor of the core value description, and combine it with the exponential decay weight of the evaluation release time to determine the first contribution value. The element labeling unit is used to assign labels to the quantified fuzzy elements based on the matching relationship between the core value description and the element evaluation description of the fuzzy evaluation after preliminary analysis, and based on the corresponding first contribution value. The labels are strong correlation labels, weak correlation labels, and no correlation labels. The secondary correction unit is used to perform secondary correction on the ternary fuzzy array based on the label results; The correction evaluation unit is used to obtain a correction evaluation based on the results of the first correction and the second correction.

5. The online real estate valuation system according to claim 1, characterized in that, The feature extraction and mining module includes: A function construction unit is used to construct an expectation function for implicit features, and to determine the depth level of potential association mining based on the expectation function, and to mine from a preset knowledge graph based on the depth level. The associated feature extraction unit is used to extract associated features of the mining results at each level based on the basic feature vector, and to perform descriptive transformation on the associated features at each level to obtain a highly associated implicit description.

6. The online property valuation system according to claim 4, characterized in that, The contribution determination unit includes: The positioning subunit is used to locate attribute words and feature words in the actual evaluation and construct a structured description. ,in, The semantic vector representing the true evaluation; Associated subunits are used to determine the structured description. Attribute correlation with the basic feature vector And retain descriptions with attribute relevance greater than the preset relevance to form a core value description set. ,in, This represents the j-th core value description; m represents the number of core value descriptions. The graph construction subunit is used to construct a three-layer heterogeneous graph G=(V,E) based on the correspondence between the core value description, spatial mapping results, and highly correlated implicit description. The first layer consists of core value description nodes, the second layer consists of spatial mapping result nodes, and the third layer consists of highly correlated implicit description nodes. V represents the nodes involved in the three layers, and E represents the edges corresponding to the nodes. The three-domain extraction subunit is used to extract three-domain features associated with the core value description from the topic discussion data of the social platform. The three-domain features include: the sentiment domain, the dissemination domain, and the spatiotemporal domain. The factor determination subunit is used to obtain the value factors describing the core value based on the three-layer heterogeneous graph and the three-domain features.

7. The online real estate valuation system according to claim 1, characterized in that, The value assessment module includes: The benchmark correction unit is used to correct the basic value benchmark line based on the spatiotemporal decay coefficient. The factor analysis unit is used to assign initial weights to the weight evaluation factors output by the deep belief network according to the feature category, and at the same time, capture the temporal correlation between the weight evaluation factors. The comprehensive acquisition unit is used to obtain the comprehensive value based on the corrected baseline, initial weights and time series correlation, the value contribution quantification value of the true evaluation, and the value contribution quantification value of the corrected evaluation.

Images