Information acquisition method and device for predicting transaction price

Abstract

The present disclosure relates to a method and device for obtaining explanatory information of a predicted transaction price, which introduces reference real estate meeting a preset similarity condition with a target real estate, takes the real estate data of the reference real estate as a comparison basis, quantifies the feature contribution value of the target real estate in each feature dimension relative to the reference real estate, and generates explanatory information of the predicted transaction price with business interpretability and a comparison basis. In this way, the reason for forming the predicted transaction price is reflected, enabling real estate agents to intuitively understand and convey the rationality of the predicted transaction price to customers, thereby improving the credibility of the predicted transaction price.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to a method and apparatus for obtaining explanatory information for predicting transaction prices. Background Technology

[0002] In real estate transactions, real estate agents typically use property valuation models to obtain predicted transaction prices for target properties to support communication and negotiation with clients. However, these technologies often only output a single price value, lacking interpretability to explain how the predicted transaction price was determined. Therefore, there is an urgent need for a method that can generate explanatory information to explain the predicted transaction price, meeting the needs of real estate agents in their actual business operations. Summary of the Invention

[0003] This disclosure provides a method and apparatus for obtaining explanatory information for predicting transaction prices.

[0004] One aspect of this disclosure provides a method for obtaining explanatory information for predicted transaction prices, the method comprising: Acquire the housing data of the target property and the housing data of the reference property. The housing data indicates the characteristics of the corresponding property in multiple feature dimensions. The housing data of the target property and the housing data of the reference property meet a preset similarity condition. Based on the housing data of the reference housing, the housing data of the target housing is processed to obtain the feature contribution value of the target housing in each feature dimension; wherein, the feature contribution value indicates the impact of the target housing on the predicted transaction price of the target housing in each feature dimension compared with the reference housing. Based on the feature contribution values ​​corresponding to the multiple feature dimensions, explanatory information is generated to explain the predicted transaction price.

[0005] Another aspect of this disclosure provides an apparatus for obtaining explanatory information for predicted transaction prices, comprising: The acquisition module is used to acquire the housing data of the target housing and the housing data of the reference housing. The housing data indicates the characteristics of the corresponding housing in multiple feature dimensions. The housing data of the target housing and the housing data of the reference housing meet the preset similarity conditions. The processing module is used to process the housing data of the target housing based on the housing data of the reference housing, and obtain the feature contribution value of the target housing in each feature dimension; wherein, the feature contribution value indicates the impact of the target housing on the predicted transaction price of the target housing in the feature dimension compared with the reference housing; The generation module is used to generate explanatory information to explain the predicted transaction price based on the feature contribution values ​​corresponding to the multiple feature dimensions.

[0006] In some embodiments, the plurality of feature dimensions include a plurality of first feature dimensions and second feature dimensions. The first feature dimension is used to represent a single attribute of the corresponding property, and the second feature dimension is used to represent a composite attribute of the corresponding property. The generation module is used to allocate the feature contribution value corresponding to the second feature dimension to each third feature dimension based on the weights corresponding to the plurality of third feature dimensions associated with the second feature dimension, thereby obtaining an allocation value corresponding to each third feature dimension. The third feature dimension is the first feature dimension corresponding to each single attribute in the composite attribute represented by the second feature dimension. The sum of the feature contribution value corresponding to the third feature dimension and the corresponding allocation value is determined as the updated feature contribution value of the third feature dimension. The explanatory information is constructed based on the updated feature contribution value of the third feature dimension and the feature contribution value corresponding to the fourth feature dimension. The fourth feature dimension is a feature dimension other than the third feature dimension among the plurality of first feature dimensions.

[0007] In some embodiments, the reference properties include multiple properties; the processing module is further configured to, for each reference property, process the property data of the reference property based on the property data of the target property and the property data of the other reference properties to obtain the feature contribution value of the reference property on each feature dimension; for each third feature dimension, determine the average value of the feature contribution values ​​of the multiple reference properties on the third feature dimension; determine the sum of the average values ​​corresponding to the multiple third feature dimensions respectively; and determine the ratio of the average value corresponding to the third feature dimension to the sum as the weight corresponding to the third feature dimension.

[0008] In some embodiments, the processing module is configured to process the housing data of the target housing based on the housing data of the reference housing using an interpretation model, to obtain the feature contribution value of the target housing in each feature dimension; the predicted transaction price is obtained through the interpretation model.

[0009] In some embodiments, the processing module is further configured to process the property data of the target property through the interpretation model to obtain the predicted transaction price of the target property.

[0010] In some embodiments, the interpretation model is a decision tree model, which includes multiple decision trees, each decision tree including a judgment node corresponding to at least one of the multiple feature dimensions; the processing module is used to process the housing data of the target housing through each decision tree in the interpretation model to obtain the output value corresponding to each decision tree; and to determine the predicted transaction price of the target housing by summing the price benchmark value corresponding to the interpretation model and the output values ​​corresponding to the multiple decision trees respectively.

[0011] In some embodiments, the processing module is configured to process the housing data of the target housing based on the housing data of the reference housing through the interpretation model, to obtain the predicted transaction price of the target housing and the feature contribution value of the target housing in each feature dimension.

[0012] In some embodiments, the processing module is configured to process the housing data of the target housing based on the housing data of the reference housing through the interpretation model to obtain a price benchmark value and the feature contribution value of the target housing in each feature dimension; and to determine the sum of the price benchmark value and the feature contribution value of the target housing in each feature dimension as the predicted transaction price of the target housing.

[0013] In some embodiments, both the target property and the reference property are unsold properties; the device further includes: The acquisition module is further configured to acquire the housing data of sample properties and the predicted transaction price of the sample properties. The sample properties are properties that have not been sold. The predicted transaction price of the sample properties is the product of the listed price and the transaction discount rate. The transaction discount rate is predicted by the prediction model based on the housing data of the sample properties. The prediction model is trained based on the housing data of sold properties and the transaction discount rate of sold properties. The transaction discount rate is the ratio of the transaction price to the listed price of the sold properties. The training module is used to train the explanatory model based on the housing data of the sample housing and the predicted transaction price of the sample housing.

[0014] In some embodiments, the preset similarity condition includes at least one of the following: The similarity between the housing data of the target property and the housing data of the reference property is greater than a first threshold. The similarity between the features of the target property and the features of the reference property in the target feature dimension is greater than the second threshold.

[0015] Another aspect of this disclosure provides an electronic device, which includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the above-described method for obtaining explanatory information for predicting transaction prices.

[0016] Another aspect of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described method for obtaining explanatory information regarding predicted transaction prices.

[0017] Another aspect of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for obtaining explanatory information for predicting transaction prices. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0019] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a method for obtaining explanatory information for predicting transaction prices, provided in an embodiment of this disclosure. Figure 2 This is a flowchart illustrating a method for obtaining explanatory information for predicting transaction prices, provided in an embodiment of this disclosure. Figure 3 This is a schematic diagram of a decision tree provided in an embodiment of this disclosure; Figure 4 This is a flowchart illustrating a method for obtaining explanatory information for predicting transaction prices, provided in an embodiment of this disclosure. Figure 5 This is a schematic diagram of the structure of a device for obtaining explanatory information for predicting transaction prices, provided in an embodiment of this disclosure. Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of this disclosure. Detailed Implementation

[0021] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0022] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.

[0023] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below. The concepts of "first", "second", etc., used in this disclosure are used only to distinguish different devices, modules, or units and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies. The modifications of "a" and "a plurality" used in this disclosure are illustrative and not restrictive, and those skilled in the art should understand that unless explicitly indicated in the context, they should be understood as "one or more".

[0024] In real estate transactions, especially in the secondary market, real estate agents typically use property valuation models to obtain predicted transaction prices for target properties. This serves as a crucial basis for quoting prices to owners, negotiating with clients, and selecting suitable properties. Property valuation models in related technologies utilize historical transaction data to train high-precision price prediction models and estimate the prices of currently available properties. However, simply predicting transaction prices is far from sufficient. Real estate agents urgently need clear, credible, and quantifiable explanatory evidence when communicating pricing with owners, explaining property value to clients, or negotiating for bargaining power. Therefore, to address these issues, this disclosure provides a method for obtaining explanatory information regarding predicted transaction prices, which will be described below with reference to specific embodiments. Figure 1 This is a flowchart illustrating a method for obtaining explanatory information for predicting transaction prices, provided in an embodiment of this disclosure. This method can be executed by an electronic device, which can be a terminal or a server. The terminal can be exemplarily understood as a device such as a mobile phone, tablet computer, laptop computer, desktop computer, or smart TV. Figure 1 As shown, the method provided in this embodiment includes the following steps: S101. Obtain the housing data of the target property and the housing data of the reference property. The housing data indicates the characteristics of the corresponding property in multiple feature dimensions. The housing data of the target property and the housing data of the reference property meet the preset similarity conditions.

[0025] In this embodiment of the disclosure, a predicted transaction price is generated for target properties that have not yet been sold, so that real estate agents can use it in actual business. In order to make this predicted transaction price interpretable, reference properties that meet the preset similarity conditions of the target property are identified. Then, the property data of the target property and the property data of the reference property are compared and analyzed to analyze the reasons for the formation of the predicted transaction price of the target property compared with the reference property, so as to explain the predicted transaction price. This makes it easier for real estate agents and customers to understand the predicted transaction price of the target property, thereby providing a basis for real estate agents' business and improving the property transaction rate.

[0026] In this context, the target property is any unsold property, meaning any property currently for sale. Reference properties can be either unsold or sold. Compared to the target property, reference properties are those that share certain similarities with it; they can be considered similar properties to the target property. Reference properties can include one or more. The preset similarity condition is a pre-defined condition used to identify similar properties for any given property. Property data describes the corresponding property and can be described from multiple feature dimensions. These dimensions may include single attributes such as area, building age, floor, elevator availability, and community name, or composite attributes such as "floor × elevator availability" or "number of rooms × area." The features of a property in a feature dimension represent its specific attributes in that dimension. For example, if the feature dimension corresponds to area, then the property's area feature is 120 square meters, which is the actual area of ​​the property; if the feature dimension corresponds to floor, then the property's floor feature is 12 floors, which is the actual floor of the property. For the property data of the target property and the reference property, the feature dimensions corresponding to both are the same. For example, both the target property data and the reference property data include the features of the corresponding property in feature dimensions 1 to 10.

[0027] In this embodiment of the disclosure, the reference properties are determined in advance based on preset similarity conditions. That is, when obtaining descriptive information for a target property, the database already stores the corresponding reference properties, so the property data of the target property and the reference properties can be obtained from the database. Alternatively, when obtaining descriptive information for a target property, the property data of the target property is obtained from the database, and using the preset similarity conditions, the property data of reference properties that meet the preset similarity conditions are determined from the database.

[0028] S102. Based on the housing data of the reference housing, process the housing data of the target housing to obtain the feature contribution value of the target housing in each feature dimension; wherein, the feature contribution value indicates the impact of the target housing on the predicted transaction price of the target housing in the feature dimension compared with the reference housing.

[0029] In this embodiment of the disclosure, in order to interpret the predicted transaction price of the target property, the property data of the reference property is used as a reference to analyze the property data of the target property, so as to analyze the impact of each feature dimension of the target property on the predicted transaction price, and then obtain the feature contribution value of the target property in each feature dimension, so as to reflect where the target property is more expensive or cheaper than the reference property, so as to fully explain the predicted transaction price.

[0030] In this context, the feature contribution value of a target property on a certain feature dimension is equivalent to the contribution of the target property's feature on that feature dimension to the predicted transaction price of the target property. In other words, it represents the marginal impact on the predicted transaction price, or the specific amount by which the target property's feature on that feature dimension increases or decreases the predicted transaction price relative to the benchmark value. For example, if the feature contribution value of a target property on the area dimension is 80,000 yuan, on the decoration dimension is 30,000 yuan, and on the elevator presence dimension is -5,000 yuan, this reflects that the target property's larger area and better decoration have a positive impact on the predicted transaction price, increasing it by 80,000 yuan and 30,000 yuan respectively; while the lack of an elevator has a negative impact, decreasing the predicted transaction price by 5,000 yuan.

[0031] In this embodiment, using the housing data of reference properties as a background distribution or comparison benchmark, attribution analysis is performed on the housing data of the target property to obtain the feature contribution value of the target property in each feature dimension. This feature contribution value does not reflect the isolated influence of the target property's own attributes, but rather quantifies the incremental or subtractive impact of the target property on the predicted transaction price in each feature dimension relative to the reference property. For example, if the area of ​​the target property is larger than the average level of the reference properties, its feature contribution value in the "area" dimension is positive, indicating that the difference makes its predicted price higher than the reference benchmark; conversely, it is negative.

[0032] S103. Based on the feature contribution values ​​corresponding to multiple feature dimensions, generate explanatory information to explain the predicted transaction price.

[0033] In this embodiment of the disclosure, after obtaining the feature contribution values ​​corresponding to multiple feature dimensions, these feature contribution values ​​are used to generate explanatory information for explaining the predicted transaction price, so as to more fully predict the reasons for the formation of the transaction price and facilitate user viewing.

[0034] The explanatory information reflects the specific amount by which the predicted transaction price of the target property increases or decreases relative to the benchmark value due to the characteristics of the target property in various feature dimensions compared to similar properties. The explanatory information can be represented in any form, such as text or images.

[0035] The solution provided in this disclosure introduces reference properties that meet preset similarity criteria to the target property. Using the reference property's data as a comparison basis, the feature contribution value of the target property relative to the reference property in various feature dimensions is quantified. This generates explanatory information on the predicted transaction price that is business-interpretable and provides a basis for comparison. This not only reflects the reasons for the formation of the predicted transaction price, enabling real estate agents to intuitively understand and convey the rationality of the predicted transaction price to clients, but also enhances the credibility of the predicted transaction price.

[0036] In the above Figure 1 Based on the embodiments shown, in this disclosure embodiment, there are feature dimensions corresponding to composite attributes among multiple feature dimensions. After obtaining the feature contribution values ​​corresponding to each feature dimension through the interpretation model, the feature contribution values ​​of the feature dimensions corresponding to the composite attributes are then allocated so that the explanatory information only indicates the feature contribution value of the feature dimension corresponding to a single attribute. For details of the process, please refer to the following embodiments. Figure 2 The diagram illustrates a flowchart of a method for obtaining explanatory information for predicting transaction prices, as provided in an embodiment of this disclosure. Figure 2 As shown, the method for obtaining explanatory information regarding predicted transaction prices may include the following steps: S201. Obtain the housing data of the target property and the housing data of the reference property. The housing data indicates the characteristics of the corresponding property in multiple feature dimensions. The housing data of the target property and the housing data of the reference property meet the preset similarity conditions.

[0037] In this embodiment, both the target property and the reference property are unsold properties. The multiple feature dimensions include multiple first feature dimensions and second feature dimensions. The first feature dimensions represent a single attribute of the corresponding property, and the second feature dimensions represent a composite attribute of the corresponding property. A single attribute refers to factors such as area, building age, floor, presence or absence of an elevator, and community name. Composite attributes, also known as cross-attributes, refer to factors such as "floor × presence or absence of an elevator" or "number of rooms × area." In addition to multiple first feature dimensions, the multiple feature dimensions also include one or more second feature dimensions. Different first feature dimensions correspond to different attributes, and different second feature dimensions correspond to different composite attributes. Furthermore, any composite attribute corresponding to a second feature dimension is composed of at least two single attributes from the multiple first feature dimensions. Therefore, for any composite attribute, the second feature dimension corresponding to that composite attribute is associated with the first feature dimensions corresponding to each of the single attributes included in that composite attribute.

[0038] In some embodiments, the preset similarity criteria include at least one of the following.

[0039] The first condition is that the similarity between the listing data of the target property and the listing data of the reference property is greater than the first threshold.

[0040] In this embodiment, the first threshold can be any value, such as 0.8 or 0.9. The similarity between the target property's listing data and the reference property's listing data can be obtained in any way, such as using cosine similarity or Euclidean distance. By pre-setting similarity conditions, the overall similarity between the target property and the reference property across all feature dimensions is limited to be higher than the first threshold. This ensures that the selected reference properties are highly comparable to the target property in terms of property attributes and market positioning. This excludes properties with excessively large attribute differences that lack reference value, ensuring that the selected reference properties truly reflect the local market environment and pricing logic of the target property. The resulting explanatory information is more targeted and credible, thereby improving the accuracy and business practicality of the overall interpretable analysis.

[0041] The second criterion is that the similarity between the features of the target property and the features of the reference property in the target feature dimension is greater than the second threshold.

[0042] In this embodiment, the target feature dimension is a key feature dimension among multiple feature dimensions, such as area, floor, unit type, or community. Based on the preset similarity condition shown in the second item above, 30 reference properties can be selected. The second threshold can be any value, such as 0.8 or 0.9. The similarity between the features of the target property and the reference property in the target feature dimension can be obtained in any way, such as using cosine similarity or Euclidean distance. By setting a preset similarity condition, requiring the target property and the reference property to have a similarity higher than the second threshold in a specific target feature dimension, it ensures a high degree of consistency in key attributes. This avoids interpretation bias caused by excessive differences in key features, making the reference property closer to the true comparable object of the target property. Feature contribution analysis or price attribution based on this is more targeted and persuasive, ensuring the accuracy, rationality, and user comprehensibility of subsequent explanations.

[0043] S202. By interpreting the model, based on the housing data of the reference housing, the housing data of the target housing is processed to obtain the feature contribution value of the target housing in each feature dimension; wherein, the feature contribution value indicates the impact of the target housing on the predicted transaction price of the target housing in the feature dimension compared with the reference housing.

[0044] The explanatory model can be any type of machine learning model. For example, it can be a decision tree model, such as one built using a gradient boosting framework like CatBoost (Categorical Boosting); or it can be a neural network model. The explanatory model has the ability to quantify the contribution of each feature dimension to the predicted transaction price through a unified attribution mechanism. This explanatory model can also obtain the predicted transaction price for a specific property; that is, the predicted transaction price of the target property is obtained through the explanatory model.

[0045] In this embodiment of the disclosure, by using an interpretive model, the housing data of a reference property is used as a background distribution or comparison benchmark to perform attribution analysis on the housing data of the target property, thereby obtaining the feature contribution value of the target property in each feature dimension. Through the interpretive model, a refined attribution analysis of the predicted transaction price of the target property is achieved, improving the business relevance and user understandability of the model interpretation.

[0046] For example, for a target property to be explained, a set of reference properties is obtained, which includes at least one reference property. The property data of the target property and the property data of each reference property in the reference set are input into the explanation model. The explanation model uses the property data of each reference property in the reference set as the background dataset to calculate the feature contribution value of the target property relative to the background dataset. This operation is equivalent to answering: "Compared to similar properties, where is the target property more expensive or cheaper?" The marginal contribution (in yuan) of each feature dimension to the predicted transaction price is output, such as: "Area: +80,000 yuan, Decoration level: +30,000 yuan, 'Floor × Elevator presence': +10,000 yuan, Building type (slab building): +5,000 yuan". If multiple feature dimensions include a first feature dimension and a second feature dimension, the explanation model outputs the feature contribution value corresponding to the first feature dimension for a single attribute and the feature contribution value corresponding to the second feature dimension for a composite attribute.

[0047] In some embodiments, the interpretation model is a decision tree model, which includes multiple decision trees, each of which includes a judgment node corresponding to at least one of the multiple feature dimensions; then step S202 includes: processing the target property's property data based on the property data of the reference property through each decision tree in the interpretation model to obtain the feature contribution value of the target property on the feature dimension associated with the decision tree; and determining the sum of the feature contribution values ​​corresponding to the same feature dimension in different decision trees as the final feature contribution value of the feature dimension.

[0048] In this embodiment of the disclosure, different decision trees may include judgment nodes corresponding to the same feature dimension. For multiple decision trees that contain judgment nodes corresponding to the same feature dimension, a local feature contribution value corresponding to the feature dimension can be obtained in each decision tree in the manner described above. The sum of the local feature contribution values ​​calculated for the feature dimension in multiple decision trees is determined as the feature contribution value of the feature dimension. The feature contribution value of each feature dimension can be calculated in the manner described above, which is the feature contribution value output in step S202.

[0049] For example, if feature dimension A has a local feature contribution value of 100 in the first tree, 10 in the second tree, and 2 in the third tree, then the feature contribution value of feature dimension A is 100 + 10 + 2 = 112.

[0050] In some embodiments, the process of determining the feature contribution value corresponding to the feature dimensions contained in a decision tree through a decision tree includes: using any decision tree of the decision tree model, based on the housing data of multiple reference housings, determining the proportion of multiple reference housings entering different branches at each decision node; based on the determined proportions and the decision tree, determining the feature contribution value corresponding to the features of the target housing on multiple feature dimension sets; and based on the feature contribution value corresponding to the target housing on multiple feature dimension sets and the weight corresponding to each feature dimension set, determining the feature contribution value corresponding to each feature dimension of the target housing in the decision tree.

[0051] Among them, the multiple feature dimension set is all possible feature dimension sets involved in the decision tree. For example, if the nodes in the decision tree involve feature dimension 1 and feature dimension 2, then the multiple feature dimension sets are empty, feature dimension set containing feature dimension 1, feature dimension set containing feature dimension 2, and feature set containing both feature dimension 1 and feature dimension 2.

[0052] For example, the target property's data: area = 120, floor = 12, elevator = yes; any decision tree in the decision tree model, such as... Figure 3 As shown, the decision tree has three levels: Level 1 root node: if (area <= 100); Level 2 left subtree (yes, proportion 60%): y=400; Level 2 right subtree (no, proportion 40%): if (floor <= 10); Level 3 left subtree (yes, proportion 30%): y=500; Level 3 right subtree (no, proportion 70%): if (is there an elevator); Leaf left node (yes, proportion 50%): y=800; Leaf right node (yes, proportion 50%): y=600. The proportions mentioned above represent the percentage of multiple reference properties falling into the corresponding node. For example, Level 2 left subtree (yes, proportion 60%) means that 60% of the multiple reference properties enter the Level 2 left subtree through this decision tree. Furthermore, among the multiple reference properties with more than 10 floors, 30% of the properties enter the second-level left subtree from the root node, and 70% enter the second-level right subtree. Among the multiple reference properties with stairs, 50% of the properties enter the second-level left subtree from the root node, and 50% enter the second-level right subtree. At the second-level right subtree node, 40% of the reference properties with elevators enter the third-level left subtree, and 60% enter the third-level right subtree. Therefore, the nodes in this decision tree involve area, floor, and elevator dimensions. The features of the target property across these multiple feature dimensions are: empty, area=120, floor=12, elevator=yes, area=120 and floor=12, area=120 and elevator=yes, floor=12 and elevator=yes, and area=120 and floor=12 and elevator=yes. Based on these proportions, the feature contribution values ​​corresponding to the features of the target property across these multiple feature dimensions are calculated as follows: f(empty) = 0.6 400 + 0.4 (0.3) 500 + 0.7 (0.5) 800+0.5 600)) = 496, which is equivalent to the price benchmark value; f (area = 120) = 0.3 500 + 0.7 (0.5) 800+0.5 600) = 640; f (floor = 12) = 0.3 400 + 0.7 (0.5) 800+0.5 600) = 610; that is, when only the floor = 12 is known, the second level left subtree is entered at the root node with a 30% probability, and the second level right subtree is entered with a 70% probability. f(elevator = yes) = 0.5 400 + 0.5 (0.4) 500 + 0.6 800) = 540, the calculation method is the same as that for the root node f (floor = 12); f(area=120, floors=12)=0.5 800+0.5 600 = 700; f(area = 120, elevator = yes) = 0.4 500 + 0.6 800 = 680; f(floor=12, elevator=yes)=0.6 400 + 0.4 800 = 560; f(Area=120, Floors=12, Elevator=Yes)=800.

[0053] Weight calculation for different feature dimension sets: For the target feature dimension whose feature contribution value needs to be calculated, a feature subset S that does not contain the target feature dimension is determined from multiple feature dimension sets. |S| is the number of feature dimensions in feature subset S, and |D| is the number of all feature dimensions involved in the housing data. The formula for calculating the weight is: w(|S|) = (|S|!(|D|-|S|-1)!) / |D|!. The weight when |S|=0 is 0!2! / 3!=2 / 6=1 / 3, the weight when |S|=1 is 1!1! / 3!=1 / 6, and the weight when |S|=2 is 2!0! / 3!=2 / 6=1 / 3.

[0054] Calculate the feature contribution value of the area dimension: g(area)1 = f(area = 120) - f(empty) = 640 - 496 = 144, weight = 1 / 3; g(area)² = f(area = 120, floor = 12) - f(floor = 12) = 700 - 610 = 90, weight = 1 / 6; g(area)3 = f(area = 120, elevator = yes) - f(elevator = yes) = 680 - 540 = 140, weight = 1 / 6; g(area)4 = f(area=120, floors=12, elevator=yes) - f(floors=12, elevator=yes) = 800 - 560 = 240, weight = 1 / 3; The feature contribution value of area in the decision tree is 44 / 3 + 90 / 6 + 140 / 6 + 240 / 3 = 166.33.

[0055] Calculate the feature contribution value of the floor dimension: g(floor)1 = f(floor=12) - f(empty) = 610 - 496 = 114; g(floor)2 = f(area=120, floor=12) - f(area=120) = 700 - 640 = 60; g(floor)3 = f(floor=12, elevator=yes) - f(elevator=yes) = 560 - 540 = 20; g(floor)4 = f(area=120, floor=12, elevator=yes) - f(area=120, elevator=yes) = 800 - 680 = 120; The feature contribution value of each floor in the decision tree is 114 / 3 + 60 / 6 + 20 / 6 + 120 / 3 = 91.33.

[0056] Calculate the feature contribution value of the elevator dimension: g(elevator)1 = f(elevator = yes) - f(empty) = 540 - 496 = 44; g(elevator)² = f(area = 120, elevator = is) - f(area = 120) = 680 - 640 = 40; g(elevator)3 = f(floor=12, elevator=is) - f(floor=12) = 560 - 610 = -50; g(elevator)4 = f(area=120, floors=12, elevator=is) - f(area=120, floors=12) = 800 - 700 = 100; The feature contribution value of the elevator in this decision tree is 44 / 3 + 40 / 6 - 50 / 6 + 100 / 3 = 46.33.

[0057] It should be noted that the above example is only used as an example of a decision tree. In another embodiment, the local feature contribution value of the target property in each feature dimension in each decision tree can be obtained by following the above method. Then, the sum of the local feature contribution values ​​of the same feature dimension in different decision trees is taken as the feature contribution value of the target property in that feature dimension.

[0058] In some embodiments, the decision tree model is constructed using CatBoost. The target property's listing data and the listing data of each reference property in the reference property set are input into the interpretive model. The interpretive model uses the listing data of each reference property in the reference property set as the background dataset and calculates the local SHAP (SHapley Additive exPlanations, an interpretable algorithm) value of the target property relative to the background dataset using the TreeSHAP (Tree-based SHapley Additive exPlanations) built into the interpretive model. The local SHAP value represents the feature contribution value of the target property in each feature dimension.

[0059] In some embodiments, before step S202, the predicted transaction price is obtained in advance using the housing data of the target property through an interpretation model. That is, the process of obtaining the predicted transaction price includes: processing the housing data of the target property through an interpretation model to obtain the predicted transaction price of the target property.

[0060] In this embodiment, the target property's listing data is input into an interpretive model, which processes the data and outputs a predicted transaction price. The interpretive model is used beforehand to obtain the predicted transaction price from the target property's listing data. Then, based on this interpretive model and combined with the listing data of reference properties, the impact of the target property on various feature dimensions is analyzed, thereby obtaining the feature contribution value of each feature to the predicted transaction price. In this way, the price prediction process is transformed into a transparent and traceable decision-making basis through the interpretive model, enabling users to clearly understand the key driving factors influencing housing prices and improving the credibility of the predicted transaction price.

[0061] In some embodiments, the interpretive model is a decision tree model, which includes multiple decision trees, each of which includes a decision node corresponding to at least one of the multiple feature dimensions. The process of obtaining the predicted transaction price of the target property using the property data through the interpretive model includes: processing the property data of the target property through each decision tree in the interpretive model to obtain the output value corresponding to each decision tree; and determining the predicted transaction price of the target property by summing the price benchmark value corresponding to the interpretive model and the output values ​​corresponding to the multiple decision trees.

[0062] Decision tree models are predictive models that mimic a tree structure, consisting of multiple decision trees. Combining different decision trees improves the accuracy and stability of the model. Each decision tree includes decision nodes and leaf nodes, classifying or regressing property data through a series of "if...then..." rules (decision nodes). Each decision node makes a judgment based on a feature dimension, branching the property data into different branches until the final output leaf node is reached. Each leaf node has an output value. Decision nodes are the decision points in the decision tree, branching property data based on a specific threshold for a feature dimension (e.g., "Is the property older than 10 years?"). Leaf nodes, also called terminal nodes or leaf nodes, are the lowest-level nodes in the decision tree, representing the end of the decision path. The price benchmark value is the starting or base value predicted by the decision tree model; it can be understood as the average price or initial price before considering all features. The price benchmark value is calculated by using the average transaction price of all properties in the training dataset of the interpreting model, the initial predicted value, or other statistical benchmarks. It serves as the starting point or adjustment benchmark for predicting transaction prices. The output value of a decision tree is an adjustment value (which may be positive or negative) that is output after the decision tree has "completed" its decision path based on the target property's data. By summing up the output values ​​of all decision trees, the overall feature adjustment amount for the price benchmark value is obtained. The price benchmark value is calculated using the average transaction price of all properties in the model training dataset, the initial predicted value, or other statistical benchmarks (such as the initial output of the model at the start of training). It serves as the starting point or adjustment benchmark for the predicted price.

[0063] In this embodiment, when price prediction for a target property is required, the property data is sequentially input into each decision tree in the decision tree model. Each decision tree, like an automated questionnaire, starts from the root node, performs a series of feature judgments based on the property data, and finally reaches a leaf node. The leaf node corresponds to a specific numerical output (i.e., the tree's "contribution value" or "adjustment value" to the predicted transaction price). After all decision trees have completed processing, the output values ​​of all decision trees are summed to obtain a total adjustment. Finally, this total adjustment is added to the inherent price benchmark value of the decision tree model, and the final result is the predicted transaction price of the target property. The judgment path of each decision tree (which features were used and what the judgment conditions were) is clearly visible, so the entire prediction process can be traced and explained. In this way, applying the integrated decision tree model to house price prediction successfully achieves an excellent balance between prediction accuracy and model interpretability. It not only provides accurate predicted transaction prices, but also provides transparent logic and detailed attribution to support the predicted transaction prices, thereby improving the interpretability and transparency of the predicted transaction prices and ensuring the accuracy of the assessment of the predicted transaction prices.

[0064] In some embodiments, the predicted transaction price is obtained when the feature contribution value is obtained by interpreting the model. That is, step S202 includes: processing the target property data based on the property data of the reference property by interpreting the model to obtain the predicted transaction price of the target property and the feature contribution value of the target property in each feature dimension.

[0065] In this embodiment of the disclosure, the interpretive model processes the target property's data based on the reference property's data to obtain feature contribution values ​​and predicted transaction prices. In this way, while using the reference property data for accurate valuation, the interpretive model automatically reflects the key driving factors of price formation, which not only improves the accuracy of the valuation results but also enhances the transparency and interpretability of the model's decision-making process, thereby improving user experience and business efficiency.

[0066] In some embodiments, the process of obtaining feature contribution values ​​and predicted transaction prices by interpreting the model based on the housing data of the target property and the housing data of the reference property includes: processing the housing data of the target property based on the housing data of the reference property using the interpreting model to obtain a price benchmark value and the feature contribution value of the target property in each feature dimension; and determining the predicted transaction price of the target property as the sum of the price benchmark value and the feature contribution value of the target property in each feature dimension.

[0067] In this embodiment, the interpretive model uses the listing data of the target property and the listing data of the reference property to calculate a price benchmark value that reflects the average level. Then, it evaluates the differences between the target property and the reference property across each feature dimension, thereby generating feature contribution values ​​for the target property in each feature dimension. Finally, the price benchmark value is added to all feature contribution values ​​to obtain the predicted transaction price of the target property. This not only achieves accurate prediction of the transaction price of the target property but also simultaneously analyzes the price impact of each feature dimension, making the interpretive model output both accurate and interpretable. This improves the transparency and practicality of the predicted transaction price.

[0068] In some embodiments, the price benchmark value is the sum of the benchmark values ​​corresponding to the multiple decision trees.

[0069] For example, according to the above Figure 3For the decision tree shown, the price benchmark value is 496, and the feature contribution values ​​corresponding to area, floor, and elevator are 166.33, 91.33, and 46.33, respectively. If the decision tree model contains only one decision tree, the predicted transaction price of the target property is 166.33 + 91.33 + 46.33 + 496 = 799.99 ≈ 800. If the decision tree model includes multiple decision trees, the predicted transaction price of the target property = price benchmark value + feature contribution value corresponding to each feature dimension. Here, the price benchmark value is the sum of the benchmark values ​​corresponding to multiple decision trees, and the feature contribution value corresponding to any feature dimension is the sum of the local feature contribution values ​​corresponding to that feature dimension in multiple decision trees.

[0070] It should be noted that the embodiments disclosed herein are illustrated by taking the example of obtaining feature contribution values ​​through an interpretation model. In another embodiment, it is not necessary to perform the above step S202. Instead, other methods are adopted to process the target property data based on the property data of the reference property to obtain the feature contribution value of the target property in each feature dimension.

[0071] S203. Based on the weights corresponding to the multiple third feature dimensions associated with the second feature dimension, the feature contribution value corresponding to the second feature dimension is assigned to each third feature dimension to obtain the allocation value corresponding to each third feature dimension. The third feature dimension is the first feature dimension corresponding to each single attribute in the composite attribute represented by the second feature dimension.

[0072] In this embodiment of the disclosure, the multiple feature dimensions include multiple first feature dimensions and second feature dimensions. The first feature dimension is used to represent a single attribute of the corresponding property, and the second feature dimension is used to represent a composite attribute of the corresponding property.

[0073] In this embodiment, based on the weights of multiple third feature dimensions (i.e., the first feature dimensions corresponding to each individual attribute constituting the composite attribute) associated with the second feature dimension (representing the composite attribute), the feature contribution value of the composite attribute is split and allocated to the third feature dimensions corresponding to each individual attribute. The allocated value is added to the feature contribution value of the original third feature dimension, and the updated third feature dimension is combined with the remaining fourth feature dimensions that do not participate in the composite to construct explanatory information. In this way, the rationality of feature contribution allocation, the completeness of explanatory information, and the consistency of user understanding are improved, thereby enhancing the user experience.

[0074] For example, if the composite attribute represented by the second feature dimension is "floor × whether there is an elevator", and the single attribute "floor" corresponds to the third feature dimension 1, and the single attribute "whether there is an elevator" corresponds to the third feature dimension 2, then the second feature dimension is associated with the third feature dimension 1 and the third feature dimension 2 respectively. The product of the weight of the third feature dimension 1 and the feature contribution value corresponding to the second feature dimension is used as the allocation value corresponding to the third feature dimension 1; the product of the weight of the third feature dimension 2 and the feature contribution value corresponding to the second feature dimension is used as the allocation value corresponding to the third feature dimension 1.

[0075] In some embodiments, the reference properties include multiple properties; the process of obtaining the weight corresponding to the third feature dimension includes: for each reference property, processing the property data of the reference property based on the property data of the target property and the property data of the other reference properties to obtain the feature contribution value of the reference property on each feature dimension; for each third feature dimension, determining the average value of the feature contribution values ​​of the multiple reference properties on the third feature dimension; determining the sum of the average values ​​corresponding to the multiple third feature dimensions respectively; and determining the ratio of the average value to the sum of the average values ​​corresponding to the third feature dimension as the weight corresponding to the third feature dimension.

[0076] The process of obtaining the feature contribution value of the reference property in each feature dimension is the same as step S202 above, and will not be repeated here.

[0077] In this embodiment, for each reference property, after determining the feature contribution value of each reference property on each feature dimension, the feature contribution value of each reference property on each third feature dimension can be determined. Then, for each third feature dimension, the feature contribution values ​​of all reference properties on that third feature dimension are aggregated and averaged. The averages of all third feature dimensions are summed to obtain a total, and the average of each third feature dimension is compared with this total to obtain a normalized weight. By dynamically calculating the weights of the third feature dimensions based on multiple reference properties, the actual influence of different single attributes in the composite attributes can be accurately reflected. Furthermore, the use of averaging and normalization not only enhances the robustness of the weight calculation but also ensures the fairness and comparability of the allocation process. Moreover, determining the weights based on reference properties makes the weights more realistic, ensuring that subsequent explanations are more targeted.

[0078] It should be noted that step S203 above is described with respect to a second feature dimension. In another embodiment, if there are multiple second feature dimensions among multiple feature dimensions, then step S203 above can be used to assign the feature contribution value corresponding to each second feature dimension to the associated third feature dimensions.

[0079] S204. The sum of the feature contribution value corresponding to the third feature dimension and the corresponding allocation value is determined as the updated feature contribution value of the third feature dimension.

[0080] In this embodiment of the disclosure, if the feature contribution value corresponding to the third feature dimension can be obtained according to the above step S202, and a portion of the allocation value can be obtained according to the above step S203, then the sum of the original feature contribution value and the allocation value is used as the updated feature contribution value of the third feature dimension.

[0081] S205. Construct explanatory information based on the updated feature contribution value of the third feature dimension and the feature contribution value corresponding to the fourth feature dimension. The fourth feature dimension is a feature dimension other than the third feature dimension among multiple first feature dimensions.

[0082] In this embodiment of the disclosure, the fourth feature dimension is equivalent to the first feature dimension, which represents a single attribute that does not belong to any composite attribute.

[0083] In this embodiment of the disclosure, explanatory information is constructed based on the updated feature contribution value of the third feature dimension and the feature contribution value corresponding to the fourth feature dimension, so that the explanatory information can indicate the impact of the feature dimensions corresponding to each single attribute of the target property on the predicted transaction price, instead of just the user representing the impact of the feature dimensions of the composite attributes on the predicted transaction price, making the explanatory information easier to understand.

[0084] For example, taking the second feature dimension used to represent "floor × whether or not there is an elevator" as an example, the feature contribution value corresponding to the second feature dimension needs to be decomposed into the "floor" dimension and the "elevator" dimension. In the reference property set, the average feature contribution value corresponding to each single attribute dimension is calculated by the interpretation model M2: the average feature contribution value corresponding to the single attribute dimension j is: mean|SHAP|_j=(1 / N)×Σ|SHAP_ij|, where i ranges from 1 to N, SHAP_ij is the feature contribution value of the i-th reference property on attribute dimension j, and N is the number of reference properties; for example, the mean|SHAP| of the "whether or not there is an elevator" dimension is 12,000 yuan; the mean|SHAP| of the "floor" dimension is 8,000 yuan. The weights are assigned as follows: w_A = mean|SHAP_A| / (mean|SHAP_A| + mean|SHAP_B|), w_B = 1 - w_A; for example, w_elevator = 12,000 / (12,000 + 8,000) = 0.6, w_floor = 0.4. The feature contribution value corresponding to the second feature dimension is shap_AB; therefore, the assigned value for the "elevator presence or absence" dimension = w_elevator × shap_AB, and the assigned value for the "floor" dimension = w_floor × shap_AB. Finally, the sum of the assigned value for the "elevator presence or absence" dimension and the original feature contribution value for the "elevator presence or absence" dimension is taken as the final feature contribution value for the "elevator presence or absence" dimension; the sum of the assigned value for the "floor" dimension and the original feature contribution value for the "floor" dimension is taken as the final feature contribution value for the "floor" dimension.

[0085] It should be noted that this embodiment of the disclosure uses the example of a feature dimension corresponding to a composite attribute among multiple feature dimensions for illustration. In another embodiment, steps S203-S205 are not required; instead, other methods are used to generate explanatory information for explaining the predicted transaction price based on the feature contribution values ​​corresponding to each of the multiple feature dimensions. In some embodiments, the feature contribution values ​​corresponding to each of the multiple feature dimensions are added to the explanatory information template to obtain the explanatory information for explaining the predicted transaction price. In this embodiment of the disclosure, by pre-setting the explanatory information template and embedding the feature contribution values ​​corresponding to each of the multiple feature dimensions into the explanatory information template, a clearly structured and semantically standardized explanatory text can be automatically generated to explain the formation logic of the predicted transaction price. In this way, the output format is unified, ensuring that the explanatory information of different properties has consistency and comparability.

[0086] The solution provided in this disclosure introduces reference properties that meet preset similarity criteria to the target property. Using the reference property's data as a comparison basis, the feature contribution value of the target property relative to the reference property in various feature dimensions is quantified. This generates explanatory information on the predicted transaction price that is business-interpretable and provides a basis for comparison. This not only reflects the reasons for the formation of the predicted transaction price, enabling real estate agents to intuitively understand and convey the rationality of the predicted transaction price to clients, but also enhances the credibility of the predicted transaction price.

[0087] Based on the embodiments shown above, the TreeSHAP algorithm can be used to obtain feature contribution values. The original SHAP algorithm suffers from high computational complexity due to the need to enumerate all feature subsets, making it unsuitable for high-dimensional scenarios. TreeSHAP, however, leverages the fixed structure of the tree model and, through dynamic programming and post-order traversal strategies, efficiently calculates the SHAP value (i.e., feature contribution value) for each feature dimension without enumerating subsets. Specifically, for any feature dimension, TreeSHAP calculates the conditional expectations f(S∪{i}) and f(S) at all nodes involved in the splitting process for two scenarios: when the feature dimension is known (the path is determined by the actual value of the target property) and when the feature dimension is unknown (the left and right subtrees are weighted according to the proportion of reference properties satisfying the splitting conditions in the background data). This yields the local feature contribution value gain_{i,S} = f(S∪{i}) for that feature dimension. f(S); then, all local contributions of this feature dimension are summed using Shapley weights to obtain the final SHAP value (i.e., the final feature contribution value) of this feature dimension. The baseline value is still the average predicted value E[f(X)] on the background data. This method reduces the computational complexity from exponential to polynomial level, balancing interpretability accuracy and computational efficiency, and is suitable for interpretability analysis of large-scale tree models.

[0088] In the above Figure 2 Based on the illustrated embodiments, the interpretive model needs to be trained in advance before use. In this embodiment, a prediction model is first trained using historical transaction data of properties and actual transaction prices; then, the trained prediction model is used to obtain the predicted transaction price of sample properties for unsold sample properties; finally, an interpretive model is trained using the sample property data and the predicted transaction price.

[0089] (a) The process of training the prediction model includes: acquiring the housing data of sold properties, actual transaction prices and listed prices; determining the ratio of actual transaction price to listed price as the transaction discount rate of sold properties; using the CatBoost regression model as the initial prediction model, taking the housing data of sold properties as the input of the CatBoost regression model, taking the transaction discount rate as the training label, and taking minimizing the error between the predicted discount rate and the actual discount rate as the optimization objective to train the CatBoost regression model.

[0090] The housing data includes features of sold properties across multiple feature dimensions. These dimensions include not only single-attribute features but also features corresponding to composite attributes (i.e., cross-attributes). These composite attributes are manually defined combinations of attributes with actual semantic meaning. The prediction model can use any loss function during training; for example, RMSE (Root Mean Square Error) can be used. The prediction model structure has arbitrary depth; for example, a depth of 7, comprising 7 decision trees, allows for complex feature interactions to maximize prediction accuracy.

[0091] In this embodiment of the disclosure, after the CatBoost regression model is trained, the correlation between each feature dimension and the predicted discount rate can be determined through the trained CatBoost regression model. This allows for optimization of the previously used transaction data or a new batch of transaction data based on the correlation, filtering out features with low correlation from the data. Then, using the optimized data, the CatBoost regression model is used as the initial prediction model to retrain a prediction model in the above manner. This achieves feature dimension selection, retaining only the feature dimensions that improve the predicted discount rate, while removing some semantically clear but non-accurate feature dimensions.

[0092] In this embodiment of the disclosure, the objective of the prediction model is to accurately predict the "discount of the transaction relative to the listing price (i.e. the asking price)". Compared to the absolute price, this objective has a smaller variance, which can reduce the difficulty of model learning.

[0093] (ii) Based on the prediction model obtained from the above training, obtain the predicted transaction price of each unsold property (i.e., property for sale): obtain the property data and asking price of the unsold property; input the property data of the unsold property into the prediction model, and the prediction model outputs the predicted discount rate; then the product of the asking price of the unsold property and the predicted discount rate is used as the predicted transaction price of the unsold property.

[0094] In this embodiment, the feature dimensions involved in the housing data of unsold properties are the same as those involved in the housing data used when training the prediction model. A predicted transaction price is determined for each property for sale in the manner described above, so that the predicted transaction price can be used as the training label for training the subsequent explanatory model.

[0095] (III) The process of training the explanatory model: Obtain the housing data and predicted transaction prices of sample properties. Sample properties are those that have not yet been sold. The predicted transaction price of a sample property is the product of its listed price and the transaction discount rate. The transaction discount rate is predicted by the prediction model based on the housing data of the sample properties. The prediction model is trained based on the housing data and transaction discount rates of sold properties. The transaction discount rate is the ratio of the transaction price to the listed price of a sold property. The explanatory model is then trained based on the housing data and predicted transaction prices of the sample properties. Here, sample properties are equivalent to properties currently for sale.

[0096] In this embodiment, a CatBoost regression model is used as the initial interpretive model. After predicting the transaction discount rate of the sample properties using the trained predictive model to obtain the predicted transaction price, the predicted transaction price is used as the label. The property data of the sample properties is used as the input to the predictive model, and the model is trained with the direct prediction of the absolute transaction price as the optimization objective. The interpretive model has an arbitrary depth, for example, a depth of 4 or 5, i.e., including 4 or 5 decision trees, which limits complexity and improves generalization and interpretive stability.

[0097] In this embodiment of the disclosure, the training of the prediction model may involve the selection of feature dimensions. Compared with the feature dimensions designed for the housing data used to train the prediction model, the feature dimensions involved in the housing data of the sample housings used to train the interpretive model may be increased. The increased feature dimensions have strong business semantics, such as apartment layout and building type. For composite attributes (i.e., cross attributes), only a few key features are retained to ensure the accuracy of the prediction results (e.g., "floor × whether there is an elevator"). This feature design principle can balance predictive ability and business interpretability.

[0098] In this embodiment of the disclosure, a prediction model is first trained, and then an explanation model is trained using the prediction model, thus achieving model distillation training. Based on the embodiments shown above, this disclosure provides a flowchart of a method for obtaining explanatory information for predicted transaction prices, as follows: Figure 4 As shown, the method includes: Step S401: Use historical transaction data and the ratio of actual transaction price to listed price to train the CatBoost model and obtain a prediction model.

[0099] Step S402: Using the trained prediction model, based on the housing data of the properties for sale, obtain the predicted discount rate of the properties for sale, and then combine it with the listed price of the properties for sale to calculate the predicted transaction price of the properties for sale, which will be used as a soft label for subsequent training of the interpretation model.

[0100] Step S403: Obtain housing data for properties for sale, use the predicted transaction price of the properties for sale as a label, and train the interpretative model using model distillation.

[0101] Step S404: For the target property, determine the reference property set. Through the interpretation model, based on the property data of each reference property in the reference property set and the property data of the target property, calculate the feature contribution value of the target property's features in each feature dimension to the predicted transaction price.

[0102] Step S405: For the feature dimension used to represent the composite attribute, calculate the weights corresponding to the feature dimensions of each individual attribute contained in the composite attribute. According to the weights, assign the feature contribution values ​​corresponding to the feature dimension used to represent the composite attribute to the feature dimensions of each individual attribute contained in the composite attribute, and output the final explanatory information used to explain the predicted transaction price.

[0103] This disclosure provides a high-precision, interpretable price prediction method for second-hand housing valuation scenarios. By constructing a dual-model decoupled architecture of a complex valuation model and a shallow interpretive model, it improves the interpretability of predicted transaction prices while ensuring prediction accuracy. Specifically, a high-precision valuation model (i.e., a prediction model) is first trained, using the ratio of transaction price to listed price as the target variable to fully capture market pricing patterns. Subsequently, the predicted transaction prices generated by this valuation model for a large number of properties for sale (e.g., tens of millions of samples) are used as soft labels to distill and train a shallower interpretive model that relies solely on feature dimensions with clear business semantics, using absolute price as the regression target. This interpretive model not only retains the core predictive capabilities of the original valuation model but also ensures that subsequent feature attribution results can be directly converted into "area contribution + 150,000 yuan" and "influence of building age." The use of intuitive and quantifiable business language, such as "80,000 yuan," facilitates communication with real estate agents and users.

[0104] Furthermore, this disclosure proposes a composite attribute attribution allocation mechanism based on feature importance. During the distillation process, the influence of composite attributes on prediction is implicitly retained in the soft label. In the interpretation stage, when using local attribution methods such as SHAP to calculate feature contributions, the SHAP values ​​corresponding to composite attributes are weighted and split according to the global importance ratio of their constituent individual attributes in the interpretation model, and reasonably mapped back to the feature dimensions used to represent individual attributes. Thus, the discriminative power of complex models is preserved while avoiding interference from uninterpretable features on the final explanatory information.

[0105] Furthermore, this embodiment of the disclosure employs a set of reference properties similar to the target property in the selection of SHAP background data, realizing a local interpretation logic of "similar comparison," making the attribution results more consistent with the actual market context. In summary, this embodiment of the disclosure not only achieves interpretability of predicted transaction prices and realizes efficient, reliable, and actionable white-box price interpretation for brokerage operations, but also improves human-machine collaboration efficiency, enhances communication efficiency between real estate agents and clients, and facilitates transaction completion.

[0106] It should be noted that the above-mentioned multiple optional embodiments can be combined in any way, and this application does not limit this.

[0107] Figure 5 This is a schematic diagram of a device for obtaining explanatory information about predicted transaction prices, provided in an embodiment of this disclosure. This device can be understood as the aforementioned electronic device or a functional module within the aforementioned electronic device. Figure 5 As shown, the device for obtaining explanatory information regarding predicted transaction prices includes: The acquisition module 501 is used to acquire the housing data of the target housing and the housing data of the reference housing. The housing data indicates the characteristics of the corresponding housing in multiple feature dimensions. The housing data of the target housing and the housing data of the reference housing meet the preset similarity conditions. The processing module 502 is used to process the housing data of the target housing based on the housing data of the reference housing, and obtain the feature contribution value of the target housing in each feature dimension; wherein, the feature contribution value indicates the impact of the target housing on the predicted transaction price of the target housing in the feature dimension compared with the reference housing. The generation module 503 is used to generate explanatory information for interpreting the predicted transaction price based on the feature contribution values ​​corresponding to multiple feature dimensions.

[0108] In some embodiments, the multiple feature dimensions include multiple first feature dimensions and second feature dimensions. The first feature dimension is used to represent a single attribute of the corresponding property, and the second feature dimension is used to represent a composite attribute of the corresponding property. The generation module 503 is used to allocate the feature contribution value corresponding to the second feature dimension to each third feature dimension based on the weights corresponding to the multiple third feature dimensions associated with the second feature dimension, thereby obtaining the allocation value corresponding to each third feature dimension. The third feature dimension is the first feature dimension corresponding to each single attribute in the composite attribute represented by the second feature dimension. The sum of the feature contribution value corresponding to the third feature dimension and the corresponding allocation value is determined as the updated feature contribution value of the third feature dimension. Based on the updated feature contribution value of the third feature dimension and the feature contribution value corresponding to the fourth feature dimension, explanatory information is constructed. The fourth feature dimension is a feature dimension other than the third feature dimension among the multiple first feature dimensions.

[0109] In some embodiments, the reference properties include multiple properties; the processing module 502 is further configured to, for each reference property, process the property data of the reference property based on the property data of the target property and the property data of the remaining reference properties to obtain the feature contribution value of the reference property on each feature dimension; for each third feature dimension, determine the average value of the feature contribution values ​​of the multiple reference properties on the third feature dimension; determine the sum of the average values ​​corresponding to the multiple third feature dimensions respectively; and determine the ratio of the average value to the sum of the average values ​​corresponding to the third feature dimension as the weight corresponding to the third feature dimension.

[0110] In some embodiments, the processing module 502 is used to process the housing data of the target housing based on the housing data of the reference housing through an interpretation model to obtain the feature contribution value of the target housing in each feature dimension; the predicted transaction price is obtained through the interpretation model.

[0111] In some embodiments, the processing module 502 is further configured to process the housing data of the target housing through an interpretation model to obtain the predicted transaction price of the target housing.

[0112] In some embodiments, the interpretation model is a decision tree model, which includes multiple decision trees, each of which includes a judgment node corresponding to at least one of the multiple feature dimensions; the processing module 502 is used to process the housing data of the target housing through each decision tree in the interpretation model to obtain the output value corresponding to each decision tree; and the sum of the price benchmark value corresponding to the interpretation model and the output values ​​corresponding to the multiple decision trees is determined as the predicted transaction price of the target housing.

[0113] In some embodiments, the processing module 502 is used to process the target property data based on the property data of the reference property through an interpretation model, so as to obtain the predicted transaction price of the target property and the feature contribution value of the target property in each feature dimension.

[0114] In some embodiments, the processing module 502 is used to process the housing data of the target housing based on the housing data of the reference housing through an interpretation model to obtain a price benchmark value and the feature contribution value of the target housing in each feature dimension; and to determine the sum of the price benchmark value and the feature contribution value of the target housing in each feature dimension as the predicted transaction price of the target housing.

[0115] In some embodiments, both the target property and the reference property are unsold properties; the device further includes: The acquisition module 501 is also used to acquire the housing data of sample housing and the predicted transaction price of sample housing. The sample housing is the housing that has not been sold. The predicted transaction price of the sample housing is the product of the listed price of the sample housing and the transaction discount rate. The transaction discount rate is predicted by the prediction model based on the housing data of sample housing. The prediction model is trained based on the housing data of sold housing and the transaction discount rate of sold housing. The transaction discount rate is the ratio of the transaction price of sold housing to the listed price. The training module is used to train the explanatory model based on the housing data of the sample housing units and the predicted transaction prices of the sample housing units.

[0116] In some embodiments, the preset similarity criteria include at least one of the following: The similarity between the listing data of the target property and the listing data of the reference property is greater than the first threshold; The similarity between the features of the target property and the features of the reference property in the target feature dimension is greater than the second threshold.

[0117] It should be noted that the apparatus provided in this embodiment can execute the explanatory information acquisition method for predicted transaction prices in any of the above embodiments, and its execution method and beneficial effects are similar, so they will not be described again here.

[0118] This disclosure also provides an electronic device, which includes: a memory storing a computer program; and a processor for executing the computer program. When the computer program is executed by the processor, it can implement the method for obtaining explanatory information for predicting transaction prices as described in the above method embodiments.

[0119] Example, Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of this disclosure. See below for details. Figure 6The diagram illustrates a structural schematic suitable for implementing the electronic device 600 in the embodiments of this disclosure. The electronic device 600 in the embodiments of this disclosure may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0120] like Figure 6 As shown, electronic device 600 may include a processing device (e.g., a central processing unit, a graphics processor, etc.) 601, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 602 or a program loaded from storage device 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of electronic device 600. Processing device 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0121] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic device 600 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 An electronic device 600 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0122] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the method for obtaining explanatory information for predicting transaction prices as shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a storage device 608, or installed from a ROM 602. When the computer program is executed by the processing device 601, it performs the functions defined in the method for obtaining explanatory information for predicting transaction prices according to embodiments of this disclosure.

[0123] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0124] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, enable the electronic device to implement the method for obtaining explanatory information regarding predicted transaction prices as described in the above embodiments.

[0125] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, and portable compact disc read-only memory (CD). The storage medium may be a ROM, an optical storage device, a magnetic storage device, or any suitable combination thereof. Embodiments of this disclosure also provide a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the method for obtaining explanatory information regarding predicted transaction prices as described in any of the above embodiments. The execution method and beneficial effects are similar and will not be repeated here.

[0126] The above are merely specific embodiments of this disclosure, enabling those skilled in the art to understand or implement this disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to these embodiments, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0127] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0128] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0129] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for obtaining explanatory information for predicting transaction prices, characterized in that, include: Acquire the housing data of the target property and the housing data of the reference property. The housing data indicates the characteristics of the corresponding property in multiple feature dimensions. The housing data of the target property and the housing data of the reference property meet a preset similarity condition. Based on the housing data of the reference housing, the housing data of the target housing is processed to obtain the feature contribution value of the target housing in each feature dimension; wherein, the feature contribution value indicates the impact of the target housing on the predicted transaction price of the target housing in each feature dimension compared with the reference housing. Based on the feature contribution values ​​corresponding to the multiple feature dimensions, explanatory information is generated to explain the predicted transaction price.

2. The method according to claim 1, characterized in that, The multiple feature dimensions include multiple first feature dimensions and second feature dimensions. The first feature dimensions represent a single attribute of the corresponding property, and the second feature dimensions represent a composite attribute of the corresponding property. The step of generating explanatory information to explain the predicted transaction price based on the feature contribution values ​​corresponding to the multiple feature dimensions includes: Based on the weights corresponding to the multiple third feature dimensions associated with the second feature dimension, the feature contribution value corresponding to the second feature dimension is assigned to each third feature dimension to obtain the assignment value corresponding to each third feature dimension. The third feature dimension is the first feature dimension corresponding to each single attribute in the composite attribute represented by the second feature dimension. The sum of the feature contribution value corresponding to the third feature dimension and the corresponding allocation value is determined as the updated feature contribution value of the third feature dimension. The explanatory information is constructed based on the updated feature contribution value of the third feature dimension and the feature contribution value corresponding to the fourth feature dimension, wherein the fourth feature dimension is a feature dimension other than the third feature dimension among the plurality of first feature dimensions.

3. The method according to claim 2, characterized in that, The reference properties include multiple ones; before assigning the feature contribution value corresponding to the second feature dimension to each third feature dimension based on the weights corresponding to the multiple third feature dimensions associated with the second feature dimension, and obtaining the allocation value corresponding to each third feature dimension, the method further includes: For each reference property, based on the property data of the target property and the property data of the other reference properties, the property data of the reference property is processed to obtain the feature contribution value of the reference property on each feature dimension. For each third feature dimension, determine the average value of the feature contribution values ​​of multiple reference properties corresponding to the third feature dimension; Determine the sum of the average values ​​corresponding to the plurality of third feature dimensions; The ratio of the average value corresponding to the third feature dimension to the sum value is determined as the weight corresponding to the third feature dimension.

4. The method according to claim 1, characterized in that, The process of processing the target property's data based on the reference property's data to obtain the target property's feature contribution value for each feature dimension includes: By using the interpretive model, the housing data of the target housing is processed based on the housing data of the reference housing to obtain the feature contribution value of the target housing in each feature dimension; the predicted transaction price is obtained through the interpretive model.

5. The method according to claim 4, characterized in that, The method further includes: The property data of the target property is processed using the interpretation model to obtain the predicted transaction price of the target property.

6. The method according to claim 5, characterized in that, The interpretation model is a decision tree model, which includes multiple decision trees, each including a decision node corresponding to at least one of the multiple feature dimensions; the step of processing the property data of the target property through the interpretation model to obtain the predicted transaction price of the target property includes: The housing data of the target housing is processed through each decision tree in the interpretation model to obtain the output value corresponding to each decision tree; The sum of the price benchmark value corresponding to the explanation model and the output values ​​corresponding to the multiple decision trees is determined as the predicted transaction price of the target property.

7. The method according to claim 4, characterized in that, The step involves processing the target property's data based on the reference property's data using an interpretive model to obtain the target property's feature contribution value for each feature dimension, including: Using the interpretation model, based on the housing data of the reference housing, the housing data of the target housing is processed to obtain the predicted transaction price of the target housing and the feature contribution value of the target housing in each feature dimension.

8. The method according to claim 7, characterized in that, The step involves processing the target property's data using the interpretation model, based on the reference property's data, to obtain the target property's predicted transaction price and the target property's feature contribution value for each feature dimension. This includes: Using the interpretation model, based on the housing data of the reference housing, the housing data of the target housing is processed to obtain the price benchmark value and the feature contribution value of the target housing in each feature dimension. The predicted transaction price of the target property is determined by summing the price benchmark value and the feature contribution value corresponding to each feature dimension of the target property.

9. The method according to claim 4, characterized in that, Both the target property and the reference property are unsold properties; before processing the property data of the target property based on the property data of the reference property using the interpretation model to obtain the feature contribution value of the target property on each feature dimension, the method further includes: The system obtains the housing data of sample properties and the predicted transaction price of the sample properties. The sample properties are those that have not yet been sold. The predicted transaction price of the sample properties is the product of the listed price and the transaction discount rate. The transaction discount rate is predicted by the prediction model based on the housing data of the sample properties. The prediction model is trained based on the housing data of sold properties and the transaction discount rate of sold properties. The transaction discount rate is the ratio of the transaction price of the sold property to the listed price. The explanatory model is trained based on the housing data of the sample properties and the predicted transaction price of the sample properties.

10. An electronic device, characterized in that, The electronic device includes: a processor; a memory for storing executable instructions of the processor; the processor being configured to read the executable instructions from the memory and execute the executable instructions to implement the method for obtaining explanatory information for predicting transaction prices as described in any one of claims 1 to 9.

Images