Land resource supply and demand intelligent matching method

Abstract

The application discloses a land resource supply and demand intelligent matching method, and belongs to the technical field of data processing and intelligent recommendation. The method comprises the following steps: acquiring the interactive behavior data of demand parties on land plots, historical land acquisition records, basic information and preset behavior type depth weights; determining a decision cycle reference value based on the historical land acquisition records; performing time decay processing on preference scores according to the decision cycle reference value, the depth weights of corresponding behavior types and behavior occurrence times; determining the similarity between demand parties according to the basic information, the historical land acquisition records and the interactive behavior data; and determining the predicted interest degree of demand parties on non-interacted land plots and generating a recommendation result according to a land plot similarity matching algorithm, the similarity, the decayed preference scores and the similarity. Through the implementation of the application, the problem that the traditional recommendation algorithm adopts a global unified time decay function to cause preference distortion can be solved, adaptive modeling of user preference timeliness can be realized, and the matching accuracy of supply and demand can be improved.

Description

Technical Field

[0001] This application relates to the fields of data processing and intelligent recommendation technology, and in particular to an intelligent matching method for land resource supply and demand. Background Technology

[0002] In the land transfer process, suppliers need to find suitable developers with corresponding development capabilities for each plot, while demanders need to efficiently select target plots from numerous options that align with their strategic plans. Whether supply and demand can achieve precise matching directly determines whether land resources flow to the most suitable development entities. Therefore, constructing a recommendation method that can intelligently match land resources with the needs of development enterprises is of great significance for improving the operational efficiency of the land factor market.

[0003] Currently, the supply and demand matching in the land factor market mainly relies on government announcements, offline promotional events, or manual matchmaking, resulting in information asymmetry between the two sides. Demand-sides struggle to efficiently filter land information that matches their own conditions, while suppliers lack the ability to accurately identify potential developers. In recent years, some platforms have attempted to digitize land information, introducing keyword search and collaborative filtering recommendation mechanisms. Regarding user preference modeling, existing methods typically construct interest scores based on interactive behavior and use time decay functions to handle the timeliness of preferences. However, land resource transactions, whether for real estate development, industrial land, or commercial service land, are generally characterized by large transaction amounts, long decision-making cycles, and a clear cyclical pattern in corporate land acquisition behavior. Existing methods' globally uniform time decay functions ignore the differences in strategic depth between different interactive behaviors and the differences in decision-making cycles among different demand-sides. For example, casual browsing and formal consultation are assigned the same decay curve; the short-term agility needs of companies with rapid capital turnover and the long-term strategic intentions of companies that are cautious in land acquisition are treated equally. This leads to some demand-sides' recent needs being submerged in outdated data, while the long-term intentions of others are forgotten due to rapid decay.

[0004] In summary, existing technologies, due to their use of a globally uniform time decay function in long-cycle, high-decision-weighted land resource transactions, cannot adapt to the differences in decision-making pace and strategic depth of various demanders and their interactions. This results in insufficient modeling of user preference timeliness, leading to a systematic deviation between recommended results and users' true strategic intentions. Therefore, there is an urgent need for an intelligent matching method that can optimize preference timeliness modeling for land resource supply and demand matching scenarios to achieve efficient matching and accurate recommendations for both supply and demand sides. Summary of the Invention

[0005] This application provides an intelligent matching method for land resource supply and demand, aiming to solve the technical problem that existing recommendation algorithms in land resource transaction scenarios cannot adapt to long-term and decision-making characteristics due to the use of a globally unified time decay function, resulting in defects in the modeling of user preference timeliness and insufficient matching accuracy between supply and demand. The method aims to achieve efficient matching and accurate recommendation of land plots between supply and demand.

[0006] Firstly, this application provides a method for intelligent matching of land resource supply and demand, including: The system acquires data on the interaction behavior of demanders with respect to land parcels, the demanders' historical land acquisition records, the demanders' basic information, and preset behavior type depth weights. The interaction behavior data includes behavior type and the time of occurrence of the behavior. The preset behavior type depth weights represent the differences in the intensity of preference indications for different interaction behavior types. Based on the historical land acquisition records, a decision cycle benchmark value for the demand side is determined; based on the decision cycle benchmark value, the preset behavior type depth weight corresponding to the behavior type, and the behavior occurrence time, the preference score corresponding to the interaction behavior data is subjected to time decay processing to obtain a decayed preference score. Based on the basic information of the demander, the demander's historical land acquisition records, and the interaction behavior data, the similarity between the demander and other demanders is determined; based on the similarity, the decay preference scores of the other demanders, and the comprehensive similarity between land parcels, the predicted interest of the demander in non-interacting land parcels is determined; and land parcel recommendation results are generated based on the predicted interest.

[0007] This application effectively addresses the shortcomings of traditional recommendation algorithms in modeling the timeliness of preferences in land resource transaction scenarios by establishing a personalized dynamic time decay mechanism. Specifically, a decision-making cycle benchmark is determined based on the historical land acquisition records of the demanders. This benchmark quantifies the frequency characteristics of land acquisition by enterprises, allowing enterprises with fast capital turnover to obtain a shorter decision-making cycle representation, while enterprises with stable turnover and prudent land acquisition obtain a longer decision-making cycle representation. Simultaneously, a preset depth weight for different behavior types is introduced to distinguish the strategic depth of different interaction behaviors; for example, the depth weight corresponding to bidding behavior is higher than that corresponding to browsing behavior. Based on this, the decision-making cycle benchmark, the corresponding depth weight, and the time of behavior occurrence are jointly applied to the time decay processing of preference scores. This allows preference scores corresponding to shallow browsing behavior to decay rapidly to avoid noise interference, while preference scores corresponding to deep bidding or consultation behavior decay slowly to retain long-term strategic intent. This mechanism fundamentally overcomes the problem of preference distortion caused by differences in decision-making cycles and strategic depth of different behaviors among enterprises, improving the alignment between preference prediction results and users' actual land acquisition intentions. Building upon accurate preference modeling, this application further integrates multi-dimensional similarity calculation and a hybrid recommendation framework. A multi-dimensional similarity measurement system based on user profile features and interactive behavior features ensures the quality of neighbor user filtering. A hybrid recommendation strategy combining plot similarity matching algorithms compensates for the information blind spots of a single recommendation source, enabling the generation of plot recommendations highly aligned with the true strategic intentions of each user. Compared to existing recommendation methods that use a globally uniform time decay function, this application introduces a benchmark value for the enterprise's personalized decision-making cycle and behavioral depth weights to achieve adaptive dynamic adjustment of the time decay half-life. Existing methods cannot distinguish between the short-term needs and long-term intentions of different enterprises, leading to outdated or distorted recommendation results. This application, through differentiated decay processing, can simultaneously and accurately capture the recent agile needs of high-turnover enterprises and effectively lock in the deep strategic intentions of stable enterprises. Furthermore, existing methods typically rely on single collaborative filtering or simple attribute matching, resulting in a single matching dimension. This application, through a hybrid recommendation framework integrating user similarity and plot similarity, further improves the comprehensiveness and accuracy of the recommendation results.

[0008] Furthermore, determining the benchmark value for the decision-making cycle of the demander based on the historical land acquisition records includes: Extract the historical land acquisition time of the demander to determine the time interval between adjacent land acquisition activities; The average time interval is obtained by averaging the time intervals. The average land acquisition time interval is normalized to obtain the decision cycle benchmark value.

[0009] By extracting and averaging the land acquisition time intervals from historical land acquisition records of demanders, the dispersed land acquisition behaviors of enterprises are transformed into a stable statistical value. This effectively eliminates the interference of single abnormal transactions on the evaluation of the decision-making cycle, enabling a quantitative understanding of the differentiated rhythms of high-turnover enterprises acquiring land frequently and stable enterprises acquiring land infrequently. Based on this, the average land acquisition time interval is normalized, mapping the absolute time span of different enterprises to a unified dimension. This allows the calculation of the decay half-life in subsequent time decay processing to be performed within a unified numerical framework, ensuring the comparability and rationality of preference decay rates among enterprises of different sizes. This process provides a personalized decision-making cycle benchmark value for the dynamic time decay mechanism, serving as a quantitative basis for achieving differentiated management of the preference maintenance cycle of different demanders.

[0010] Further, the step of performing time-decay processing on the preference score corresponding to the interaction behavior data based on the decision cycle benchmark value, the preset behavior type depth weight corresponding to the behavior type, and the behavior occurrence time to obtain a decayed preference score includes: The adaptive decay half-life is determined based on the decision cycle baseline value, the depth-sensitive adjustment parameter, and the preset behavior type depth weight corresponding to the behavior type; wherein, the higher the preset behavior type depth weight, the longer the corresponding half-life, and the depth-sensitive adjustment parameter is used to control the degree of influence of the preset behavior type depth weight on the half-life. The time interval between the occurrence of the behavior and the current time is determined based on the difference between the time when the behavior occurred and the current time. Using the adaptive decay half-life and the time interval since the behavior, the preference score corresponding to the interaction behavior data is subjected to time decay processing to obtain the decayed preference score.

[0011] By combining the decision-making cycle benchmark with the preset depth weights of corresponding behavior types, the personalized land acquisition pace of enterprises and the strategic depth of the behavior itself are integrated into a unified decay half-life. This allows different enterprises and different behavior types to obtain differentiated decay rate control parameters, fundamentally overcoming the shortcomings of traditional methods that use the same decay function for all users and all behaviors. Based on this, an exponential decay function is used to calculate the time decay factor by relating the behavior occurrence time to the decay half-life. The mathematical properties of the exponential function give higher weight to recent behaviors while gradually reducing the impact of long-term behaviors. The decay rate is precisely controlled by the decay half-life, conforming to the regular changes in user preferences over time. Finally, the time decay factor is used to weight preference scores, transforming the original preference scores into decayed preference scores. This allows for long-term retention of memories of high-value, deep-level behaviors while rapidly fading memories of low-value, shallow-level behaviors, ultimately achieving a high degree of consistency between preference prediction results and users' actual strategic intentions in the current period.

[0012] Furthermore, determining the similarity between the demander and other demanders based on the demander's basic information, historical land acquisition records, and interaction behavior data includes: Extract profile attributes from the basic information of the demander and the demander’s historical land acquisition records, and construct a profile feature vector; Explicit interaction behaviors are extracted from the interaction behavior data, and explicit behavior feature vectors are constructed. Implicit interaction behaviors are extracted from the interaction behavior data to construct implicit behavior feature vectors; Calculate the similarity of portrait features, explicit behavioral features, and implicit behavioral features respectively; Adjustable weights are assigned to the portrait feature similarity, explicit behavioral feature similarity, and implicit behavioral feature similarity; the similarity is obtained by weighted summation of the portrait feature similarity, explicit behavioral feature similarity, and implicit behavioral feature similarity.

[0013] This application constructs profile feature vectors, explicit behavioral feature vectors, and implicit behavioral feature vectors respectively, establishing the similarity measurement of demanders on three dimensions: static attributes, explicit intent, and potential interests. Profile feature vectors reflect long-term, stable investment attributes such as a company's capital strength, business preferences, and strategic geographical layout; explicit behavioral feature vectors capture high-value interactive information that explicitly expresses intent, such as collections, inquiries, and bidding; implicit behavioral feature vectors mine implicit interest signals such as browsing time and click frequency, compensating for the sparsity of explicit behavioral data. Calculating the similarity of the three types of features separately allows each dimension to be calculated independently within its respective suitable metric space, avoiding dimensional differences and feature overload caused by directly mixing features of different natures. Furthermore, by assigning adjustable weights to the three types of similarity and performing a weighted summation, the contribution ratio of each dimension can be flexibly adjusted according to the actual scenario. The final fused similarity simultaneously considers the convergence of strategic backgrounds, consistency of explicit intent, and similarity of potential interests, providing a more comprehensive and reliable basis for selecting neighboring users for collaborative filtering recommendations.

[0014] Furthermore, the step of extracting profile attributes from the basic information of the demander and the demander's historical land acquisition records to construct a profile feature vector includes: The profile attributes are extracted from the basic information of the demander and the demander's historical land acquisition records; wherein, the profile attributes include at least one of the following: enterprise asset status, business type, and intended region; The extracted portrait attributes are converted into label vectors to obtain the portrait features.

[0015] By extracting profile attributes such as enterprise asset status, business type, and intended region from basic information and historical land acquisition records, the static information of demanders is condensed into key feature dimensions with representational capabilities. This makes the core investment attributes of enterprises, such as capital strength, development expertise, and spatial preferences, explicit. Based on this, the extracted profile attributes are converted into label vectors, uniformly representing the attribute differences between different demanders in a structured numerical form. This allows profile features to directly participate in subsequent similarity calculations, avoiding the computational complexity and matching bias caused by direct comparison of unstructured information. This provides a standardized feature foundation for constructing a high-quality set of neighboring users.

[0016] Furthermore, the explicit interactive behavior includes at least one of bidding, consultation, and favorites; the implicit interactive behavior includes at least one of page browsing, repeated visits, and browsing duration.

[0017] This application provides a clear and operational definition of data sources for constructing behavioral feature vectors by explicitly defining the specific types of explicit and implicit interactive behaviors. Classifying bidding, inquiries, and favorites as explicit interactive behaviors ensures the accurate capture of high-value intent signals; these behaviors directly reflect the clear investment intentions of the demanders and have a high intensity of preference indication. Classifying page browsing, repeated visits, and browsing duration as implicit interactive behaviors enables the extraction of potential interest signals from the demanders' natural browsing trajectories, filling the gap in preference information when demanders have not actively expressed their intentions. These two types of behaviors each have their own focus and complement each other, together forming a complete foundation for behavioral feature data, avoiding the problem of one-sided or sparse preference information caused by relying on only a single behavior type.

[0018] Further, determining the predicted interest of the demander in non-interacting land parcels based on the similarity score, the decay preference score of the other demanders, and the comprehensive similarity between land parcels includes: Based on the similarity and the decay preference scores of the other demanders, a first predicted interest level is obtained; The second predicted interest degree is obtained based on the similarity matching algorithm of the plots; The first predicted interest score and the second predicted interest score are weighted and fused to obtain the predicted interest score; wherein the weights of the weighted fusion are dynamically adjusted according to historical transaction data.

[0019] This application constructs a hybrid recommendation framework by splitting the determination of predicted interest into two independent paths and then weighting and fusing them. The first predicted interest, obtained based on user similarity and neighbor preference scores, identifies potentially interesting plots of land from the historical behavior of similar demanders, reflecting the collaborative recommendation logic of "similar companies, similar choices." The second predicted interest, obtained based on a plot similarity matching algorithm, recommends other plots similar to those historically preferred by the target user, considering the plot's own attributes and spatial value, reflecting the content recommendation logic of "similar plots, similar value." The two paths independently predict from the user and plot dimensions respectively, effectively compensating for the information blind spots of a single recommendation source. Furthermore, the weights of the weighted fusion are dynamically adjusted based on historical transaction data, allowing the two prediction results to adaptively tilt towards better actual transaction results during the fusion process. This avoids the problem of insufficient adaptability of fixed weights in different market environments, ultimately improving the alignment between the fused predicted interest and actual transaction intentions.

[0020] Further, obtaining the first predicted interest level based on the similarity and the decay preference scores of the other demanders includes: Using the similarity as a weight, the decay preference scores of the other demanders are summed in a weighted manner to obtain a weighted preference score. The absolute values ​​of the similarities are summed to obtain the similarity normalization factor; The weighted preference score is normalized using the similarity normalization factor to obtain the first predicted interest level.

[0021] In calculating the first predicted interest, this application uses similarity as a weight to perform a weighted summation of neighbor preference scores. This ensures that the more similar a neighbor is to the target user, the greater the influence of their preferences on the prediction result, avoiding prediction bias caused by treating all neighbors equally. A normalization factor is obtained by summing the absolute values ​​of similarity, and this factor is used to normalize the weighted summation result. This eliminates the impact of differences in the size of different user neighbor sets and the magnitude of differences in the absolute values ​​of similarity on the prediction result. This ensures that the first predicted interest between different users and different plots has a unified numerical dimension and comparability, guaranteeing numerical fairness when subsequently merging with the second predicted interest.

[0022] Further, the step of obtaining the second predicted interest degree based on the similarity matching algorithm of the land parcels includes: Obtain data on supporting facilities around the land parcel, and determine the scarcity weight of each type of facility based on the global distribution density of the supporting facility data; Calculate the spatial potential energy value of the land parcel based on the scarcity weight and the synergistic effect between facilities. The spatial similarity bandwidth is determined based on the spatial potential energy value. The spatial similarity between land parcels is determined based on the spatial similarity bandwidth. Determine the similarity of land parcel attributes based on the numerical attributes of the land parcels; The spatial similarity and the land parcel attribute similarity are weighted and fused to obtain the comprehensive land parcel similarity. The second predicted interest level is determined based on the comprehensive similarity of the land parcels and the historical preferences of the demanders for land parcels.

[0023] In calculating the second predicted interest degree, this application first acquires data on surrounding facilities of the land parcel and determines the scarcity weight of various facilities based on global distribution density. Borrowing the concept of inverse document frequency, it assigns a greater weight to scarce facilities with sparser distribution, overcoming the shortcomings of traditional methods that treat all facility types equally and fail to reflect the differences in real market scarcity value. Based on this, it introduces the synergistic effect between facilities to calculate the spatial potential value of the land parcel. Through cross-product terms, the nonlinear added value effect generated by the combination of core facilities is explicitly quantified, making the spatial potential value more realistically reflect the comprehensive locational value of the land parcel. An adaptive spatial similarity bandwidth is determined based on the spatial potential value, allowing high-potential, bustling areas to use a smaller bandwidth for greater sensitivity to spatial distance, while low-potential, remote areas use a larger bandwidth for greater tolerance to spatial distance, achieving differentiated processing of spatial similarity calculation. Finally, the spatial similarity and land parcel attribute similarity are weighted and fused to obtain the comprehensive land parcel similarity. This, combined with the historical preferences of the demand side, determines the second predicted interest degree, ensuring that the recommendation results reflect both the objective value of the land parcel and align with the historical land parcel preferences of individual users.

[0024] Further, generating land parcel recommendation results based on the predicted interest level includes: Based on the predicted interest levels from high to low, a preset number of non-interactive land parcels are selected to generate a recommendation list; or non-interactive land parcels with predicted interest levels exceeding a preset threshold are marked on a GIS map for display.

[0025] This application generates a recommendation list by selecting a preset number of non-interactive land parcels in descending order of predicted interest. This allows demanders to prioritize a small number of high-quality parcels that best match their strategic intentions and preferences from a vast amount of land parcel information, significantly reducing manual screening costs and time investment, and improving the efficiency of investment and development decisions. In another implementation, non-interactive land parcels with predicted interest exceeding a preset threshold are marked on a GIS map for display. This allows the recommendation results to be intuitively combined with geospatial information. Demanders can quickly identify the concentrated distribution of high-quality land parcels within the target area from a spatial perspective. Simultaneously, by combining heat maps and other visualization methods, they can perceive regional heat and provide intuitive spatial reference for site selection decisions. These two recommendation generation methods respectively meet the different use case requirements of quantitative sorting and screening and intuitive spatial positioning, improving the usability of the recommendation results and user experience. Attached Figure Description

[0026] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a flowchart of an embodiment of the intelligent matching method for land resource supply and demand provided in this application; Figure 2 This is a flowchart illustrating the prediction of interest determination according to an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0030] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0031] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0032] In the field of land resource transactions, land transactions are generally characterized by large single-transaction amounts, long corporate decision-making chains, and cyclical land acquisition behavior. Whether it's real estate development land, industrial land, or commercial service land, the land acquisition behavior of demanders requires a lengthy internal evaluation and decision-making process, and different companies differ in their capital turnover speed and land acquisition frequency. Existing land information service platforms, when introducing collaborative filtering recommendation mechanisms, typically use a globally uniform time decay function to handle the timeliness of user preferences. However, this approach ignores the differences in strategic depth between different interactive behaviors and the differences in decision-making cycles among different demanders. This results in the short-term agility needs of companies with fast capital turnover being submerged in outdated data, while the long-term strategic intentions of companies that acquire land prudently are forgotten due to rapid decay, leading to a systematic deviation between the recommendation results and the users' true strategic intentions. To address this, this application provides an intelligent matching method for land resource supply and demand.

[0033] Please refer to Figure 1 , Figure 1 This is a flowchart of an embodiment of the intelligent matching method for land resource supply and demand provided in this application. The method includes steps S1 to S5, each step of which is detailed below: Step S1: Obtain the interaction behavior data of the demander on the land parcel, the demander's historical land acquisition records, the demander's basic information, and the preset behavior type depth weight; wherein, the interaction behavior data includes the behavior type and the time of the behavior occurrence; the preset behavior type depth weight represents the difference in the intensity of preference indication of different interaction behavior types; Step S2: Determine the decision cycle benchmark value of the demander based on historical land acquisition records; based on the decision cycle benchmark value, the preset behavior type depth weight of the corresponding behavior type and the behavior occurrence time, perform time decay processing on the preference score corresponding to the interaction behavior data to obtain the decayed preference score. Step S3: Based on the basic information of the demander, the demander's historical land acquisition records and interaction behavior data, determine the similarity between the demander and other demanders; Step S4: Determine the predicted interest of demanders in non-interacting land parcels based on similarity scores, decay preference scores of other demanders, and comprehensive similarity scores between land parcels. Step S5: Generate land parcel recommendation results based on predicted interest levels.

[0034] The interactive behavior data includes behavior type and behavior occurrence time. Behavior type refers to various interactive operations performed by the demander on the platform regarding land parcels, including explicit and implicit interactive behaviors. Explicit interactive behaviors are actions by the demander actively expressing investment intentions, including at least one of bidding, inquiring, and adding to favorites; implicit interactive behaviors are the behavioral patterns generated by the demander during browsing, including at least one of page browsing, repeated visits, and browsing duration. Behavior occurrence time is the specific point in time when the demander performs the above interactive operations, accurate to the date.

[0035] The demander's historical land acquisition records are the historical transaction data of the land parcels successfully won or acquired by the demander in the past, including at least the transaction date of each land acquisition, used to calculate the average land acquisition interval for the demander. The demander's basic information includes corporate asset information, main business type, intended development area, and other registration information and self-reported data, used to construct a static profile of the demander.

[0036] Preset depth weights for different behavior types are used to characterize the differences in the strength of preference indications conveyed by various interaction types. The more explicit the investment intention reflected by the interaction, the greater its corresponding depth weight. For example, the depth weight of bidding behavior is greater than that of consultation behavior, and the depth weight of consultation behavior is greater than that of browsing behavior.

[0037] It should be noted that although some embodiments of this application use real estate land parcels as the main example scenario, its core personalized time decay mechanism is also applicable to supply and demand matching scenarios for other types of land resources such as industrial land, commercial service land, and warehousing and logistics land. The transactions of the aforementioned land resources are also characterized by large sums of money, long cycles, and heavy decision-making, with the demand side's land acquisition behavior exhibiting significant periodicity. When applied to different types of land resources, those skilled in the art can adaptively adjust the specific values ​​of the preset behavior type depth weights, the category labels of business preferences in the profile attributes, and the numerical attribute set in the land parcel attribute similarity calculation, according to the specific transaction characteristics of the land type, thereby achieving intelligent supply and demand matching in different land transaction scenarios.

[0038] In the specific implementation, in step S1, the interaction behavior data of the demander with the land plot, the historical land acquisition records of the demander, the basic information of the demander, and the preset behavior type depth weight are obtained to construct a multi-dimensional profile system of the demander, which provides a multi-source data foundation for subsequent personalized preference modeling and collaborative filtering recommendation. Its core processing process includes two sub-steps: multi-source data collection and data standardization and vectorization processing.

[0039] Regarding demand-side data collection, raw data directly related to demanders is aggregated from multiple data sources. This data originates from enterprise registration information, published demand information, and historical transaction records on the platform. Specifically, it covers basic enterprise information, business preferences, asset status, intended regions, and historical land acquisition records. Basic information includes registration data such as registered capital and enterprise type; business preferences refer to the types of products the enterprise primarily develops; for example, real estate development enterprises may include residential, commercial, and office real estate types, while industrial enterprises may include manufacturing, processing, warehousing, and logistics types; intended regions refer to the administrative divisions of cities the enterprise is interested in or plans to enter; published demand information includes the enterprise's self-reported proposed development scale and budget range; historical land acquisition records include the transaction dates, locations, and transaction prices of all successfully acquired or bid-for land parcels. This data is the core basis for subsequently estimating the time interval for demanders to acquire land and determining the benchmark value for the decision-making cycle.

[0040] Regarding interactive behavior data, the platform collects the behavioral trajectories of potential buyers regarding land parcels through data tracking. Key fields include behavior type and behavior occurrence time. Behavior type encompasses two main categories: explicit and implicit interactions. Explicit interactions refer to actions taken by potential buyers to actively express investment intentions, including at least one of bidding, inquiring, or adding to favorites. Implicit interactions refer to the behavioral trajectories generated by potential buyers during browsing, including at least one of page browsing, repeated visits, and browsing duration. The behavior occurrence time is the specific point in time when the potential buyer performs the aforementioned interactive actions, accurate to the date, providing a time anchor for subsequent timeliness modeling.

[0041] It should be noted that this application will also incorporate supplier-side land parcel data and third-party data for collaborative calculations in subsequent steps. The supplier-side land parcel data originates from a land resource pool and includes basic attributes such as parcel location, area, planned use, plot ratio, starting price, building height limit, and the coordinates of the parcel's center point. The third-party data includes vector data of supporting facilities such as hospitals, schools, and subway stations, as well as population grid data used to assess regional population coverage. The scope of the aforementioned land parcel attribute data can be flexibly adjusted according to different land resource types. For example, industrial land can include attributes such as factory area and industry category, while commercial land can include attributes such as business district level and pedestrian traffic.

[0042] Regarding data standardization and vectorization, to eliminate format heterogeneity among data from different sources, the original data collected from the requesting parties undergoes unified standardization processing. For tabular data, data processing tools are used to convert it into N-dimensional vectors; for vector data, after completing relevant business overlay operations, its attribute table is exported and vectorized using the same processing method to ensure homogeneity and consistency across data dimensions, providing a standardized input format for subsequent user profile tag modeling, preference score construction, and similarity calculation.

[0043] Specifically, in some embodiments, in step S2, the decision-making cycle benchmark value of the demander is determined based on historical land acquisition records, including: Extract the historical land acquisition time of the demand side to determine the time interval between adjacent land acquisition activities; The average land acquisition time interval is obtained by averaging the time intervals. The average land acquisition time interval is normalized to obtain the decision cycle benchmark value.

[0044] The average land acquisition interval reflects the typical time span from one land acquisition to the next in the history of this demander, and is an absolute time quantity in days or months. Normalization aims to map the absolute time spans of different enterprises to a unified numerical range, eliminating the incomparability caused by differences in absolute time magnitudes between enterprises. This allows enterprises with significantly different capital turnover cycles to perform subsequent attenuation parameter calculations under a unified dimension. The decision cycle benchmark value is the dimensionless value obtained after normalization, used to quantitatively characterize the land acquisition frequency characteristics of this demander: a smaller value indicates more frequent land acquisition and a shorter decision cycle; a larger value indicates more cautious land acquisition and a longer decision cycle.

[0045] Specifically, traditional recommendation algorithms typically employ a globally uniform exponential decay function when dealing with the timeliness of user preferences, applying the same forgetting rate to all users and all interactions. However, in land resource transactions, land recommendations exhibit both strong timeliness and long cycles. Different demanders have varying capital turnover cycles; companies with fast capital turnover and large project reserves make decisions quickly and acquire land in short intervals, while companies that are cautious in land acquisition and have longer turnover cycles acquire land in longer intervals. A globally uniform decay function treats all companies equally, failing to differentiate between these rhythmic differences, resulting in a disconnect between the decay rate and the actual decision-making cycle of each company. Therefore, this application extracts benchmark values ​​for decision-making cycles from the historical land acquisition records of demanders to quantitatively characterize the land acquisition rhythm characteristics of different companies.

[0046] In its implementation, the process begins by extracting the transaction dates of each land acquisition from the client's historical land acquisition records and calculating the time interval between adjacent acquisitions in chronological order. By averaging all time intervals, the dispersed land acquisition activities of enterprises are transformed into a stable statistical value, effectively eliminating the interference of single abnormal transactions on the decision-making cycle assessment. This allows for the accurate quantification of the differentiated rhythms of high-frequency land acquisition by enterprises with fast capital turnover and low-frequency land acquisition by enterprises with prudent land acquisition practices. Based on this, the average land acquisition time interval is normalized, mapping the absolute time span of different enterprises to a unified dimension, thus obtaining a benchmark value for the decision-making cycle. Normalization makes enterprises of different sizes comparable, providing a quantitative numerical basis for the subsequent differentiated calculation of decay half-life. Through the above processing, the differences in enterprises' land acquisition rhythms are transformed into a numerical parameter that can be directly used by the algorithm, fundamentally solving the problem that traditional uniform decay functions cannot adapt to the differences in decision-making cycles among different enterprises.

[0047] Specifically, in some embodiments, in step S2, based on the decision cycle baseline value, the preset behavior type depth weight corresponding to the behavior type, and the behavior occurrence time, the preference score corresponding to the interaction behavior data is subjected to time decay processing to obtain a decayed preference score, including: The adaptive decay half-life is determined based on the decision cycle baseline value, the depth-sensitive adjustment parameter, and the preset behavior type depth weight of the corresponding behavior type. The higher the preset behavior type depth weight, the longer the corresponding half-life. The depth-sensitive adjustment parameter is used to control the degree of influence of the preset behavior type depth weight on the half-life. The time interval since the action is committed is determined by the difference between the time the action occurred and the current time. By using an adaptive decay half-life and the time interval between behaviors, the preference scores corresponding to the interactive behavior data are subjected to time decay processing to obtain decayed preference scores.

[0048] The depth-sensitive adjustment parameter is a preset global adjustment coefficient used to control the impact of preset behavior type depth weights on the decay half-life. A larger value for the depth-sensitive adjustment parameter results in more significant differences in the decay half-life between different behavior types, and a stronger ability to distinguish behavior depth. The adaptive decay half-life is a differentiated parameter generated by combining the enterprise's personalized decision-making cycle benchmark value with the behavior type depth weights: enterprises with shorter decision-making cycle benchmark values ​​have generally shorter decay half-lives, and preference scores are forgotten more quickly; behavior types with higher depth weights have relatively longer decay half-lives, and corresponding preference scores are retained more persistently. Through this adaptive mechanism, different enterprises and different behavior types obtain decay speeds that match their characteristics, achieving personalized management of preference timeliness. The core logic of time decay processing is that the ratio of the time interval between the behavior and the adaptive decay half-life is used as the decay basis. The longer the behavior is in the past, the more obvious the decay; the shorter the half-life, the faster the decay, thus ensuring that recent high-depth behaviors are remembered, while distant shallow behaviors are decisively forgotten.

[0049] In the specific implementation, the baseline value of the decision cycle is first obtained. Combined with preset behavior type depth weights and preset global depth-sensitive adjustment parameters For specific demanders and specific interaction types, the adaptive decay half-life is calculated using the following formula. : in, For the demand side The baseline value of the decision-making cycle, For the first Preset depth weights corresponding to different interaction behavior types This is a depth-sensitive adjustment parameter used to control... The degree of impact on half-life. The higher the preset depth weight of behavior type, the longer the corresponding adaptive decay half-life; the shorter the decision cycle benchmark value of enterprises, the shorter their adaptive decay half-life is overall. Through this combined calculation, the shallow browsing behavior of enterprises with fast capital turnover obtains a shorter half-life, and the preference score is quickly forgotten to agilely capture their short-term needs; the deep bidding behavior of enterprises that are cautious in acquiring land obtains a longer half-life, and the preference score is retained for a long time to lock in their strategic intentions.

[0050] Next, based on the difference between the time of the action and the current time in the interaction data, the time interval since the action was performed is calculated. . The larger the value, the further back in time the behavior occurred, and the lower its reference value for current preferences. An adaptive decay half-life is used. Time interval between behaviors The preference scores corresponding to the interaction behavior data are subjected to time decay processing. Specifically, using... and The ratio of the original preference score to the actual preference score is used as the criterion for decay; the older the behavior, the more pronounced the decay; the shorter the half-life, the faster the decay. The processing employs an exponential decay function to reduce the original preference score. Multiply by the time decay factor and apply to the demand side. land parcel The summation of all generated interactions yields the overall preference score after time decay. The calculation formula is: in, For the demand side land parcel The set of all generated interactive behaviors, For the demand side, the first The original preference score corresponding to this behavior. For the first The time interval since this behavior occurred. For the first The adaptive decay half-life corresponds to each behavior. Through the above dynamic time decay mechanism, different enterprises and different behavior types obtain a forgetting rate that matches their characteristics. Recent high-depth behaviors are remembered, while distant shallow behaviors are decisively forgotten, thereby improving the accuracy of preference prediction in long-term, high-decision scenarios.

[0051] It should be noted that the raw preference scores involved in the above formulas... The data originates from the interactive behavior data collected in step S1. The original preference scores of the demanders for the land parcels are determined from both explicit and implicit feedback. Explicit feedback reflects the investment intentions actively expressed by the demanders, and is assigned a fixed score based on the behavior type; for example, bidding is worth 1.0 point, consultation is worth 0.8 points, and adding to favorites is worth 0.6 points. Implicit feedback reflects the interest signals implicitly conveyed by the demanders during browsing, and is accumulated according to the quantitative rules of the behavioral indicators; for example, browsing time exceeding a preset threshold is worth 0.2 points, and each repeated visit to the same land parcel adds 0.1 points, with a cumulative maximum of 0.5 points. These two types of scores together constitute the demanders'... Original preference score matrix for land parcels Each score record in the matrix is ​​associated with a timestamp attribute of the corresponding interaction behavior, i.e., the specific date the behavior occurred. When time decay processing is required for a particular interaction behavior, it is performed based on the type of the behavior, starting from the original preference score matrix. Extract the corresponding raw score Simultaneously, the time interval between the behavior and the present is calculated based on the timestamp associated with the behavior. Substituting the two values ​​above into the aforementioned time decay weighted formula, the decay calculation for the preference score is completed. Specifically, in some embodiments, in step S3, the similarity between the demander and other demanders is determined based on the demander's basic information, historical land acquisition records, and interaction behavior data. This includes steps S31 to S35, each step of which is as follows: Step S31: Extract profile attributes from the basic information of the demander and the demander's historical land acquisition records, and construct a profile feature vector; Step S32: Extract explicit interaction behaviors from the interaction behavior data and construct explicit behavior feature vectors; Step S33: Extract implicit interaction behaviors from the interaction behavior data and construct implicit behavior feature vectors; Step S34: Calculate the similarity of portrait features, explicit behavioral features, and implicit behavioral features respectively; Step S35: Assign adjustable weights to the similarity of portrait features, explicit behavioral features, and implicit behavioral features; sum the similarity of portrait features, explicit behavioral features, and implicit behavioral features by weight to obtain the similarity score.

[0052] Among them, profile feature similarity is a similarity value obtained by cosine similarity calculation based on profile feature vectors, reflecting the degree of similarity between demanders in static attributes such as enterprise assets and business types. Explicit behavioral feature similarity is a similarity value obtained by cosine similarity calculation based on explicit behavioral feature vectors, reflecting the degree of similarity between demanders in terms of land preferences. Implicit behavioral feature similarity is a similarity value obtained by cosine similarity calculation based on implicit behavioral feature vectors, reflecting the deep similarity between demanders in terms of browsing behavior habits. Adjustable weights include... , , The contribution ratios of explicit behavioral similarity, portrait feature similarity, and implicit behavioral similarity to the overall similarity are controlled separately to satisfy... It can be dynamically adjusted through Bayesian optimization, so that the contribution ratio of each dimension can be adaptively optimized according to the actual recommendation effect.

[0053] Specifically, in some embodiments, in step S31, profile attributes are extracted from the basic information of the demander and the demander's historical land acquisition records to construct a profile feature vector, including: The profile attributes are extracted from the basic information of the demanders and their historical land acquisition records; among them, the profile attributes include at least one of the following: enterprise asset status, business type, and intended region; The extracted portrait attributes are converted into label vectors to obtain portrait features.

[0054] Among them, the profile attributes are features extracted from the basic information and historical land acquisition records of the demand side, reflecting the long-term investment attributes of the enterprise, including at least one of the enterprise's asset status, business type, and intended region. The profile feature vector is a numerical vector generated after transforming the extracted profile attributes. The transformation method is as follows: all possible profile attributes are integrated into a label vocabulary, and One-Hot encoding is used to convert the extracted profile attributes into label vectors. If the demand side belongs to a certain type, the corresponding dimension is set to 1; otherwise, it is 0. Thus, each demand side is mapped to a profile feature vector consisting of 0s and 1s.

[0055] Specifically, in some embodiments, in steps S32 and S33, explicit interactive behaviors include at least one of bidding, consultation, and favorites; implicit interactive behaviors include at least one of page browsing, repeated visits, and browsing duration.

[0056] Explicit interactive behavior refers to actions taken by the demand side to actively express investment intentions, including at least one of bidding, consulting, and adding to favorites. The explicit behavior feature vector is a vector constructed based on the demand side's explicit preference score for the land parcel, reflecting the investment intentions explicitly expressed by the demand side through the aforementioned proactive actions. Implicit interactive behavior refers to the behavioral trajectory generated by the demand side during browsing, including at least one of page browsing, repeated visits, and browsing duration. The implicit behavior feature vector is a vector constructed based on the demand side's implicit preference score for the land parcel, used to capture the deep-seated characteristics of the interest signals and behavioral habits implicit in the demand side's aforementioned browsing behaviors.

[0057] In the specific implementation, when constructing the profile feature vector, profile attributes are extracted from the basic information and historical land acquisition records of the demander. These profile attributes include at least one of the following: enterprise asset status, business type, and intended region. All possible profile attributes are integrated into a tag vocabulary, and the extracted profile attributes are converted into tag vectors using One-Hot encoding. For example, the tag vocabulary includes tags such as "Assets_Large Enterprises," "Assets_Small Enterprises," "Business_Residential," "Business_Commercial," "Intended Region_Haizhu District," and "Intended Region_Tianhe District." If the demander belongs to this type, the corresponding dimension has a value of 1; otherwise, it is 0. When the solution is applied to different types of land resources, the business preference categories in the profile attributes can be adaptively adjusted according to the actual scenario. For example, in an industrial land scenario, business types such as manufacturing and processing, warehousing and logistics can be added; in a commercial land scenario, business types such as retail, hotels, and offices can be added. Thus, each demander is mapped to a profile feature vector composed of 0s and 1s, realizing the conversion of non-numerical categorical attributes into structured numerical vectors.

[0058] When constructing explicit behavior feature vectors, explicit interactive behaviors of the demand side are extracted from the interaction behavior data, including at least one of bidding, consultation, and collection. The explicit behavior feature vectors are constructed based on the demand side's explicit preference scores for land parcels, using the explicit preference scores of the demand side for each interacting land parcel as vector elements, reflecting the investment intentions explicitly expressed by the demand side through active operations. When constructing implicit behavior feature vectors, implicit interactive behaviors of the demand side are extracted from the interaction behavior data, including at least one of page browsing, repeated visits, and browsing duration. The implicit behavior feature vectors are constructed based on the demand side's implicit preference scores for land parcels, using the implicit preference scores of the demand side for each interacting land parcel as vector elements, used to capture the deep features of the interest signals and behavioral habits implicitly contained by the demand side during browsing. Both the explicit and implicit behavior feature vectors originate from the original preference score matrix. The former extracts the score column corresponding to explicit feedback, while the latter extracts the score column corresponding to implicit feedback.

[0059] Land transactions are characterized by large sums of money, long cycles, and heavy decision-making, making it difficult to comprehensively depict the similarity between demanders using information from a single dimension. Therefore, a multi-dimensional weighted similarity model is constructed, calculating similarity from three dimensions: profile features, explicit behavioral features, and implicit behavioral features. These are then weighted and fused using adjustable weights to obtain a quantitative index that comprehensively reflects the similarity between demanders. Cosine similarity is used in calculating the similarity of each of the three feature types. Cosine similarity measures the similarity between two vectors by calculating the cosine of the angle between them, with a value ranging from [0,1]. A value closer to 1 indicates greater similarity between the two demanders in that feature dimension. For demanders... and demand side Let the characteristics of the two be respectively =[ , ..., ]and =[ , ..., The formula for calculating cosine similarity is: The numerator is the dot product of the two vectors, and the denominator is the product of the magnitudes of the two vectors. Based on the above formula, the image feature similarity is calculated. Explicit behavioral feature similarity and implicit behavioral feature similarity After obtaining the similarity of the above three categories, the following multi-dimensional weighted fusion formula is used to calculate the demand side. and Overall similarity between : in, The similarity is based on explicit interaction behavior vectors, reflecting the degree of similarity between two demanders in their land preferences; Based on the similarity of profile feature vectors, it reflects the degree of similarity between two demanders in static attributes such as corporate assets and business types; The similarity is based on implicit interaction behavior vectors, reflecting the deep similarity between two users in their browsing habits. , , The adjustable weights of the above three are respectively, satisfying The aforementioned weights can be dynamically adjusted through Bayesian optimization, allowing the contribution ratio of each dimension to be adaptively optimized based on the actual recommendation effect.

[0060] Specifically, in some embodiments, please refer to Figure 2 , Figure 2 This is a flowchart of a predicted interest determination embodiment provided in this application. In step S4, the predicted interest of the demander in the non-interacting land parcels is determined based on the similarity, the decay preference score of other demanders, and the comprehensive similarity between land parcels. This includes steps S41 to S43, each of which is as follows: Step S41: Obtain the first predicted interest level based on similarity and decay preference scores with other demanders; Step S42: Obtain the second predicted interest degree based on the similarity matching algorithm of the land parcels; Step S43: The first predicted interest degree and the second predicted interest degree are weighted and fused to obtain the predicted interest degree; wherein, the weight of the weighted fusion is dynamically adjusted according to historical transaction data.

[0061] Weighted fusion refers to a hybrid recommendation strategy that linearly combines the first and second predicted interest levels using an adjustable weight λ. The first predicted interest level is calculated based on the similarity between demanders, its core logic being collaborative filtering based on the idea that "similar companies have similar preferences," using the historical preferences of other similar demanders as the basis for prediction. The second predicted interest level is calculated based on the comprehensive similarity between land parcels, its core logic being content matching based on the idea that "similar land parcels have similar value," extending the demander's own interest in historically preferred land parcels to other similar land parcels. Through this weighted fusion, the advantages of both recommendation logics can be combined, compensating for the information blind spots of a single recommendation source. The fusion weight λ ranges from [0,1]. When λ approaches 1, the prediction result is more biased towards the recommendation opinion based on demander similarity; when λ approaches 0, the prediction result is more biased towards the recommendation opinion based on land parcel similarity.

[0062] Dynamic adjustment of historical transaction data means that the aforementioned fusion weight λ is not a fixed value, but rather is dynamically optimized through online learning based on historical transaction data from different recommendation channels. Specifically, it continuously tracks the transaction records ultimately achieved by the demand side. If more transactions originate from recommendation channels based on demand side similarity, the value of λ is gradually increased; conversely, if more transactions originate from recommendation channels based on land parcel similarity, the value of λ is gradually decreased. Through this adaptive adjustment mechanism, it can automatically and in real-time tilt towards recommendation paths with better actual transaction results, thereby continuously improving the alignment between recommendation results and actual transaction intentions.

[0063] Specifically, in some embodiments, in step S41, a first predicted interest level is obtained based on similarity and decay preference scores with other demanders, including: Using similarity as a weight, the decay preference scores of other demanders are summed in a weighted manner to obtain a weighted preference score. The similarity normalization factor is obtained by summing the absolute values ​​of the similarities. The weighted preference score is normalized using a similarity normalization factor to obtain the first predicted interest level.

[0064] The weighted preference score is an intermediate result obtained by weighting and summing the attenuation preference scores of other demanders for the target plot, using the similarity value between the target demander and all other demanders as weighting coefficients. The other demanders involved in the weighted calculation are the neighboring users. After calculating the comprehensive similarity between the target demander and all other demanders, a preset number of demanders with the highest similarity are selected to form the neighbor set. The similarity normalization factor is a normalized denominator obtained by summing the absolute similarity values ​​of all neighboring users in the neighbor set. Its function is to eliminate the influence of differences in the size of the neighbor sets of different demanders on the prediction results, making the prediction results comparable for demanders with a large number of neighbors and those with a small number of neighbors. The first predicted interest score is a predicted score obtained based on the collaborative filtering path. It is normalized by dividing the weighted preference score by the similarity normalization factor, reflecting the potential interest of the target demander in non-interacting plots inferred from the historical preferences of similar demanders.

[0065] Specifically, in some embodiments, in step S42, obtaining the second predicted interest degree according to the parcel similarity matching algorithm includes: Obtain data on supporting facilities around the land parcel, and determine the scarcity weight of each type of facility based on the global distribution density of the supporting facility data; The spatial potential energy value of the land parcel is calculated based on the scarcity weight and the synergistic effect between facilities. Determine the spatial similarity bandwidth based on the spatial potential energy value; The spatial similarity between land parcels is determined based on the spatial similarity bandwidth. Determine the similarity of land parcel attributes based on the numerical attributes of the land parcels; The spatial similarity and the similarity of land parcel attributes are weighted and fused to obtain the comprehensive similarity of land parcels; The second predicted interest level is determined based on the overall similarity of the land parcels and the historical preferences of the demand side.

[0066] Among them, supporting facility data refers to the spatial location information of various public service facilities around the plot, which comes from third-party vector data, including facility categories such as hospitals, schools, and subway stations. Global distribution density refers to the distribution density of a certain type of facility within the entire urban space; the lower the distribution density, the scarcer the facility. Scarcity weight is the facility value weight quantified through an inverse density scarcity factor. Its basic rule is: the sparser the distribution of a scarce facility, the greater its scarcity weight, and the higher its contribution to the spatial value of the plot.

[0067] The synergistic effect between facilities refers to the non-linear added value generated when multiple types of core facilities cluster together around the same plot of land. It is quantified through cross-product terms. A subset of core facilities exemplarily includes the combination of subway stations and large commercial districts. Spatial potential value is a quantitative indicator of the locational value of a plot of land, which integrates the scarcity weights of various facilities, the influence of distance attenuation, and the synergistic effect of core facilities. It is used to measure the comprehensive endowment of a plot of land in the spatial dimension.

[0068] Spatial similarity bandwidth refers to the parameter in the Gaussian kernel function used to control the rate of distance similarity decay. It is adaptively determined by the normalized spatial potential energy value: the higher the spatial potential energy value, the smaller the bandwidth, and the more sensitive the spatial similarity is to distance changes; the lower the spatial potential energy value, the larger the bandwidth, and the more tolerant the spatial similarity is to distance changes. Spatial similarity is based on adaptive spatial similarity bandwidth, using joint bandwidth and the Gaussian kernel function to convert the Euclidean distance between the center points of land parcels into a similarity metric.

[0069] Land parcel attribute similarity is a similarity measure of attribute dimensions calculated based on the numerical attributes of the land parcels. These numerical attributes include at least one of the following: area, plot ratio, starting bid price, and building height limit. The similarity measure is calculated using weighted normalized Euclidean distance, and the weights of each attribute are determined using the entropy weight method. When the scheme is applied to different types of land resources, the selection of numerical attributes can be adaptively adjusted according to the actual scenario. For example, for industrial land, factory area indicators can be added, and for commercial land, business district level or pedestrian traffic indicators can be added.

[0070] Weighted fusion refers to the linear combination of spatial similarity and parcel attribute similarity using adjustable weights, which can be dynamically adjusted based on users' historical filtering behavior. The overall parcel similarity is the overall similarity index between parcels obtained after weighted fusion of spatial similarity and parcel attribute similarity. Historically preferred parcels refer to the set of highly preferred parcels that demanders have historically interacted with, serving as the reference basis for calculating the second predicted interest score. The second predicted interest score is a predicted score obtained based on the parcel similarity matching path. It is calculated by combining the overall similarity between candidate parcels and historically preferred parcels by demanders, and weighted by attenuated preference scores, reflecting the potential interest of demanders in uninterrupted parcels inferred from the objective attributes and spatial value dimensions of the parcels.

[0071] In its implementation, step S4 employs a hybrid recommendation strategy to determine the predicted interest of demanders in non-interacting land parcels. This involves independently calculating predicted scores through collaborative filtering based on demander similarity and content matching based on land parcel similarity, then weighted and fused together. Traditional single-path recommendation algorithms have inherent limitations: relying solely on collaborative filtering significantly reduces prediction reliability if historical data on similar demanders is sparse; relying solely on content matching completely severs the preference transmission relationship between demanders. Step S4, through dual-path fusion, preserves the collective wisdom of converging preferences among similar enterprises while also considering the objective assessment of land parcel attributes and spatial value. Furthermore, it dynamically adjusts the fusion weights using historical transaction data, adaptively seeking an optimal balance between the two types of logic. The specific implementation of this hybrid recommendation strategy includes the following steps.

[0072] In step S41, a first predicted interest level based on demand-side similarity is calculated. Following the collaborative recommendation logic that "similar enterprises have similar preferences," land parcels that the target demander might be interested in are discovered from the historical behavior of similar demanders. The calculated demanders... After considering the overall similarity with other demanders, the results are sorted from highest to lowest similarity, and the highest similarity is selected. Each demander constitutes a set of neighboring users. .

[0073] Subsequently, the demand side With neighboring users Overall similarity between As a weight, for each neighboring user land parcel Decay preference score A weighted sum is performed to obtain a weighted preference score. Next, the absolute similarity values ​​of all neighboring users are summed to obtain a similarity normalization factor. Finally, this normalization factor is used to normalize the weighted preference score, yielding the first predicted interest level. The calculation formula is as follows: in, In order to meet with the demand side The set of neighboring users with the highest similarity; For the demand side With neighboring users Overall similarity between them; For neighboring users land parcel Decaying preference score. By using similarity as a weight, the more similar the neighboring user to the target demander, the greater their contribution to the prediction results; through normalization, the impact of differences in the size of the neighboring sets of different demanders on the comparability of prediction results is eliminated.

[0074] In step S42, a second predicted interest degree based on parcel similarity matching is calculated. Starting from the parcel's own attributes and spatial value dimensions, other parcels similar to those historically preferred by the target demander are recommended. Existing spatial similarity calculation methods often only consider absolute Euclidean distance, neglecting the scarcity of supporting facilities and the synergistic effect between facilities. Therefore, step S42 constructs a multi-source facility synergy and scarcity potential field model to dynamically solve for the adaptive spatial similarity bandwidth.

[0075] Specifically, prior to calculating land parcel similarity matching, land parcel attribute modeling and data preprocessing were completed. Specifically, numerical attributes (such as area, plot ratio, starting price, building height limit, etc.) were selected for attribute similarity matching, and spatial attributes were selected for spatial similarity matching. For generating spatial distance attributes, third-party vector data (such as hospitals, schools, subway stations, etc.) was spatially overlaid with the land parcel vector data to calculate the Euclidean distance from the center point of the land parcel to each surrounding facility, and this distance value was assigned to the corresponding land parcel's vector data. Simultaneously, third-party population grid data was overlaid with the land parcel vector data. For each grid covered or involved by a land parcel, the average population value was taken, and this population number was assigned to the corresponding land parcel as an auxiliary economic indicator for subsequent land parcel value assessment. The above preprocessing work provides a standardized data foundation for subsequent spatial and attribute similarity calculations.

[0076] Subsequently, data on supporting facilities surrounding the site were acquired, and the scarcity weights of various facilities were quantified. The supporting facility data originated from third-party vector data, including spatial location information of facilities such as hospitals, schools, and subway stations. A mechanism similar to Term Frequency-Inverse Document Frequency (TF-IDF) was introduced to calculate the global distribution density of facility category c across the entire urban space. And its scarcity weight is defined as the inverse density scarcity factor. : Among them, global density The lower the value, the scarcer this type of facility is; its inverse density scarcity factor... The larger the plot, the higher its contribution to the spatial value of the land.

[0077] Next, calculate the land parcel. Influenced by the local distance attenuation of its surrounding Category C facilities : in, For the plot of land To the The actual Euclidean distance of each Class C facility This is the attenuation coefficient of the impact of this type of facility. For the plot of land The number of surrounding Class C facilities. This formula means that facilities closer to the site contribute more, and their influence decreases exponentially with distance.

[0078] Introducing cross-product terms to characterize the synergistic effects between different facilities, and constructing plots Comprehensive potential energy : in, For all third-party facility categories, A subset of core facilities with strong synergistic effects, exemplarily including a combination of subway stations and large shopping malls; This is the synergistic effect amplification factor. The cross-product term explicitly quantifies the nonlinear value-added effect generated when the core facilities coexist.

[0079] Determining the adaptive spatial similarity bandwidth based on comprehensive energy, using comprehensive energy Perform normalization, denoted as Substituting this into the bandwidth calculation of the Gaussian kernel function: in, Based on bandwidth, This is a sensitivity parameter used to control the degree to which spatial potential energy affects bandwidth. Through this adaptive mechanism, high-potential-energy, busy areas receive smaller bandwidth and are more sensitive to changes in spatial distance; while low-potential-energy, remote areas receive larger bandwidth and are more tolerant of changes in spatial distance.

[0080] After determining the adaptive bandwidth, the plot is calculated using the joint bandwidth and Gaussian kernel function. With the plot of land Spatial similarity between : in, Indicates land parcel With the plot of land The Euclidean distance between the center points. This refers to bandwidth parameters, which can be set differently for different city scales or different business applications. Values, for example, can be set to smaller values ​​for land use types that have a significant impact on location value. To enhance sensitivity to distance, a larger value can be set for land use types that are not sensitive to spatial distance. This is to increase tolerance for distance.

[0081] Secondly, based on the numerical attributes of the land parcels, weighted normalized Euclidean distance is used to calculate the land parcels. With the plot of land Attribute similarity between : in, This is the set of all numerical attributes, including area, floor area ratio, starting bid price, building height limit, etc. and plots of land With the plot of land In attributes The value of ; and Attributes Maximum and minimum values ​​among all plots; For attributes The weights are determined using the entropy weight method.

[0082] The spatial similarity and attribute similarity are weighted and fused to obtain the comprehensive similarity of the land parcels. : in, and For adjustable weights, satisfying This system dynamically adjusts based on the filtering actions of demanders using historical behavioral data, achieving online adaptive optimization. For example, if demanders click on filtering by spatial criteria more frequently, the spatial similarity weight is increased. If the client clicks to filter by attribute criteria more frequently, then increase the weight of attribute similarity. .by For example, the adjustment method is as follows: in, This indicates the number of clicks made by the demand side when filtering by attribute criteria. This indicates the number of clicks made by the demand side when filtering by spatial criteria. As a smoothing parameter, it is usually set to 1 to prevent the weight from being 0 when it has never been filtered (0 times).

[0083] Finally, obtain the demand side. Set of high-preference land parcels with historical interactions For each candidate plot Based on the comprehensive similarity of land parcels Demand-side preference score for historical land parcels Calculate the second predicted interest level : In step S3, the first predicted interest level is... Compared with the second predicted interest Perform linear weighted fusion to obtain the demand side For non-interactive plots Comprehensive predicted interest : in, The weights are adjustable, and their values ​​range from [value range missing]. Based on the recommendation results, adjustments and optimizations are dynamically made through online learning. This is based on the needs of the user. weight For example: in, Indicates demand side The number of transactions achieved through recommendation channels based on demand-side similarity. Indicates demand side The number of transactions achieved through recommendation channels based on land parcel similarity matching. To smooth out the parameters, the aforementioned dynamic adjustment mechanism assigns higher integration weights to recommendation channels with higher transaction volumes, thus adaptively favoring recommendation paths with better actual transaction results.

[0084] It should be noted that after generating the candidate recommendation set, a rule-based constraint filtering mechanism can be introduced for further screening and re-ranking. Specifically, in some embodiments, rule-based constraint filtering includes at least one of the following: policy constraint filtering, used to remove plots that do not comply with urban planning or land use restrictions, ensuring that the recommendation results comply with policy and regulatory requirements; and enterprise capability filtering, used to filter land beyond the financial capacity of the demander, ensuring that the recommendation results match the actual development capacity of the demander. All recommendation results undergo filtering processing by a unified rule engine before being output to the application layer, ensuring compliance and feasibility of the recommendation results while achieving personalized recommendations.

[0085] Specifically, in some embodiments, in step S5, generating land parcel recommendation results based on predicted interest includes: Based on the predicted interest level from high to low, a preset number of non-interactive land parcels are selected to generate a recommendation list; or non-interactive land parcels with predicted interest levels exceeding a preset threshold are marked on a GIS map for display.

[0086] In practice, generating land parcel recommendation results based on predicted interest levels includes at least one of the following two optional methods: Method 1: Based on the predicted interest level from high to low, a preset number of non-interactive land parcels are selected to generate a recommendation list. This recommendation list is presented to the demand side in the form of a sorted list. The demand side can browse the high-quality land parcels selected by the system in order, quickly focusing on a small number of land parcels that best match their strategic intentions and preferences, significantly reducing the manual cost and time investment of sifting through massive amounts of land parcel information one by one.

[0087] Method 2: Mark non-interactive plots with predicted interest exceeding a preset threshold on a GIS map for display. This map marking method intuitively combines recommendation results with geospatial information, allowing potential buyers to quickly identify the concentrated distribution of high-quality plots within their target area. Simultaneously, heat maps can be used to reflect areas of high plot popularity, providing intuitive spatial reference for buyers' site selection decisions. Alongside GIS map marking, detailed information about the marked plots is also provided, including basic attributes such as location, area, planned use, floor area ratio, starting price, and building height restrictions, enabling buyers to fully grasp key information about recommended plots without leaving the map interface.

[0088] It should be noted that while outputting recommendation results, the system continuously records the interactive behavior of demanders on the platform, including clicks, favorites, inquiries, and bidding on recommended plots. This behavioral data is fed back into the user preference model. Combining transaction data and matching effect evaluation, the system regularly analyzes the conversion rate and matching degree of the recommendation results, and updates the adjustable weights of demander profile features and the parameter configuration of the recommendation algorithm model accordingly, achieving continuous iteration and dynamic optimization of recommendation performance.

[0089] The above are preferred embodiments of this application. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method for intelligent matching of land resource supply and demand, characterized in that, include: The system acquires data on the interaction behavior of demanders with respect to land parcels, the demanders' historical land acquisition records, the demanders' basic information, and preset behavior type depth weights. The interaction behavior data includes behavior type and the time of occurrence of the behavior. The preset behavior type depth weights represent the differences in the intensity of preference indications for different interaction behavior types. Based on the historical land acquisition records, a decision cycle benchmark value for the demand side is determined; based on the decision cycle benchmark value, the preset behavior type depth weight corresponding to the behavior type, and the behavior occurrence time, the preference score corresponding to the interaction behavior data is subjected to time decay processing to obtain a decayed preference score. Based on the basic information of the demander, the demander's historical land acquisition records, and the interaction behavior data, the similarity between the demander and other demanders is determined; based on the similarity, the decay preference scores of the other demanders, and the comprehensive similarity between land parcels, the predicted interest of the demander in non-interacting land parcels is determined; and land parcel recommendation results are generated based on the predicted interest.

2. The intelligent matching method for land resource supply and demand according to claim 1, characterized in that, The step of determining the decision cycle benchmark value of the demander based on the historical land acquisition records includes: Extract the historical land acquisition time of the demander to determine the time interval between adjacent land acquisition activities; The average time interval is obtained by averaging the time intervals. The average land acquisition time interval is normalized to obtain the decision cycle benchmark value.

3. The intelligent matching method for land resource supply and demand according to claim 1, characterized in that, The step of performing time decay processing on the preference score corresponding to the interaction behavior data based on the decision cycle benchmark value, the preset behavior type depth weight corresponding to the behavior type, and the behavior occurrence time to obtain a decayed preference score includes: The adaptive decay half-life is determined based on the decision cycle baseline value, the depth-sensitive adjustment parameter, and the preset behavior type depth weight corresponding to the behavior type; wherein, the higher the preset behavior type depth weight, the longer the corresponding half-life, and the depth-sensitive adjustment parameter is used to control the degree of influence of the preset behavior type depth weight on the half-life. The time interval between the occurrence of the behavior and the current time is determined based on the difference between the time when the behavior occurred and the current time. Using the adaptive decay half-life and the time interval since the behavior, the preference score corresponding to the interaction behavior data is subjected to time decay processing to obtain the decayed preference score.

4. The intelligent matching method for land resource supply and demand according to claim 1, characterized in that, The step of determining the similarity between the demander and other demanders based on the demander's basic information, historical land acquisition records, and interaction behavior data includes: Extract profile attributes from the basic information of the demander and the demander’s historical land acquisition records, and construct a profile feature vector; Explicit interaction behaviors are extracted from the interaction behavior data, and explicit behavior feature vectors are constructed. Implicit interaction behaviors are extracted from the interaction behavior data to construct implicit behavior feature vectors; Calculate the similarity of portrait features, explicit behavioral features, and implicit behavioral features respectively; Adjustable weights are assigned to the portrait feature similarity, explicit behavioral feature similarity, and implicit behavioral feature similarity; the similarity is obtained by weighted summation of the portrait feature similarity, explicit behavioral feature similarity, and implicit behavioral feature similarity.

5. The intelligent matching method for land resource supply and demand according to claim 4, characterized in that, The step of extracting profile attributes from the basic information of the demander and the demander's historical land acquisition records to construct a profile feature vector includes: The profile attributes are extracted from the basic information of the demander and the demander's historical land acquisition records; wherein, the profile attributes include at least one of the following: enterprise asset status, business type, and intended region; The extracted portrait attributes are converted into label vectors to obtain the portrait features.

6. The intelligent matching method for land resource supply and demand according to claim 4, characterized in that, The explicit interactive behaviors include at least one of bidding, consultation, and favorites; the implicit interactive behaviors include at least one of page browsing, repeated visits, and browsing duration.

7. The intelligent matching method for land resource supply and demand according to claim 1, characterized in that, The step of determining the predicted interest of the demander in non-interacting land parcels based on the similarity, the decay preference scores of the other demanders, and the comprehensive similarity between land parcels includes: Based on the similarity and the decay preference scores of the other demanders, a first predicted interest level is obtained; The second predicted interest degree is obtained based on the similarity matching algorithm of the plots; The first predicted interest score and the second predicted interest score are weighted and fused to obtain the predicted interest score; wherein the weights of the weighted fusion are dynamically adjusted according to historical transaction data.

8. The intelligent matching method for land resource supply and demand according to claim 7, characterized in that, The step of obtaining the first predicted interest degree based on the similarity and the decay preference scores of the other demanders includes: Using the similarity as a weight, the decay preference scores of the other demanders are summed in a weighted manner to obtain a weighted preference score. The absolute values ​​of the similarities are summed to obtain the similarity normalization factor; The weighted preference score is normalized using the similarity normalization factor to obtain the first predicted interest level.

9. The intelligent matching method for land resource supply and demand according to claim 7, characterized in that, The second predicted interest degree obtained based on the land parcel similarity matching algorithm includes: Obtain data on supporting facilities around the land parcel, and determine the scarcity weight of each type of facility based on the global distribution density of the supporting facility data; Calculate the spatial potential energy value of the land parcel based on the scarcity weight and the synergistic effect between facilities. The spatial similarity bandwidth is determined based on the spatial potential energy value. The spatial similarity between land parcels is determined based on the spatial similarity bandwidth. Determine the similarity of land parcel attributes based on the numerical attributes of the land parcels; The spatial similarity and the land parcel attribute similarity are weighted and fused to obtain the comprehensive land parcel similarity. The second predicted interest level is determined based on the comprehensive similarity of the land parcels and the historical preferences of the demanders for land parcels.

10. The intelligent matching method for land resource supply and demand according to claim 1, characterized in that, The process of generating land parcel recommendation results based on the predicted interest level includes: Based on the predicted interest levels from high to low, a preset number of non-interactive land parcels are selected to generate a recommendation list; or non-interactive land parcels with predicted interest levels exceeding a preset threshold are marked on a GIS map for display.

Images