House source recommendation method, device, equipment, medium and program product

Abstract

This application provides a method, apparatus, device, medium, and program product for recommending housing listings, relating to the field of artificial intelligence technology. It is used to accurately recommend housing listings to users. The specific technical solution is as follows: Based on the spatiotemporal behavior data of a target user within a predetermined time period, at least one cluster of persistent points for the target user and semantic tags for each cluster of persistent points corresponding to different life scenarios are determined; based on the target user's behavior data on a housing transaction platform, the target user's housing transaction intent information is determined; based on the user's spatiotemporal behavior data, the semantic tags for each cluster of persistent points corresponding to different life scenarios, and the housing transaction intent information, user feature information of the target user is generated; the user feature information and the attribute feature information of each candidate housing listing are input into a dual-tower model to obtain a first recommendation score for each candidate housing listing; based on the first recommendation score of each candidate housing listing, housing listings are recommended to the target user. This application is applied in rental recommendation scenarios.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, medium, and program product for recommending housing listings. Background Technology

[0002] With the rapid development of internet and artificial intelligence technologies, rental services are gradually transforming towards digitalization and intelligence. As the core carrier connecting users and housing listings, the personalization and contextualization of housing recommendation systems directly affect user experience and service efficiency. Therefore, improving the accuracy of housing recommendation systems has become an important development trend in this field.

[0003] Currently, conventional property recommendation systems primarily rely on explicit user behavior data (such as search keywords, filter criteria, click history, and favorites) and basic geographic location information (such as the distance between the user's current location and the property) to recommend properties to users. Specifically, when recommending properties to users, these systems typically require users to actively input keywords. The system then extracts the user-input keywords, matches them with existing properties, and then recommends matching properties to the user.

[0004] However, since the accuracy of the aforementioned property recommendation system depends on the accuracy of the search terms entered by the user, if the user enters vague keywords, the system will be unable to accurately identify the user's intent. Consequently, the property recommendations will not match the user's needs, resulting in low accuracy. Summary of the Invention

[0005] This application provides a property recommendation method for more accurate property recommendations to users.

[0006] Firstly, embodiments of this application provide a method for recommending housing listings, the method comprising:

[0007] Based on the spatiotemporal behavior data of the target users within a predetermined time period, at least one cluster of permanent locations of the aforementioned target users and semantic tags of life scenarios corresponding to each cluster of permanent locations are determined.

[0008] Based on the behavioral data of the aforementioned target users on the housing transaction platform, the housing transaction intention information of the aforementioned target users is determined;

[0009] Based on the above user spatiotemporal behavior data, the semantic tags of life scenarios corresponding to each of the above permanent location clusters, and the above housing transaction intent information, user characteristic information of the above target users is generated.

[0010] Input the above user feature information and the attribute feature information of each candidate property into the dual-tower model to obtain the first recommendation score for each candidate property;

[0011] Based on the first recommendation score of each candidate property, properties are recommended to the target users mentioned above.

[0012] The technical solution provided in this application brings at least the following beneficial effects: Based on the spatiotemporal behavior data of the target user within a predetermined time period, at least one cluster of permanent locations of the target user is determined, and each cluster of permanent locations is labeled with a corresponding semantic tag for a life scenario. Then, the semantic tag for the life scenario corresponding to each cluster of permanent locations is fused with the housing transaction intent information tag, so that the obtained user feature information can reflect the type of life scenario corresponding to the user at different permanent locations. Therefore, when recommending housing based on the user feature information and the attribute feature information of candidate housing, housing that is more in line with the life scenario of the target user can be recommended, thereby improving the accuracy of housing recommendation.

[0013] One possible implementation is that the aforementioned user spatiotemporal behavior data includes at least one location information of the aforementioned target user within the aforementioned predetermined time period, and each location information includes a time point and the coordinates of the location of the aforementioned target user at that time point;

[0014] Based on the spatiotemporal behavior data of the target user within a predetermined time period, the above-mentioned method determines at least one cluster of persistent points for the target user and semantic tags for each cluster of persistent points, including:

[0015] Clustering is performed on at least one location information of the target user within the aforementioned predetermined time period to obtain at least one cluster of permanent points.

[0016] Based on at least one preset semantic judgment rule, the first life scene semantic label contained in the first semantic judgment rule satisfied by the first permanent point cluster is determined as the life scene semantic label corresponding to the first permanent point cluster; the first permanent point cluster is one of the at least one permanent point cluster.

[0017] Each semantic judgment rule corresponds to a time period and a semantic label for a life scenario.

[0018] Another possible implementation is that each semantic judgment rule includes at least one of the following: the time period definition corresponding to the semantic judgment rule, the scene description information of the life scenario to which the time period applies, the judgment condition, and the corresponding life scenario semantic label.

[0019] Another possible implementation, based on the first recommendation score of each candidate property, recommends properties to the target user, including:

[0020] The user-property interaction graph is aggregated using a graph neural network (GNN) to update the user feature information corresponding to the user node and the property feature information corresponding to the property node in the user-property interaction graph. The user-property interaction graph is generated based on the interaction data between at least one user and at least one candidate property.

[0021] Based on the updated user feature information corresponding to the user node and the property feature information corresponding to the property node in the above user-property interaction graph, a second recommendation score is obtained for each candidate property.

[0022] Based on the first and second recommendation scores of each candidate property, the target recommendation score for each candidate property is obtained;

[0023] Based on the target recommendation score for each candidate property, properties are recommended to the aforementioned target users.

[0024] Another possible implementation involves inputting the aforementioned user characteristic information and the attribute characteristic information of each candidate property into the dual-tower model to obtain the first recommendation score for each candidate property, including:

[0025] The aforementioned user feature information is input into the user embedding network in the aforementioned dual-tower model for feature encoding, and the user feature vector is output.

[0026] The attribute feature information of each candidate property is input into the property embedding network in the above dual-tower model for feature encoding, so as to obtain the property feature vector of each candidate property.

[0027] Based on the user feature vector and the property feature vector of each candidate property, the first recommendation score for each candidate property is output.

[0028] Secondly, embodiments of this application provide a housing recommendation device, which includes:

[0029] The determination module is used to determine at least one cluster of permanent locations of the target user and the semantic tags of life scenarios corresponding to each cluster of permanent locations based on the user's spatiotemporal behavior data within a predetermined time period.

[0030] The aforementioned determining module is also used to determine the housing transaction intent information of the aforementioned target user based on the behavioral data of the aforementioned target user on the housing transaction platform;

[0031] The generation module is used to generate user feature information of the target user based on the above-mentioned user spatiotemporal behavior data, the life scene semantic tags corresponding to each of the above-mentioned permanent point clusters, and the above-mentioned housing transaction intent information.

[0032] The aforementioned generation module is also used to input the aforementioned user feature information and the attribute feature information of each candidate property into the dual-tower model to obtain the first recommendation score for each candidate property.

[0033] The recommendation module is used to recommend properties to the target users based on the first recommendation score of each candidate property.

[0034] One possible implementation is that the aforementioned user spatiotemporal behavior data includes at least one location information of the aforementioned target user within the aforementioned predetermined time period, and each location information includes a time point and the coordinates of the location of the aforementioned target user at that time point;

[0035] The aforementioned determining module is specifically used for:

[0036] Clustering is performed on at least one location information of the target user within the aforementioned predetermined time period to obtain at least one cluster of permanent points.

[0037] Based on at least one preset semantic judgment rule, the first life scene semantic label contained in the first semantic judgment rule satisfied by the first permanent point cluster is determined as the life scene semantic label corresponding to the first permanent point cluster; the first permanent point cluster is one of the at least one permanent point cluster.

[0038] Each semantic judgment rule corresponds to a time period and a semantic label for a life scenario.

[0039] Another possible implementation is that each semantic judgment rule includes at least one of the following: the time period definition corresponding to the semantic judgment rule, the scene description information of the life scenario to which the time period applies, the judgment condition, and the corresponding life scenario semantic label.

[0040] Another possible implementation, the aforementioned recommended module, is specifically used for:

[0041] The user-property interaction graph is aggregated using a graph neural network (GNN) to update the user feature information corresponding to the user node and the property feature information corresponding to the property node in the user-property interaction graph. The user-property interaction graph is generated based on the interaction data between at least one user and at least one candidate property.

[0042] Based on the updated user feature information corresponding to the user node and the property feature information corresponding to the property node in the above user-property interaction graph, a second recommendation score is obtained for each candidate property.

[0043] Based on the first and second recommendation scores of each candidate property, the target recommendation score for each candidate property is obtained;

[0044] Based on the target recommendation score for each candidate property, properties are recommended to the aforementioned target users.

[0045] Another possible implementation, the aforementioned generation module, is specifically used for:

[0046] The aforementioned user feature information is input into the user embedding network in the aforementioned dual-tower model for feature encoding, and the user feature vector is output.

[0047] The attribute feature information of each candidate property is input into the property embedding network in the above dual-tower model for feature encoding, so as to obtain the property feature vector of each candidate property.

[0048] Based on the user feature vector and the property feature vector of each candidate property, the first recommendation score for each candidate property is output.

[0049] Thirdly, this application provides an electronic device comprising: a processor and a memory; the memory stores a program or instructions executable on the processor, wherein the program or instructions, when executed by the processor, implement the method of the first aspect described above.

[0050] Fourthly, this application provides a readable storage medium on which a program or instructions are stored, which, when executed by a computer, implement the method of the first aspect described above.

[0051] Fifthly, this application provides a computer program product stored in a storage medium, which, when executed by a computer, implements the method described in the first aspect.

[0052] In a sixth aspect, embodiments of this application provide a chip including a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the method described in the first aspect.

[0053] The beneficial effects of the second to sixth aspects mentioned above are described in the corresponding description of the first aspect and will not be repeated here. Attached Figure Description

[0054] Figure 1 A schematic diagram of the network architecture for a housing recommendation method provided in this application embodiment;

[0055] Figure 2 A flowchart illustrating a housing recommendation method provided in an embodiment of this application;

[0056] Figure 3 A flowchart illustrating a housing recommendation method provided in an embodiment of this application;

[0057] Figure 4A flowchart illustrating a housing recommendation method provided in an embodiment of this application;

[0058] Figure 5 A flowchart illustrating a housing recommendation method provided in an embodiment of this application;

[0059] Figure 6 This is a schematic diagram of the structure of a housing recommendation system provided in an embodiment of this application;

[0060] Figure 7 This is a schematic diagram illustrating the working principle of the user spatiotemporal behavior profile construction module provided in the embodiments of this application;

[0061] Figure 8 This is a schematic diagram illustrating the working principle of the user rental intent recognition module provided in this application embodiment;

[0062] Figure 9 A schematic diagram illustrating the working principle of the multimodal user demand vector fusion module provided in this application embodiment;

[0063] Figure 10 A schematic diagram illustrating the working principle of the housing feature and recommendation ranking module provided in this application embodiment;

[0064] Figure 11 A schematic diagram illustrating the working principle of the graph neural network-enhanced collaborative filtering module provided in this embodiment of the application;

[0065] Figure 12 A flowchart illustrating the implementation process of a housing recommendation method provided in this application embodiment;

[0066] Figure 13 This is a schematic diagram of a housing recommendation device provided in an embodiment of this application;

[0067] Figure 14 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0068] The following section will describe in detail the housing recommendation method, apparatus, equipment, media, and program products provided in this application, with reference to the accompanying drawings.

[0069] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0070] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0071] The terms "at least one," "at least one of," etc., used in the specification and claims of this application refer to any one, any two, or a combination of two or more of the included items. For example, at least one of a, b, and c can mean: "a," "b," "c," "a and b," "a and c," "b and c," and "a, b, and c," where a, b, and c can be single or multiple. Similarly, "at least two" refers to two or more items, and its meaning is similar to that of "at least one."

[0072] In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0073] The embodiments of this application provide a method, apparatus, device, medium, and program product for recommending housing listings, which can be applied to rental recommendation scenarios.

[0074] Currently, property recommendation systems primarily rely on explicit user behavior data (such as search keywords, filtering criteria, click history, and favorites) and basic geographic location information (such as the distance between the user's current location and the property). In terms of technical implementation, mainstream solutions include collaborative filtering, content-based recommendation, and traditional machine learning ranking models (such as logistic regression and gradient boosting trees). With the development of deep learning technology, some systems have begun to introduce sequence models (such as GRU and LSTM) to model user behavior sequences, or to adopt a dual-tower neural network architecture to achieve feature matching between users and properties.

[0075] However, when recommending properties to users, property recommendation systems typically require users to actively input keywords. The system then extracts these keywords and matches them with existing properties before recommending matching listings. Therefore, such systems often fail to accurately identify user intent, leading to a mismatch between recommended properties and user needs, resulting in low recommendation accuracy.

[0076] To address the aforementioned technical issues, this application provides a method, apparatus, device, medium, and program product for recommending housing listings. Based on spatiotemporal behavior data of a target user within a predetermined time period, at least one cluster of the target user's permanent locations is determined. Each cluster of permanent locations is labeled with a corresponding semantic tag representing a lifestyle scenario. Then, the semantic tags representing the lifestyle scenario of each cluster of permanent locations are fused with housing transaction intent information tags. This allows the obtained user characteristic information to reflect the type of lifestyle scenario corresponding to the user at different permanent locations. Therefore, when recommending housing listings based on this user characteristic information and the attribute characteristics of candidate housing listings, listings that better match the target user's lifestyle scenario can be recommended, improving the accuracy of housing listing recommendations.

[0077] The following description, in conjunction with the accompanying drawings, details the housing recommendation method, apparatus, equipment, medium, and program products provided in the embodiments of this application.

[0078] Figure 1 This illustration shows the network architecture of a housing recommendation method provided in an embodiment of this application. For example... Figure 1 As shown, the network architecture includes a property recommendation device 101 and a terminal device 102. The property recommendation device 101 and the terminal device 102 are interconnected.

[0079] In some embodiments, the property recommendation device 101 may be a server, a computer, or a processor or processing unit within a server or computer. The server may be a single server or a server cluster comprising multiple servers. It should be noted that this application does not limit the specific device form of the property recommendation device 101. Figure 1 The example shown is a single server, 101, which is used as a housing recommendation device.

[0080] In some embodiments, the terminal device may be a mobile phone, tablet computer, laptop computer, handheld computer, in-vehicle electronic device, mobile internet device (MID), augmented reality (AR) / virtual reality (VR) device, robot, wearable device, personal computer (PC), ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc., and the embodiments of this application do not specifically limit it. Figure 1 The example shown is a mobile phone, with terminal device 102 as an example.

[0081] In some embodiments, the property recommendation device 101 collects the location information and historical data of the terminal device 102 to obtain the target recommendation score for each candidate property, and recommends the property information with the highest target recommendation score to the terminal device 102.

[0082] It should be noted that the network architecture described in the embodiments of this application is for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and does not constitute a limitation on the technical solutions provided in the embodiments of this application. As network architectures evolve, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0083] See Figure 2 This is a flowchart illustrating a housing recommendation method provided in an embodiment of this application. Figure 2 As shown, the housing recommendation method provided in this application embodiment can be implemented by the above-mentioned housing recommendation device, specifically including the following steps 201 to 205.

[0084] Step 201: The housing recommendation device determines at least one cluster of permanent locations of the target user and the semantic tags of the life scene corresponding to each cluster of permanent locations based on the user's spatiotemporal behavior data within a predetermined time period.

[0085] In some embodiments, user spatiotemporal behavior data is the result of binding time, space, and behavior, indicating that a user appears at a certain time and place and performs a certain behavior.

[0086] In some embodiments, a resident cluster refers to a dense group of points formed by clustering location information that a user has been using repeatedly for a long time and for extended periods.

[0087] In some embodiments, life scenario semantic tags are used to characterize the degree of user dependence on a certain life scenario, including tags such as user residence, user workplace, and user school-age family.

[0088] In some embodiments, the aforementioned user spatiotemporal behavior data includes at least one location information of the target user within the predetermined time period, each location information including a time point and the coordinates of the target user's location at that time point, combined with... Figure 2 ,like Figure 3 As shown, step 201 above can be implemented through steps 201a and 201b.

[0089] Step 201a: The housing recommendation device performs clustering processing on at least one location information of the target user within a predetermined time period to obtain at least one resident point cluster.

[0090] In some embodiments, at least one location information of the user includes the latitude and longitude coordinates of the user's location, the timestamp of the time when the user's location information was collected, etc.

[0091] In some embodiments, clustering processing refers to grouping user location information by density using clustering algorithms (such as DBSCAN clustering algorithm, K-Means clustering algorithm, hierarchical clustering algorithm, Gaussian mixture model GMM clustering algorithm).

[0092] Step 201b: The housing recommendation device determines the first life scene semantic tag contained in the first semantic judgment rule satisfied by the first permanent point cluster as the life scene semantic tag corresponding to the first permanent point cluster based on at least one preset semantic judgment rule.

[0093] The first resident cluster mentioned above is one of the at least one resident clusters mentioned above.

[0094] Each semantic judgment rule corresponds to a time period and a semantic label for a life scenario.

[0095] In some embodiments, the semantic judgment rule is based on the user's stay in three preset time windows and the corresponding dwell ratio threshold.

[0096] For example, each semantic judgment rule includes at least one of the following: the time period definition corresponding to the semantic judgment rule, the scene description information of the life scene to which the time period applies, the judgment condition, and the corresponding life scene semantic label.

[0097] For example, the judgment condition is to distinguish the life scenarios suitable for users in the preset overlapping time periods, and to perform detailed segmentation and judgment on the life scenarios. The judgment condition is a dual verification of geofence verification and stay mode verification.

[0098] For example, the preset time period is 22:00–06:00 (nighttime). This time period is applicable to the life scenarios of sleep and home. If the proportion of time a user records in the permanent location cluster during this time period is greater than the proportion threshold, it means that the semantic tag of the life scenario corresponding to the permanent location cluster is the place of residence.

[0099] For example, the preset time period is 09:00–18:00 (daytime working hours). This time period is applicable to office and business activity scenarios. If the proportion of time a user records in the cluster of permanent locations during this time period is greater than the proportion threshold, then the semantic tag of the life scenario corresponding to the cluster of permanent locations is the workplace.

[0100] For example, the preset time period is 08:00–16:00 (weekday daytime). This time period is applicable to life scenarios such as dropping off / picking up students and school scenarios. If a user appears for 5 consecutive weekdays during this time period and meets the dual verification of geofencing and dwell pattern, it means that the semantic tag of the life scenario corresponding to the permanent cluster of points during this time period is the school-age family tag.

[0101] For example, geofence verification indicates that there is an educational POI (kindergarten, primary school, junior high school) within 200 meters of the user's permanent location coordinates, and that the POI is marked with the "school district" attribute in the real estate database.

[0102] For example, the dwell pattern verification indicates that the dwell time distribution within the user's permanent dwell point cluster is a "bimodal pattern" (i.e., two dwell time peaks occur during the morning drop-off period 07:30–08:30 and the afternoon pick-up period 15:00–16:30), and the user has another daytime permanent dwell point cluster (i.e., workplace) during other time periods (09:00–15:00).

[0103] In some embodiments, a triple verification mechanism of semantic judgment rule verification, geofence verification, and dwell pattern verification is satisfied, and the cluster of permanent dwell points within the user's preset time period of 08:00–16:00 is marked as a school-age family tag.

[0104] In this way, by collecting users' location information and transforming it into semantic preference tags with recommendation decision-making value, namely life scenario semantic tags, it can more intuitively reflect users' living preferences, thereby recommending housing to users more accurately and improving the accuracy of housing recommendations.

[0105] Step 202: The housing recommendation device determines the housing transaction intention information of the target user based on the target user's behavioral data on the housing transaction platform.

[0106] In some embodiments, user behavior data on the housing transaction platform may include user search history, viewing history, interaction history, etc. on the housing transaction platform. The specific data may be determined according to the use case, and this application does not impose any specific limitations.

[0107] In some embodiments, a user's intention to transact in a property transaction can be calculated using the following formula:

[0108] ;

[0109] (1)

[0110] in, Indicates the intensity factor of user behavior. Indicates the normalized search frequency. Indicates the normalized dwell time. Represents the balance coefficient. This indicates linear weighted fusion.

[0111] Step 203: The housing recommendation device generates user characteristic information of the target user based on user spatiotemporal behavior data, semantic tags of life scenarios corresponding to each permanent location cluster, and housing transaction intent information.

[0112] In some embodiments, user characteristic information is obtained by concatenating user spatiotemporal behavior data, semantic tags of living scenarios corresponding to each cluster of permanent locations, and housing transaction intent information. The spatiotemporal behavior data includes place of residence. Workplace School-age family tag Each resident cluster corresponds to a semantic tag for a specific living scenario, including a vector representing the neighborhood's grade. Supporting demand vectors Commuting time School-age needs Intent information includes intent score. wait.

[0113] Step 204: The housing recommendation device inputs user feature information and attribute feature information of each candidate housing into the dual-tower model to obtain the first recommendation score for each candidate housing.

[0114] In some embodiments, the attribute characteristics of each candidate property include price, area, unit type, location, surrounding amenities, etc.

[0115] In some embodiments, combined with Figure 2 ,like Figure 4 As shown, step 204 above can be implemented through steps 204a to 204c.

[0116] Step 204a: The housing recommendation device inputs user feature information into the user embedding network in the dual-tower model for feature encoding and outputs user feature vectors.

[0117] In some embodiments, the dual-tower model can be used to perform similarity matching between user feature information and the property attribute information of each candidate property to obtain a similarity score, i.e., the first recommendation score.

[0118] In some embodiments, feature encoding can be performed using the following formula:

[0119] (2)

[0120] in, This indicates that the above user is embedded in the network. The output of the layer, i.e., the aforementioned user embedding network... Input of layer +1, This indicates the number of network layers in the user-embedded network (this module is set to 3 layers). Indicates the first The weight matrix of the layer, Indicates the first Layer bias vector, This represents the activation function. The last layer (the first) The output of the layer is the user feature vector, where The embedding dimension has a value of 64.

[0121] Step 204b: The housing recommendation device inputs the attribute feature information of each candidate housing into the housing embedding network in the dual-tower model for feature encoding, and obtains the housing feature vector of each candidate housing.

[0122] In some embodiments, each candidate property has attribute feature information (price, area, unit type, geographical location, surrounding amenities, etc.), and the property feature vector of each candidate property is obtained through a property embedding network (with the same structure as the user embedding network but with independent parameters).

[0123] Step 204c: The housing recommendation device outputs the first recommendation score for each candidate housing unit based on the user feature vector and the housing feature vector of each candidate housing unit.

[0124] In some embodiments, the first recommendation score is obtained by taking the inner product of the user feature vector and the property feature vector of each candidate property, specifically calculated using the following formula:

[0125] (3)

[0126] in, This indicates the first recommendation score. Represents the user feature vector. Represents the feature vector of a property

[0127] In this way, by using user characteristic information and the property attribute information of each candidate property, a first recommendation score is output for each candidate property. This recommendation score is then used to recommend properties to users more accurately, thereby improving the accuracy of property recommendations.

[0128] Step 205: The housing recommendation device recommends housing to the target user based on the first recommendation score of each candidate housing.

[0129] In some embodiments, combined with Figure 2 ,like Figure 5 As shown, step 205 above can be implemented through steps 205a to 205d.

[0130] Step 205a: The housing recommendation device performs graph neural network (GNN) information aggregation on the user-housing interaction graph and updates the user feature information corresponding to the user node and the housing feature information corresponding to the housing node in the user-housing interaction graph.

[0131] The above user-property interaction graph is generated based on the interaction data between at least one user and the at least one candidate property.

[0132] In some embodiments, the update of the user-property interaction map can be achieved using the following formula:

[0133] ;

[0134] (4)

[0135] in, This indicates the updated user node. This indicates the updated property listing node. Indicates the user identifier subscript. Indicates the subscript of the property listing. This represents the set of properties that user u has interacted with. Indicate its quantity, Indicates housing resources The set of users who have interacted with the platform Indicate its quantity, express The node represents a vector. express The node representation vectors are initially input as user embeddings and property embeddings: , This represents the coefficients after normalization. That is, for each element in the input vector, if it is positive, it remains unchanged, and if it is negative, it is set to zero. This function is widely used in deep neural networks. Its advantages are simple computation, effective mitigation of the gradient vanishing problem, and introduction of non-linear expressive power.

[0136] Step 205b: The property recommendation device obtains a second recommendation score for each candidate property based on the user feature information corresponding to the user node and the property feature information corresponding to the property node in the updated user-property interaction graph.

[0137] In some embodiments, the second recommendation score is obtained by performing an inner product operation on the user nodes and property nodes aggregated by the GNN.

[0138] Step 205c: The housing recommendation device obtains the target recommendation score for each candidate housing unit based on the first recommendation score and the second recommendation score of each candidate housing unit.

[0139] In some embodiments, the target recommendation score can be calculated using the following formula:

[0140] (5)

[0141] in, This indicates the second recommended score. The fusion coefficient is used to control the relative weights of the dual-tower direct matching signal and the GNN collaborative filtering signal. L represents the number of GCN layers, which is set to 2 layers in this system.

[0142] Step 205d: The housing recommendation device recommends housing to the target user based on the target recommendation score of each candidate housing.

[0143] In some embodiments, the property recommendation device sorts all candidate properties by their target recommendation scores from highest to lowest, and then recommends at least one property with the highest target recommendation score to the target user.

[0144] In this way, by using the first and second recommendation scores of each candidate property, a target recommendation score is obtained for each candidate property. Based on the target recommendation score of each candidate property, properties are recommended to the target user, thereby improving the accuracy of property recommendations.

[0145] The housing recommendation method provided in this application transforms user location information into semantic tags of living scenarios that more intuitively reflect user's residential preferences. By using historical data of user interactions with each candidate housing and housing attribute information of each candidate housing, a target recommendation score is obtained for each candidate housing. This recommendation score is then used to recommend housing to users more accurately, thereby improving the accuracy of housing recommendations.

[0146] The following specific embodiments illustrate the housing recommendation method of this application.

[0147] This application proposes an intelligent rental housing recommendation system based on spatiotemporal behavior perception and multimodal data fusion, the overall architecture of which is as follows: Figure 6 As shown, the system mainly consists of five core modules: a user spatiotemporal behavior profile construction module, a user rental intent recognition module, a multimodal user demand vector fusion module, a housing resource feature and recommendation ranking module, and a graph neural network-enhanced collaborative filtering module.

[0148] The intelligent rental housing recommendation system can integrate multi-source data from operators (location signaling, internet access behavior), construct multi-dimensional profiles of users' housing needs, and introduce a ranking mechanism that combines graph neural networks and a dual-tower model to achieve accurate and scenario-based rental recommendations.

[0149] The purpose of this embodiment is to provide an intelligent rental housing recommendation system based on spatiotemporal behavior perception and multimodal data fusion. It should be noted that user residential preferences, compared to general user preferences in e-commerce, content, and other fields, have the following unique technical characteristics, which render existing graph neural network and dual-tower model combination schemes unsuitable for direct application:

[0150] Multi-semantic conflicts of spatial anchors: User housing preferences are constrained by multiple geographical anchors (residence, workplace, school), and the decision weight of each anchor changes dynamically with the user's life stage. Existing dual-tower models compress user-side features into a single embedding vector, failing to decouple and independently model the differentiated contributions of different anchors to housing matching in the vector space. This results in the irreversible dilution of information regarding the trade-off between commuting distance and school district requirements during the embedding process.

[0151] Spatiotemporal heterogeneity of interaction graphs: User-property interaction graphs in rental scenarios exhibit strong spatial locality and relevance to different life stages. For example, the interaction behavior of school-aged families is highly clustered within a specific school district, while the interaction behavior of young professionals is distributed along commuting corridors. Existing message passing mechanisms in graph neural networks do not consider this spatiotemporal heterogeneity. When aggregating neighbors, they mix interaction signals from different life scenarios with equal weight, leading to "semantic oversmoothing" of node representations—that is, the school district preference of school-aged families is overwhelmed by the interaction signals of users outside the school district.

[0152] The non-interactive observability of implicit preferences: Users' preferences for attributes such as neighborhood grade and residential social circle are often not expressed through explicit clicks, but are implicit in the objective attributes of their long-term residence. Existing dual-tower models rely on user-property interaction data as supervisory signals, lacking a perceptual channel for preference dimensions (such as quietness and community quality) that users have never interacted with but which are already reflected in their residences, resulting in user embedding vectors that cannot be aligned with their real-life standards.

[0153] In view of the unique characteristics of the aforementioned user residential preferences, the purpose of this invention is to solve the following technical problems:

[0154] Problem 1: The dual-tower model's embedding space cannot accommodate multi-anchor semantics. This invention addresses this by constructing a user spatiotemporal behavior profile module, explicitly extracting multi-anchor information (residential grade, commuting time, school district requirements) as structured features, and then using a multimodal fusion network to perform feature cross-referencing. This allows the user embedding vector in the dual-tower model to retain the relative importance ranking of multiple anchors in a compressed form, rather than simply mixing them.

[0155] Problem 2: Semantic oversmoothing in spatiotemporally heterogeneous graphs by graph neural networks. This invention introduces attention weights based on life stage perception during graph convolution aggregation, enabling node updates to distinguish the importance differences of neighbors in different life scenarios and preventing school district preferences from being diluted by commuting preference signals.

[0156] Problem 3: The lack of supervised signals for implicit residential preferences. This invention utilizes operator location signaling to extract objective semantic labels of the user's current residence, using this as an auxiliary task for self-supervised learning to guide the user's embedding vector towards their real-life standards, thus compensating for the lack of implicit preference signals in the interaction data.

[0157] In summary, the purpose of this embodiment is to overcome the inherent limitations of existing model architectures in user residential preference scenarios through the specially designed technical means described above, to achieve in-depth mining of users' implicit residential preferences and accurate identification of life scenarios, and ultimately to achieve accurate and scenario-based rental property recommendations.

[0158] The system first collects multi-source user data, including operator location signaling data (latitude and longitude, timestamps), digital behavior trajectory data (access to rental platforms, clicks, searches, favorites, inquiries, etc.), basic property data (price, area, unit type, location, surrounding amenities, etc.), and a real estate database (neighborhood grade, school district division, etc.). Through the collaborative processing of these modules, a personalized property recommendation list is finally output.

[0159] The user spatiotemporal behavior profile building module uses operator location signaling data to extract users' long-term spatiotemporal behavior patterns and generate tags for key life stages such as residence, workplace, and school-age family.

[0160] Step 1: Extraction of resident points

[0161] Set user In the time interval The internal location record set is ,in Latitude and longitude coordinates The timestamp is used. The DBSCAN clustering algorithm is used to extract resident clusters.

[0162] (6)

[0163] in The neighborhood radius is set to 50 meters. The minimum number of points required to form the core point (set to 10).

[0164] Step 2: Semantic Tag Generation

[0165] The time period rule in this step is defined as the following three sets of time windows and corresponding dwell ratio thresholds, which are used to map the cluster of resident points to life scene semantics:

[0166]

[0167] For each resident cluster The semantic label determination is performed by applying the above time period rules in sequence:

[0168] Nighttime residence: If rule 1 is met, then it is marked as a place of residence. ;

[0169] Daytime permanent residence: If rule 2 is met, then mark it as a workplace. ;

[0170] School Resident Verification: To avoid misidentifying office workers as school-aged families, this plan employs a triple verification mechanism: "time-of-day characteristics + geofencing + dwelling pattern".

[0171]

[0172] When all three of the above checks are satisfied, the school-age family label is generated. .

[0173] Distinguishing between teachers and office workers: If only the first and second layer of verification are met, but the third layer of stay pattern shows a single peak of stay throughout the day (09:00–17:00 continuously around the school), and the cluster is also marked as a candidate workplace, then the user is classified as an education industry practitioner, without being labeled with a school-age family tag, but a separate "education practitioner" profile can be generated for other scenarios.

[0174] Step 3: Semantic Enhancement of Residence

[0175] Obtain residence information from real estate databases Corresponding residential community level (e.g., high-end, mid-range, standard) and related information vectors Features include transportation (distance to subway stations), commerce (number of shopping malls), and medical care (distance to top-tier hospitals).

[0176] Conventional geolocation processing can only output coordinate-level information (such as latitude and longitude, grid code), but coordinates themselves do not have the semantic interpretation ability of user residential preferences. For example, two communities within the same H3 grid may differ significantly in price, neighborhood, and amenities, and cannot be distinguished by coordinates alone. This invention transforms the coordinate-level residential location into semantic preference tags with recommendation decision-making value through the following steps:

[0177] (1) Construction of the semantic mapping layer

[0178] Residual coordinates are obtained through reverse geocoding Match to a specific neighborhood in the real estate database and extract the objective attribute vector of that neighborhood. ,include:

[0179]

[0180] (2) The semantic transformation principle from "objective attributes" to "implicit preferences"

[0181] The above objective attribute vector Originally descriptive data, this invention transforms it into users' implicit residential preferences through the following technical logic:

[0182] Inverse inference of preference strength: A user's long-term residence in a particular neighborhood implies that they have actually accepted and adapted to the neighborhood's amenities in their housing choices. According to revealed preference theory, a user's actual housing choices under the same budget constraints more accurately reflect their minimum tolerance threshold and relative weighting of various amenities than their explicitly expressed preferences (such as screening criteria). For example, a user living in a neighborhood 1.5km from a subway station indicates that their tolerance for commuting convenience is at least 1.5km; this can be used to broaden or narrow the screening range when recommending new housing.

[0183] Learning the relative weights of preferences: In the subsequent multimodal fusion module, The model will cross-reference this data with user behavior data from other dimensions (such as search filter criteria and clicked property attributes). If a user's residential area has average amenities but they have never searched for "near subway," the model can infer that "commuting convenience" has a low weight in the user's preference. Conversely, if the residential area has high amenities and the user frequently filters high-priced properties, the model can confirm that "residential circle" is their core preference dimension.

[0184] (3) Technical effects of semantic enhancement

[0185] Through the semantic enhancement steps described above, location information is upgraded from "where the user is" (coordinate level) to "what kind of lifestyle the user is accustomed to" (semantic level). This qualitative improvement enables the recommender system to:

[0186] When users do not explicitly express their preferences, the system automatically aligns with their real-life standards to avoid recommending properties that are too different from their daily experiences.

[0187] In cold start or behavior sparse scenarios, the semantics of residence are used as prior knowledge to make up for the lack of interaction data.

[0188] Transforming abstract conditions such as "commuting distance" and "school district quality" into concrete values ​​that users can perceive and that align with their current lives improves the interpretability of recommendations.

[0189] Step 4: Commute Feature Calculation

[0190] Calculate residence With workplace Commuting time between (Based on the map API, predict the driving / public transportation time during the weekday morning rush hour) as a quantitative indicator of commuting tolerance.

[0191] Final output: The user spatiotemporal behavior profile contains the following structured fields for use by the subsequent multimodal fusion module:

[0192]

[0193] Process description: Figure 7 It demonstrates the complete process of building a user's spatiotemporal behavioral profile.

[0194]

[0195]

[0196] External database description: The real estate database called in step ④ includes the following external data sources:

[0197] Community basic information database (community name, coordinate boundaries, building year, property type)

[0198] Property valuation database (historical average transaction price, listing price distribution)

[0199] Supporting POI database (coordinates and ratings of subway stations, shopping malls, hospitals, and schools).

[0200] School district zoning database (mapping relationship between residential communities and their corresponding schools)

[0201] By spatially matching and associating the coordinates of the residence with the aforementioned database, the semantic enhancement from "latitude and longitude values" to "living standards" is achieved, enabling the system to measure the relative merits of candidate properties in terms of amenities based on the user's current residence.

[0202] The user rental intent recognition module constructs an intent scoring model based on the user's behavioral sequence on the rental platform to identify the user's real rental needs.

[0203] Step 1: Construction of Behavioral Sequences

[0204] Organize user interactions on rental platforms (or partner platforms) into a sequence based on time. Each interaction is encoded as a vector, containing two parts: the behavior type and the attributes of the interaction object (property listing).

[0205] (1) Definition of Behavior Type

[0206] This module defines the following six types of user interaction behaviors, each of which is assigned a unique type code to distinguish different levels of user interest in property listings:

[0207]

[0208] The above intensity weight reference values ​​are used for the calculation of subsequent behavior intensity factors. The higher the weight, the stronger the intention to rent a house reflected by the behavior.

[0209] (2) Description of the attributes of the interactive object (property listing)

[0210] For each property listing involved in an interaction, the following attribute information is extracted to construct an attribute vector (continuous values ​​normalized to...). (interval)

[0211]

[0212] (3) Example of interactive behavior vector encoding

[0213] Taking a single "details page click" action as an example, the interaction vector The encoding format is:

[0214]

[0215] The "behavior type encoding" is a one-hot vector of length 6 (corresponding to the 6 behaviors mentioned above), and the attribute vector is formed by concatenating the 10 fields mentioned above. After concatenation... The dimension is The dimension is 10 for continuous attributes; if categorical attributes use one-hot encoding, the dimension will be expanded accordingly. For search behavior (without a specific property listing), the property attribute vector is filled with 0.

[0216] (4) Examples of behavioral sequences

[0217] The following is an example of a user's behavior sequence over a week:

[0218]

[0219] The sequence is input into the GRU network in chronological order to extract user behavior patterns and the strength of their rental intention.

[0220] Step 2: Sequence Modeling

[0221] A gated recurrent unit (GRU) is used to model user behavior sequences and extract user behavior patterns and temporal features of intent. GRU effectively alleviates the gradient vanishing problem in long sequence training through its gating mechanism and is suitable for variable-length time-series data such as user behavior sequences.

[0222] (1) GRU cell forward propagation formula

[0223] Let the first The input for each step is an interaction behavior vector. The input feature dimension is determined according to the encoding method in step 1. The forward propagation calculation for a single GRU unit is as follows:

[0224] Reset door :

[0225] (7)

[0226] Update Gate :

[0227] (8)

[0228] Candidate hidden state :

[0229] (9)

[0230] Final hidden state :

[0231] :;(10)

[0232] in:

[0233] This is the hidden state from the previous moment. For the hidden layer dimension (set to this module) dimension);

[0234] This is a trainable weight matrix;

[0235] It is the bias vector;

[0236] It is the Sigmoid activation function. It is the hyperbolic tangent activation function;

[0237] Represents element-wise product;

[0238] This indicates a vector concatenation operation.

[0239] (2) Initial hidden state Definition

[0240] The sequence recursion starts from $t=1$, and an initial hidden state needs to be set. This module serves as the starting point for GRU computation. The initial hidden state is initialized using an all-zero vector:

[0241] (11)

[0242] The rationale for using zero initialization is as follows: before the first observation of user behavior, the model has no prior knowledge of the user's behavior pattern, and the zero vector represents an unbiased initial state; at the same time, GRU's reset and update gate mechanisms can adjust the initial state based on the input in the first time step. By dynamically adjusting the dependence on the initial state, the hidden state can be quickly converged to an effective data-driven representation.

[0243] (3) Double-layer GRU stacking structure

[0244] To enhance the expressive power of sequence features, this module adopts a two-layer GRU stacked structure, with the hidden state sequence output by the first layer serving as the input to the second layer:

[0245] First-layer GRU:

[0246] (12)

[0247] Second-layer GRU:

[0248] (13)

[0249] (4) Sequence feature extraction

[0250] Let the length of the user behavior sequence be $N$ (the length varies for different users; in actual training, padding and masking mechanisms are used to handle variable-length inputs). The hidden state at the last time step of the second-layer GRU is taken as the compressed representation of the entire behavior sequence, i.e., the sequence feature vector. :

[0251] (14)

[0252] choose The basis for this is that the timing recursion process of GRU makes... Integrating from arrive It collects all historical interaction information, and the gating mechanism naturally assigns a higher weight to recent behaviors, which can effectively reflect the user's current rental intention pattern.

[0253] (5) Setting key parameters

[0254]

[0255] (6) Variable-length sequence processing

[0256] For insufficient sequence length For users with a sequence length exceeding a certain threshold, zero-vector padding is used to achieve a uniform length, and the padding positions are ignored during GRU computation via a mask vector to ensure that the padding portion does not participate in hidden state updates. Users who have recently used the service will only retain the most recent data. This is a secondary interaction.

[0257] (7) Technical effects

[0258] Two-level GRU sequence modeling can automatically capture the temporal dependencies between user behaviors. For example, if a user searches for "2 bedrooms and 1 living room," clicks on listings in the same area consecutively, and eventually initiates an inquiry, The encoding of the composite intent pattern of "regional preference + housing type preference + high urgency" will provide a high-quality temporal feature representation for the intent score calculation in the subsequent step 3.

[0259] Step 3: Intent Score Calculation

[0260] The sequence feature vector obtained in step 2 (i.e., the hidden state at the last time step of the two-layer GRU, dimension) Input a fully connected layer and pass it through a sigmoid activation function to obtain the user's intent probability based on the behavioral sequence. :

[0261] (15)

[0262] in: This is the weight matrix of the fully connected layer. For bias terms, The sigmoid function has an output range of 1 / 2. The physical meaning of this step is: mapping the temporal behavioral pattern extracted by GRU (which includes contextual information about behavioral type sequences and interactive property attribute sequences) into a scalar, representing a probability estimate of the user's current state of "having a clear intention to rent a house". The closer this value is to 1, the higher the urgency, focus, and clarity of needs implied in the behavioral sequence.

[0263] At the same time, considering that sequence modeling alone may not directly reflect the magnitude of intent intensity (for example, a user who only browses once in a very short period of time may have different sequence characteristics in terms of intent intensity compared to a user who operates for a long time and at a high frequency, but the model output...), (While it may not be able to completely distinguish between them), this invention introduces a behavior intensity factor to correct the intent score. The behavior intensity factor is determined by search frequency. Total valid browsing time Combined structure.

[0264] To ensure that the dimensions of the behavior intensity factor and the intention probability $p_u$ are consistent (both are dimensionless scalars), for and The following normalization process is performed:

[0265] The total number of searches performed by a user within the observation period (e.g., 7 days). Divide this by an empirically referenced upper limit. (Based on statistics, over 95% of active rental users search no more than 20 times per week), resulting in the normalized search frequency.

[0266] : The cumulative effective dwell time of users on the property details page (in minutes). Divide this by the normalized reference value. (That is, setting 60 minutes as the upper limit of the effective browsing time for high-intent users per day), to obtain the normalized dwell time. .

[0267] The formula for calculating the behavior intensity factor is the above formula (1), and will not be repeated here to avoid repetition.

[0268] The range of values ​​is This comprehensively reflects the frequency of users' active searches and the depth of their browsing engagement within a unit observation period. When users both search frequently and spend a long time reading property details, It approaches 1; conversely, it approaches 0.

[0269] Final weighted intention score From the sequence intention probability With behavioral intensity factor The linear weighted fusion is used to obtain the result. The specific calculation formula is the above formula (1). To avoid repetition, it will not be repeated here.

[0270] in This is the balance coefficient. Regarding... Basis for determining the value: This controls the model's relative trust in the "quality of behavioral patterns" and the "quantity of behavioral intensity." If... Too high (e.g.) If the intent score relies excessively on GRU's judgment of sequence patterns, then in scenarios where user behavior is sparse but intent is clear (e.g., a new user immediately inquires after their first search), GRU may output a low score due to a lack of sufficient temporal context. This leads to the intention being underestimated; if Too low (e.g.) If the intent score is primarily governed by the intensity of the behavior, it is easy to misjudge "high-frequency, low-quality browsing" (such as repeatedly refreshing but never clicking on details) as high intent. This invention uses offline A / B testing, with click-through rate (CTR) and final conversion rate (CVR) as optimization targets on the validation set, to... exist A grid search was performed within the specified range. Experimental results show that... At that time, the combined improvement in CTR and CVR was the greatest (compared to The CTR increased by approximately 2.1%, and the CVR increased by approximately 1.8%. (The CTR decreased by approximately 1.5% at that time), so 0.6 was selected as the system default parameter. This value indicates that while considering the intensity of behavior, the model places greater trust in the deep behavioral patterns mined by the sequence model.

[0271] Process description: Figure 8 This demonstrates the complete workflow of the intent recognition module. User behavior logs are preprocessed to form a sequence, which is then input into a two-layer GRU network to extract temporal features, resulting in a sequence feature vector. This vector is then passed through a fully connected layer and activated by a sigmoid function to output the intent probability. On the other hand, the behavior intensity factor is calculated by combining the normalized search frequency and dwell time. The two values ​​are weighted and fused together to output the final intent score. .

[0272] The multimodal user demand vector fusion module concatenates user spatiotemporal profile features, intent scores, and residential preference features into an original feature vector, and then performs feature cross-interaction through a deep neural network to generate user demand embeddings.

[0273] Step 1: Feature splicing

[0274] The following features are concatenated to form the original feature vector. :

[0275] -Community Grade Vector (one-hot encoding)

[0276] - Supporting demand vector (Including preferences for transportation, commerce, and healthcare, obtained through statistics of users' historical behavior)

[0277] - Commute time (After normalization)

[0278] -School-age needs (0 / 1)

[0279] -Intention Score

[0280] Step 2: Deep Neural Network Feature Crossing (User Embedding Network)

[0281] The concatenated original feature vector Input a multi-layer fully connected neural network (hereinafter referred to as the user embedding network), and realize the high-order feature cross through layer-by-layer nonlinear transformation. The specific calculation formula adopts the above formula (2). To avoid repetition, it will not be repeated here.

[0282] in: The number of network layers (this module is set to 3 layers); and The first Layer weight matrix and bias vector; The activation function. The last layer (the...) The output of the layer is the user requirement embedding vector. ,in The embedding dimension has a value of 64.

[0283] The criteria for selecting the embedding dimension $k=64$ are as follows:

[0284] Embedded Dimension The hyperparameter determines the expressive power of the user demand vector and the complexity of subsequent matching calculations, and is a key hyperparameter in the dual-tower recommendation model. This invention determines this hyperparameter through offline comparative experiments. The specific values ​​are determined based on the following:

[0285]

[0286] Experimental data show that: when When the AUC increased from 32 to 64, the AUC improved by approximately 3.2%, resulting in a significant improvement in sorting accuracy; when Increasing the embedding dimension from 64 to 128 only improved the AUC by 0.7%, while the online inference latency increased by approximately 55%, exceeding the system's set latency limit of 20ms. Considering both sorting accuracy and computational efficiency, the 64-dimensional embedding, while maintaining sufficient expressive power, meets the low latency requirements of online recommendation services and was therefore selected as the standard dimension for embedding user needs.

[0287] In addition, 64 dimensions are also a common dimension setting for embedding vectors in industry recommendation systems, which is consistent with mainstream engineering practices and facilitates model deployment and engineering optimization.

[0288] Process description: Figure 9 The process of multimodal fusion is demonstrated. Multi-source features are concatenated and input into the user embedding network. After three layers of fully connected nonlinear transformation, a 64-dimensional user demand embedding vector is output.

[0289] The property feature and recommendation ranking module uses a dual-tower model to generate user embeddings and property embeddings respectively, calculates recommendation scores through inner product, and uses Pairwise ranking loss for training.

[0290] Step 1: Embedding property features

[0291] Each property With attribute vectors (Price, area, unit type, location, surrounding amenities, etc.) are used to embed property listings through an embedded network (with the same structure as the user's embedded network but with independent parameters). .

[0292] Step 2: Calculate the recommended score

[0293] The recommended score is the inner product of user embedding and property embedding. The specific calculation formula is the above formula (3). To avoid repetition, it will not be repeated here.

[0294] Step 3: Model Training

[0295] Pairwise ranking loss is used, with user-to-positive samples from the exposed samples. (Clicks / Inquiries) and Negative Samples (Pairings that were not clicked)

[0296] (16)

[0297] in For the set of all positive and negative sample pairs, This is the Sigmoid function.

[0298] (1) Definition of exposure sample

[0299] In this system, an "exposure" refers to the act of a user seeing a listing card on the search results page or in the recommendation feed of a rental platform. Each exposure records the following information: User ID. Property listings Exposure timestamp, and whether the user engaged in positive interaction after exposure (clicking on the details page, initiating an inquiry, adding to favorites, etc.).

[0300] (2) Composition of positive samples

[0301] Positive samples Defined as: User There was proactive positive interaction with the exposed property listings. This includes the following types of behavior:

[0302]

[0303] If any of the above behaviors occur after an exposure, then the exposure will be... Samples marked as positive are included in the training set. .

[0304] (3) Composition of negative samples

[0305] negative samples Defined as: User There was no positive interaction with the displayed property listings. In other words, the listings were shown to users, but they neither clicked on the details page nor performed any subsequent actions such as adding the property to their favorites or making inquiries.

[0306] However, directly treating all "exposure without clicks" samples as negative samples leads to the following problems:

[0307] Imbalanced sample: In recommendation systems, the number of samples that have been exposed but not clicked is far greater than the number of positive samples. Directly using all negative samples will cause the model training to be biased towards negative examples, affecting the ranking effect.

[0308] Negative sample noise: A user's failure to click on a property listing may not be due to "lack of interest," but rather to reasons other than preference, such as not noticing it later in the browsing or temporarily leaving the page.

[0309] Therefore, this module uses the following negative sampling strategy to construct training sample pairs:

[0310]

[0311] (4) Training sample pair construction example

[0312] Suppose a user In a session on March 1, 2026, the system exposed five listings to him, as recorded below:

[0313]

[0314] For positive samples From the candidate negative sample pool Two are randomly selected from the middle (according to the negative sampling ratio). (part of the text) and Forming training pairs: .

[0315] For positive samples Similarly, two negative samples are randomly selected from the candidate negative sample pool to form training pairs: and .

[0316] (5) Loss function calculation

[0317] For each training sample pair ,in These are positive sample properties. For negative sample properties, the loss function is defined as:

[0318] (17)

[0319] in: The set of all positive and negative sample pairs; The recommended score output by the dual-tower model; This is the Sigmoid function.

[0320] The optimization objective of this loss function is to maximize the difference between the scores of positive and negative samples, thus forcing the model to rank positive samples before negative samples. Through the above negative sampling and loss function design, the model can effectively learn the relative preference ranking of users for different properties with limited training overhead.

[0321] Process description: Figure 10 The structure of the dual-tower model is shown. The left-hand user tower receives user features to generate user embeddings, and the right-hand property tower receives property features to generate property embeddings. The inner product is then used to output the ranking score.

[0322] To further utilize the high-order collaborative information in the user-property interaction graph, the graph neural network-enhanced collaborative filtering module introduces a graph neural network (GNN) to aggregate neighbor information for user embeddings and property embeddings, and then fuses it with the output of the dual-tower model.

[0323] Step 1: Construct the user-property interaction graph

[0324] Construct a bipartite graph with user nodes and property nodes, where edges represent interactions between users and properties (clicks, inquiries, etc.).

[0325] Step 2: GNN Information Aggregation

[0326] A Graph Convolutional Network (GCN) is used for message passing, simultaneously updating the representations of user nodes and property nodes. Let the set of user nodes be... The set of housing nodes is ,user With housing The interactions between them form a bipartite graph , among which the side Indicates user Previously with housing resources Positive interactions occur (clicks, favorites, inquiries, etc.).

[0327] User node updates are represented by the formula described in formula (4) above. To avoid repetition, this will not be repeated here. Property node updates are represented by the formula described in formula (4) above. To avoid repetition, this will not be repeated here.

[0328] in:

[0329] The collection of properties that user $u$ has interacted with. Its quantity;

[0330] For housing The set of users who have interacted with the platform Its quantity;

[0331]

[0332] The initial input consists of user embeddings and property listing embeddings:

[0333] That is, for each element in the input vector, if it is positive, it remains unchanged, and if it is negative, it is set to zero. This function is widely used in deep neural networks. Its advantages are simple computation, effective mitigation of the gradient vanishing problem, and introduction of non-linear expressive power.

[0334] The above two equations use symmetric normalization coefficients. The effect of message passing is that when nodes with higher degrees (i.e., users with more interactions or popular listings) aggregate neighbor information, their contribution is appropriately suppressed, preventing popular nodes from dominating the entire representation space, thereby better preserving the characteristic information of long-tail users and long-tail listings.

[0335] Step 3: Integrate Recommendation Scores

[0336] The final recommendation score is obtained by weighted fusion of the inner integral value output by the dual-tower model and the inner product of the node representations aggregated by the GNN. The specific calculation formula is the above formula (5), and will not be repeated here to avoid repetition.

[0337] in:

[0338] Subscript for user identifier, Add a label to the property listing;

[0339] The user requirement embedding vector output by the multimodal fusion module The housing embedding vector output by the housing embedding network;

[0340] and They were respectively through User node representation and property node representation after GCN aggregation at layer 1;

[0341] The fusion coefficient controls the relative weights of the dual-tower direct matching signal and the GNN collaborative filtering signal;

[0342] $L$ represents the number of GCN layers, which is set to 2 layers in this system.

[0343] Regarding the fusion coefficient Selection criteria:

[0344] This determines the extent to which the final recommendation score depends on the direct matching ability of the dual-tower model. (partial) and the collaborative filtering capability of graph neural networks ( (Partial). If Too large (e.g.) If the model degenerates into a pure two-tower model that hardly uses interaction graph information, it cannot fully utilize the collaborative signal that "similar users like similar listings"; if Too small (e.g.) If the model relies too heavily on neighbor information aggregated by GNN, it may cause the user's personalized preferences to be diluted by the group's average preferences (i.e., the "oversmoothing" problem), resulting in decreased recommendation accuracy in scenarios where user behavior is sparse or preferences are unique.

[0345] In offline experiments, this invention uses the normalized loss cumulative gain NDCG@10 as the core evaluation index to assess... exist A grid search was performed within the specified range. The experimental results are as follows:

[0346]

[0347] Experimental data show that: NDCG@10 reached its highest value, representing an improvement of approximately 8.2% compared to the pure twin-tower model. When the value drops to 0.6, due to the increased weights of the GNN, some users experience a dilution of their personalized signals, resulting in a slight decline in the metric. When the value rises to 0.8, the utilization of collaborative information is insufficient, and the improvement narrows. Therefore, the following was selected. As the system default parameter, it indicates that direct matching between the two towers is the primary method (accounting for 70% weight), while high-order collaborative information aggregated by GNN is used as an important supplement (accounting for 30% weight). The two complement each other to achieve the optimal ranking effect.

[0348] Process description: Figure 11 This demonstrates the process of using GNN to enhance collaborative filtering. After multi-layer graph convolution, the user-property interaction graph obtains a representation of the aggregated neighbor information for both user and property nodes. This representation is then fused with the output of the dual-tower model to obtain the final score.

[0349] Key technical parameter setting instructions

[0350] To ensure the full disclosure and feasibility of the technical solution, the following supplementary explanation is provided regarding the selection criteria for the core parameters:

[0351] 1. DBSCAN clustering parameters

[0352] Neighborhood radius ε=50 meters: Based on the spatial scale of "the same community" in urban residential scenarios, and matched with the positioning accuracy of operator location signaling (30-100 meters), it can effectively distinguish between adjacent buildings and external road stops.

[0353] Minimum sample size min_samples=10 times: Based on the statistical threshold of at least 10 stays within 30 days, this effectively filters out accidental stays and ensures the statistical significance of the extracted permanent locations.

[0354]

[0355] Table 1 DBSCAN Clustering Parameters

[0356] 2. Intent Recognition Module Parameters

[0357] Behavior intensity normalization factor =1440 minutes (24 hours): Used to normalize the user's cumulative dwell time to the [0,1] interval, aligned with the dimension of the intent probability p_u.

[0358] Intent balance coefficient α=0.6: Determined through A / B testing, this value achieves a balance between sequence modeling results and behavioral intensity factors, and is the optimal comprehensive index for recommended click-through rate and conversion rate.

[0359]

[0360] Regarding the intention balance coefficient Detailed experimental basis for the values:

[0361] (1) Experimental setup

[0362] Dataset: User behavior logs collected from a platform from September to November 2025, totaling three months, including search, browsing, click, favorite, inquiry and other behavior sequences of approximately 87,000 active users, as well as corresponding exposure and conversion tags.

[0363] Evaluation metrics: Click-through rate (CTR = number of clicks / number of impressions) and final conversion rate (CVR = number of successful contracts / number of impressions) are used as the core evaluation metrics, while also paying attention to the normalized cumulative gain (NDCG@10).

[0364] Comparison range: [To be filled in] exist A grid search is performed using 6 discrete values. The model is trained independently for each value and evaluated on the same test set.

[0365] (2) Experimental results

[0366]

[0367] (3) Selection of the optimal value

[0368] Experimental data show that, At that time, CTR, CVR, and NDCG@10 all reached their optimal levels. When the value increases from 0.5 to 0.6, the CTR increases by approximately 1.2%, and the CVR increases by approximately 2.7%; Increasing the value from 0.6 to 0.7 resulted in a decrease of approximately 2.0% in CTR and approximately 3.9% in CVR. Therefore, 0.6 was selected as the system default parameter.

[0369] (4) The range of suboptimal values ​​and the scope of protection

[0370] As can be seen from the experimental data, when In the interval When taking internal values, the decrease in CTR and CVR relative to the optimal value does not exceed 3%, and the decrease in NDCG@10 does not exceed 2%, indicating that the recommendation effect remains at an acceptable high level (user satisfaction remains above 85%). Specifically:

[0371] when At the time (interpolation), the expected CTR is approximately 2.51%, and the CVR is approximately 0.76%, with similar effects to... Basically equivalent;

[0372] when At the time (interpolation), the expected CTR was approximately 2.50%, and the CVR was approximately 0.75%, with no significant degradation in performance.

[0373] Criteria for determining obvious degradation:

[0374] This invention uses the "difference between recommendation result lists" as an objective threshold to measure parameter degradation. The Jaccard similarity between two recommendation lists is defined as:

[0375] (18)

[0376] in This indicates that the balance coefficient takes the value of At that time, a Top-20 recommended property list is generated for the user. Experimental calculations show that:

[0377]

[0378]

[0379] If the Jaccard similarity of the recommended list is less than 0.70 (i.e., the difference is more than 30%) and the CTR drops by more than 3%, users can clearly perceive a decline in recommendation quality, so it is judged as a significant deterioration in performance.

[0380] In summary, the intended balance coefficient of this invention The optimal value is 0.6, and its acceptable suboptimal range is... Any value within this range that achieves a similar recommended effect falls within the scope of protection of this invention; when When the value exceeds this range (less than 0.55 or greater than 0.65), the recommendation effect will deteriorate significantly.

[0381]

[0382] Table 2 Parameters of the Intent Recognition Module

[0383] 3. Parameters of the multimodal feature fusion module

[0384] DNN with L=3 hidden layers: A balance is achieved between feature representation capability and model complexity. The 3-layer structure achieves the best AUC of 0.84 on the validation set.

[0385] User embedding dimension k=64: A common setting in industrial recommendation systems, achieving the optimal balance between sorting accuracy and computational efficiency.

[0386] Detailed experimental evidence regarding the user embedding dimension $k=64$:

[0387] (1) Experimental setup

[0388] Dataset: The same offline evaluation dataset as the intent recognition module is used, which contains spatiotemporal profile features, behavioral sequence features and corresponding exposure / conversion tags of approximately 87,000 active users.

[0389] Evaluation indicators:

[0390] Ranking accuracy: AUC (Area Under ROC Curve) on the validation set, which measures the model's ranking ability;

[0391] Online inference latency: Under the same hardware environment (Intel Xeon Gold 6248 CPU, no GPU acceleration), the average latency (P99 quantile) of simulating online requests to calculate the inner product of user embedding and property embedding.

[0392] Overall Score: Defining the overall efficiency index This is used to balance the trade-off between accuracy and latency.

[0393] Comparison range: [To be filled in] exist A total of 8 values ​​were used for independent training and evaluation.

[0394] (2) Experimental results

[0395]

[0396] (3) Selection of the optimal value

[0397] Experimental data show that:

[0398] when When the AUC was increased from 32 to 64, it improved from 0.801 to 0.827 (an improvement of 3.2%), with a significant improvement in sorting accuracy, while the inference latency only increased by 6ms, which is still within the low latency requirement of 20ms.

[0399] when When the value was increased from 64 to 128, the AUC only increased from 0.827 to 0.833 (an increase of 0.7%), with diminishing marginal returns. Meanwhile, the inference latency increased from 18ms to 28ms (an increase of 55.6%), exceeding the 20ms latency limit set by the system.

[0400] Comprehensive efficiency indicators exist Although it was not the highest ( Due to extremely low latency The value is relatively high, but the AUC is too low to meet the accuracy requirements. However, it reaches the optimal value under the constraints of AUC≥0.82 and delay≤20ms.

[0401] Therefore, selected As the default value for the user embedding dimension, this value achieves the best balance between sorting accuracy and computational efficiency.

[0402] (4) The range of suboptimal values ​​and the scope of protection

[0403] As can be seen from the experimental data, when In the interval When taking values ​​from the inside, AUC is relatively The variation range does not exceed ±1.5%, and the inference latency is controlled within 22ms. The recommendation effect is basically equivalent to the optimal value. Specifically:

[0404] when At that time, the AUC was 0.818 (a decrease of only 1.1%) and the latency was 15ms (a decrease of 16.7%).

[0405] when At that time, the AUC was 0.830 (an improvement of 0.4%) and the latency was 22ms (an increase of 22.2%).

[0406] Criteria for determining obvious degradation:

[0407] This invention uses two dimensions to comprehensively determine whether the value of the embedding dimension leads to a significant degradation in performance:

[0408] ①Accuracy degradation threshold: When AUC is relatively When the drop exceeds 3% (i.e., AUC < 0.802), the ranking accuracy is significantly insufficient, and users can perceive a decline in recommendation quality.

[0409] ② Latency exceeding the threshold: When the P99 inference latency exceeds 25ms, it exceeds the upper limit of the online service response time set by this system (1.25 times 20ms), affecting the user experience.

[0410] According to the above standards:

[0411] when Time (e.g.) (AUC decreased by 3.2%), accuracy deteriorated beyond the threshold, and performance deteriorated significantly;

[0412] when Time (e.g.) The latency was 25ms, which exceeded the standard and significantly degraded the performance.

[0413] Therefore, user embedding dimension The optimal value is 64, and its acceptable suboptimal range is... Any integer value within this range that achieves similar recommendation accuracy and computational efficiency falls within the scope of protection of this invention; when When the value exceeds this range (less than 48 or greater than 80), the recommendation effect will deteriorate significantly.

[0414]

[0415] Table 3 Parameters of Multimodal Feature Fusion Module

[0416] 4. Graph Neural Network Augmentation Module Parameters

[0417] GCN has 2 layers: It can effectively aggregate second-order neighbor information and avoid oversmoothing while preserving the individual characteristics of nodes.

[0418] Fusion coefficient β=0.7: controls the weight distribution between the output of the dual-tower model and the output of the GNN. At this value, the NDCG@10 index is improved by about 8.2%.

[0419]

[0420] Table 4 shows the parameters of the neural network enhancement module.

[0421] Regarding the fusion coefficient Detailed experimental basis:

[0422] (1) Experimental setup

[0423] Dataset: The same offline evaluation dataset as the aforementioned modules is used, containing exposure / click / conversion logs of approximately 87,000 active users, and a user-listing interaction bipartite graph is constructed for GCN training.

[0424] Evaluation indicators:

[0425] NDCG@10: Normalized depreciation cumulative gain, which measures the ranking quality of the Top-10 of the recommendation list and is the core indicator of this experiment;

[0426] Recall@20: Recall rate, which measures the proportion of properties in the Top-20 recommendation list that actually interact with the user;

[0427] Recommendation list dissimilarity: definition and The Jaccard similarity of the Top-20 recommendation list at that time is used to quantify the degree of change in results caused by parameter deviation.

[0428] Comparison range: [To be filled in] exist A total of 8 values ​​were used for independent training and evaluation. This represents a pure two-tower model that does not use the GNN module. This represents the extreme case where the system is completely dependent on the output of the GNN.

[0429] (2) Experimental results

[0430]

[0431] (3) Selection of the optimal value

[0432] Experimental data show that:

[0433] when At that time, NDCG@10 and Recall@20 both reached their highest values, at 0.197 and 0.280 respectively, compared to the pure two-tower model ( They increased by 8.2% and 7.3% respectively.

[0434] when When the value decreased from 0.7 to 0.5, NDCG@10 decreased from 0.197 to 0.191 (a decrease of about 3.0%), indicating that when the GNN weights were too high, the personalized matching signal was over-diluted, and the ranking quality began to decline.

[0435] when When the value increased from 0.7 to 0.9, NDCG@10 decreased from 0.197 to 0.188 (a decrease of about 4.6%), indicating insufficient utilization of collaborative information and a decline in recommendation accuracy.

[0436] when In the case of (pure GNN), NDCG@10 is only 0.164, which is 9.9% lower than that of the pure dual-tower model, indicating that relying entirely on graph structure while ignoring content matching will lead to serious information loss.

[0437] Therefore, selected As the system default parameter, this value allows the dual-tower direct matching and GNN cooperative signal to achieve the best complementary state.

[0438] (4) Range of suboptimal values ​​and stability of indicators

[0439] As can be seen from the experimental data, when When the value is taken within the range of [0.6, 0.8], the NDCG@10 indicator remains relatively stable, specifically as follows:

[0440]

[0441] The maximum drop in NDCG is defined as the relative decrease between the lowest and highest NDCG values ​​within the interval, specifically 0.197. Within the range, the maximum drop shall not exceed 2.0%; in a more conservative... Within the range, the maximum elevation difference shall not exceed 1.0%. Meanwhile, with... The Jaccard similarity of the recommended lists is no less than 0.85, indicating that the recommendation results are stable.

[0442] therefore, The recommended suboptimal value range is any value within this range that can achieve a recommended effect almost equivalent to the optimal value. The difference between NDCG@10 and the recommended value is controlled within 1%, which is within the protection scope of this invention.

[0443] (5) Criteria for judging obvious deterioration

[0444] This invention makes a comprehensive judgment based on two dimensions. Does the chosen value lead to a significant deterioration in performance?

[0445] ① Sorting accuracy degradation threshold: When NDCG@10 is relative When the decrease exceeds 5% (i.e., NDCG@10 < 0.187), the quality of recommendation ranking deteriorates significantly, and users can perceive a decrease in the relevance of the recommendation results. According to experimental data, at $\beta \leq 0.5$, NDCG = 0.191 (a decrease of 3.0%). At that time, NDCG = 0.188 (a decrease of 4.6%), which is close to or has reached the threshold.

[0446] ② Recommendation list difference threshold: when compared with When the Jaccard similarity of the Top-20 recommendation list is below 0.70 (i.e., the difference exceeds 30%), it indicates that parameter changes have caused a substantial shift in the recommendation strategy, resulting in a loss of stability in the recommendation results. Based on experimental data... At that time, Jaccard = 0.67 (33% difference). At that time, Jaccard = 0.78 (difference 22%). Jaccard = 0.73 (27% difference).

[0447] Taking into account both accuracy and stability requirements, the significant degradation boundary is set as follows:

[0448] when or At that time, the NDCG@10 decreased by more than 3% and the Jaccard similarity was below 0.75, indicating a noticeable and significant deterioration in the recommendation performance.

[0449] Therefore, the fusion coefficient The optimal value is 0.7, and its acceptable suboptimal range is... The scope of protection can be extended to .when Values ​​Exceeding If this is not done, the recommended effect will be significantly degraded, and it will not constitute an effective implementation of the present invention.

[0450] VIII. Complete Implementation Examples

[0451] The following specific embodiment illustrates the implementation process of the present invention in detail.

[0452] Example scenario: A user The user is a China Unicom user who spends his / her days in a technology park and lives in Community A at night. He / she has been frequently accessing rental platforms recently, but has not submitted specific search criteria.

[0453] Step 1: Spatiotemporal Behavioral Profile Construction

[0454] The system collects 30 days of user location signaling data and uses DBSCAN clustering to obtain two main resident clusters:

[0455] Cluster 1: Appears between 22:00 and 06:00, located in Community A. According to the real estate database, Community A is a "mid-range community". The surrounding amenities include a subway station (500 meters away), a shopping mall (1 kilometer away), and a hospital (2 kilometers away).

[0456] Cluster 2: Appears between 09:00 and 18:00, located in B Science Park, and marked as a workplace.

[0457] No school-based activities were detected, therefore .

[0458] Calculate commute time from residence to workplace Minutes (based on real-time traffic).

[0459] Step 2: Intent Recognition

[0460] The user's behavior sequence on the rental platform over the past 7 days (observation period is 7 consecutive calendar days) is as follows: searched for "two-bedroom apartment" 2 times, browsed the listing page 5 times, clicked the details page 3 times, and the total effective time spent on the details page was 12 minutes. The GRU network outputs the intent probability. .

[0461] The behavior intensity factor is calculated according to the defined method:

[0462] Observation period Heaven. Search frequency. The average number of searches per day during the observation period, i.e. Times / day. Divide $f_{search}$ by the upper limit of the experience reference. Search frequency per day (this value is based on platform statistics: over 95% of active users have an average of no more than 20 searches per day), resulting in the normalized search frequency:

[0463]

[0464] Total effective stay This refers to the sum of effective browsing time on the property details page by the user within the observation period. The calculation rule for effective browsing time is: starting from when the user enters the details page, until they leave the details page (closing the page, switching to another application, or no operation on the page for more than 30 seconds), only single dwell times exceeding 3 seconds are included in the cumulative value (filtering out accidental touches). In this embodiment, the user's cumulative effective dwell time over 7 days is 12 minutes. This is divided by the daily normalized reference value. The normalized dwell time is obtained by calculating minutes (i.e., the upper limit of each 24-hour statistical period):

[0465]

[0466] Behavioral intensity factor Defined as the product of normalized search frequency and normalized dwell time:

[0467]

[0468] Final weighted intention score From the sequence intention probability With behavioral intensity factor Perform linear weighted fusion, balance coefficient :

[0469]

[0470] Explanation of alternative search counts for the behavior intensity factor:

[0471] In the definition of behavior intensity factor, normalized search frequency This reflects the frequency of users actively expressing their needs. Besides search counts, this invention also allows for the use of other behaviors that reflect user intent for substitution or joint calculation. Specifically:

[0472] Page views (list page views): The number of times a user browses a list page can reflect their level of engagement in finding a property to some extent, but the intensity of their intent is weaker than that of a proactive search. If page views are used to replace search counts, the normalization reference upper limit needs to be adjusted accordingly (e.g., ...). (times / day), its protection scope is equivalent to that of this scheme.

[0473] Click-through rate: The act of clicking on a details page to view full information; its intent strength lies between searching and browsing. If click-through rate is used, the normalized reference upper limit can be set to... Once / day.

[0474] Weighted combination: Different behavior types such as search, click, favorite, and consultation can be weighted and summed according to their intensity weights (see the behavior intensity weight reference value defined in step 1 of Part 3) to obtain a comprehensive "frequency of proactive intent behavior", which is then normalized by dividing by the corresponding comprehensive reference upper limit.

[0475] The above-mentioned alternative solutions all fall within the protection scope of this invention, as long as their core idea is: to combine the frequency of behaviors reflecting the user's active intent with the effective browsing time to generate a behavior intensity factor with uniform dimensions, and to perform weighted fusion with the intent probability output by the sequence model.

[0476] Explanation regarding normalization truncation:

[0477] When a user's actual behavior data within the observation period exceeds the reference upper limit, the normalized value will be truncated to 1 using the $\min(\cdot,1)$ function to ensure... For example, if a user searches 25 times in a single day (exceeding...) ),but Instead of 1.25, this is to avoid the normalized value exceeding the expected range.

[0478] Step 3: Multimodal fusion

[0479] Feature vector concatenation: Community tier encoding (mid-range = 0.5), supporting facilities vector (transportation = 0.8, commerce = 0.6, medical = 0.4), commuting time (35 minutes, normalized to 0.35), school-age requirement 0, intent score 0.474. User embedding is obtained through a 3-layer DNN. .

[0480] The user spatiotemporal behavior profile features generated in step 1 are concatenated with the intent score output in step 2 to construct the original feature vector. In this embodiment, the feature values ​​and encodings are as follows:

[0481] The neighborhood tier (mid-range) uses one-hot encoding, with a value of 1 corresponding to the mid-range category and 0 for the rest, which is the neighborhood tier vector. (Assuming the quality is divided into three categories: high, medium, and ordinary);

[0482] Supporting demand vector These correspond to normalized scores for three dimensions: transportation convenience, commercial maturity, and medical resources.

[0483] Commuting time After normalization to the maximum reference value (e.g., 120 minutes), the commuting characteristic component is set to... ;

[0484] School-age needs ;

[0485] Intention Score .

[0486] The above components are concatenated to form the original feature vector:

[0487] This represents a vector concatenation operation. The vector is then input into a user embedding network consisting of three fully connected layers, which generates the final user requirement embedding through layer-by-layer nonlinear transformations.

[0488] This embedding vector integrates the semantics of the user's life stage (neighborhood level, commuting tolerance, school district requirements) with the strength of real-time rental intent, providing a unified user representation for subsequent dual-tower matching and GNN-enhanced ranking.

[0489] Step 4: Property Recommendation and GNN Enhancement

[0490] The system embeds candidate housing resources (such as Community C, Community D, etc.) through the housing resource tower. The inner product of the two towers yields an initial score. Simultaneously, the user-property interaction graph, after passing through two layers of GCN, aggregates neighbor information to obtain a GNN-enhanced representation. Finally, after merging and ranking the scores, the top three recommended properties are: Community C (25-minute commute, mid-range community), Community E (40-minute commute, upscale community), and Community F (30-minute commute, mid-range community). Since users have a high tolerance for commuting time (35 minutes) and prefer mid-range communities, the recommendation results align with the users' potential needs.

[0491] Recommended effect verification

[0492] To verify the actual technical effect of the present invention, we conducted an offline comparative experiment on the same dataset, comparing the effect of the present invention with the following baseline methods:

[0493] Baseline 1 (Pure Dual-Tower Model): Only a dual-tower neural network is used for matching user and property features, without introducing spatiotemporal behavioral profile features and GNN enhancement modules.

[0494] Baseline 2 (the LTPM-TRSP method in the paper): The spatiotemporal fusion long-term preference mining method in the POI recommendation field is transferred to the rental recommendation scenario.

[0495] The present invention fully implements the spatiotemporal behavior perception, multimodal fusion, dual-tower matching, and GNN-enhanced ranking mechanism of the present invention.

[0496] The experiment used user behavior log data from September to November 2025, covering approximately 87,000 active users. Click-through rate (CTR), conversion rate (CVR), and normalized depreciation cumulative gain (NDCG@10) were used as the core evaluation metrics. The experimental results are as follows:

[0497]

[0498] Experimental results show that:

[0499] Compared to the pure dual-tower model, the proposed solution improves CTR and CVR by 59.1% and 63.8% respectively, demonstrating that introducing user spatiotemporal behavioral semantic features significantly enhances the accurate matching capability of the recommendation system.

[0500] Compared to the LTPM-TRSP method, which relies solely on spatiotemporal trajectory sequence modeling, this invention improves CTR and CVR by 46.2% and 51.0% respectively through semantic separation of nighttime / daytime / school permanent locations and semantic mapping of residences, verifying the actual gain of the technological leap "from geographic coordinates to life scene semantics".

[0501] The GNN enhancement module improved NDCG@10 by 8.2% compared to the pure dual-tower model, demonstrating that the introduction of higher-order collaborative information from the user-property interaction graph effectively optimized the recommendation ranking quality.

[0502] The above quantitative comparative data strongly demonstrates that the present invention can make full use of operator spatiotemporal data and user behavior sequences to achieve in-depth mining of users' implicit residential preferences and accurate identification of life scenarios, and finally output personalized housing recommendation results, which has clear technological progress and industrial applicability.

[0503] In summary, this application has the following key points and protection points.

[0504] I. Spatiotemporal behavioral semantic modeling method based on operator location signaling

[0505] Key points: Utilize operator location signaling data, extract user's permanent location clusters through DBSCAN clustering algorithm, and automatically identify user's residence, workplace, and school-age family needs based on time period rules (22:00–06:00 at night, 09:00–18:00 during the day, and 08:00–16:00 during school hours on consecutive weekdays) to generate multi-dimensional life stage tags.

[0506] Protection point:

[0507] A method for identifying the grade of a residential community and constructing a profile of its supporting facilities based on nighttime frequent locations;

[0508] A method for identifying workplaces and modeling commuting preferences based on daytime locations;

[0509] Based on the behavior of school staff, we can determine the needs of school-age families and generate school district preference labels.

[0510] Beneficial effects: It solves the problem that existing technologies can only obtain the user's current location or simple geographic location, and cannot identify the semantics of the user's long-term life scenarios.

[0511] II. A Multimodal User Demand Identification Mechanism Integrating Location Semantics and Behavioral Intent

[0512] Key point: Multimodal fusion of user spatiotemporal behavioral data (residential grade, supporting facilities profile, commuting time, school age tag) and digital behavioral trajectory data (rental platform visits, clicks, searches, favorites, inquiries, etc.) to construct a multidimensional vector of user housing needs.

[0513] Protection point:

[0514] A multimodal user demand identification mechanism that integrates location semantics and behavioral intent;

[0515] A user rental intent scoring method based on a weighted fusion of behavioral sequences (GRU) and behavioral intensity (search frequency, dwell time).

[0516] Beneficial effects: It solves the problems of incomplete multi-source data fusion and single user profile dimensions in existing technologies, and enables joint modeling of users' explicit behaviors and implicit life scenarios.

[0517] III. A User Needs Profile Construction Method Driven by "Life Stage + Behavioral Intent"

[0518] Key points: Identify users' life stages (such as working professionals, school-age families, and retirees) through spatiotemporal behavioral profiles, and combine this with an intent recognition model based on behavioral sequences to construct a dual-driven profile of users' residential needs.

[0519] Protection point:

[0520] A collaborative modeling mechanism for user life stage (semantic properties of residence, semantic properties of workplace, and tags of school-age families) and behavioral intentions;

[0521] The feature cross-fusion method based on deep neural networks (DNN) nonlinearly combines life stage features and behavioral intention features to generate user demand embeddings.

[0522] Beneficial effects: It solves the problem that existing technologies cannot distinguish the different living needs of users at different stages of life, and realizes truly personalized scenario-based recommendations.

[0523] IV. A Housing Recommendation and Ranking Model Based on Graph Neural Network Enhancement

[0524] Key takeaway: Based on the dual-tower recommendation ranking model, a graph neural network (GCN) is introduced to perform high-order information aggregation on the user-property interaction graph, and the GNN output is weighted and fused with the inner product result of the dual-tower model to improve the accuracy of recommendation ranking.

[0525] Protection point:

[0526] Based on a collaborative filtering mechanism enhanced by graph neural networks, neighbor information in the user-property interaction graph is aggregated through multi-layer graph convolution.

[0527] A recommended score calculation method that combines the inner product of two towers with the weighted fusion of GNN output.

[0528] Beneficial effects: It solves the problems of insufficient utilization of user-property interaction graphs and limited ranking accuracy in existing recommendation models, and makes full use of collaborative filtering information to improve recommendation performance.

[0529] V. Complete Technical System Architecture

[0530] Protection Point: A complete intelligent rental property recommendation system consisting of the following five modules:

[0531] User spatiotemporal behavior profile building module;

[0532] User rental intent recognition module;

[0533] Multimodal user demand vector fusion module;

[0534] Property Features and Recommendation Ranking Module (Dual Tower Model);

[0535] A collaborative filtering module enhanced by graph neural networks.

[0536] Beneficial effects: The above modules work together to form a complete technical chain of "spatiotemporal behavior perception → multimodal fusion → dual-wheel drive profiling → GNN enhanced ranking", realizing full-process automation from data collection to accurate recommendation.

[0537] VI. Explanation of the Independent Protection Scope of Each Key Technology Point

[0538]

[0539] like Figure 12 As shown, the implementation process of the housing recommendation method provided in this application embodiment includes the following S1 to S9:

[0540] S1. Collect 30 days of user location signaling data, obtain two main persistent point clusters through DBSCAN clustering, and construct a user spatiotemporal behavior profile based on the persistent point clusters and the life scene semantic tags corresponding to each persistent point cluster.

[0541] S2. Based on the user's behavior sequence on the rental platform over the past 7 days (the observation period is 7 consecutive natural days), identify intent and output intent information;

[0542] S3. Combine the user's spatiotemporal behavioral profile with intent information to obtain user feature information;

[0543] S4. Input the user feature information and the attribute feature information of each candidate property into the dual-tower model, and perform inner product operation to obtain the first recommendation score of each candidate property.

[0544] S5. Construct a user-property interaction graph based on the interaction data between user nodes and each candidate property node;

[0545] S6. Use a graph convolutional network (GCN) to aggregate neighbor information in the user-property interaction graph to obtain the node representation after GNN aggregation;

[0546] S7. The inner integral value output by the dual-tower model is weighted and fused with the inner integral value of the node representation after aggregation by GNN to obtain the second recommendation score for each candidate property.

[0547] S8. Based on the first recommendation score and the second recommendation score of each candidate property, obtain the target recommendation score for each candidate property.

[0548] S9. Based on the target recommendation score of each candidate property, recommend properties to the target user.

[0549] In this way, by transforming user location information into semantic tags that more intuitively reflect users' living preferences, and by using historical data of user interactions with each candidate property and the property attribute information of each candidate property, a target recommendation score is obtained for each candidate property. This recommendation score is then used to more accurately recommend properties to users, thereby improving the accuracy of property recommendations.

[0550] It should be noted that the descriptions of each step S1 to S9 in this embodiment can be found in the descriptions in the above embodiments, and will not be repeated here.

[0551] It should be noted that the above-described method embodiments, or the various possible implementations of the method embodiments, can be executed individually, or, provided there is no conflict, they can be combined with each other. The specific implementation can be determined according to actual usage requirements, and this application embodiment does not impose any restrictions on this.

[0552] Figure 6 This is a schematic diagram of a housing recommendation system provided in an embodiment of this application. Figure 6 As shown, the housing recommendation system 600 may include: a user spatiotemporal behavior profile construction module 601, a user rental intention recognition module 602, a multimodal user demand vector fusion module 603, a housing feature and recommendation ranking module 604, a graph neural network-enhanced collaborative filtering module 605, a housing feature extraction module 606, and a neighborhood grade / school district mapping module 607.

[0553] Among them, the user spatiotemporal behavior profile construction module 601 and the community grade / school district mapping module 607 are used to construct the user spatiotemporal behavior profile and are applied to the relevant scheme of the above step S1.

[0554] The user rental intent recognition module 602 is used to recognize the user's rental intent and is applied to the relevant scheme in step S2 above.

[0555] The multimodal user demand vector fusion module 603 is used to splice user spatiotemporal behavior profiles and intent information to obtain user feature information, which is then applied to the relevant scheme in step S3 above.

[0556] The property feature and recommendation ranking module 604 is used to obtain the first recommendation score for each candidate property and apply it to the relevant scheme in step S4 above.

[0557] The graph neural network-enhanced collaborative filtering module 605 is used to obtain the target recommendation score for each candidate property and is applied to the relevant schemes in steps S5 to S9 above.

[0558] The property feature extraction module 606 is used to extract the property attribute information of each candidate property.

[0559] It should be noted that for a detailed explanation of the steps performed by each module and their beneficial effects, please refer to the description in the above embodiments, which will not be repeated here.

[0560] As can be seen, the above mainly describes the solutions provided by the embodiments of this application from a methodological perspective. To achieve the above functions, the embodiments of this application provide corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the modules and algorithm steps of the various examples described in the embodiments disclosed herein, the embodiments of this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0561] This application embodiment can divide the housing recommendation device into functional modules according to the above method example. For example, each function can be divided into a separate functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. Optionally, the module division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0562] In some embodiments, this application also provides a property recommendation device. This property recommendation device may include one or more functional modules for implementing the property recommendation method described in the above embodiments.

[0563] For example, Figure 13 This is a schematic diagram of a housing recommendation device provided in an embodiment of this application. Figure 13 As shown, the property recommendation device 1300 includes: a determination module 1301, a generation module 1302, and a recommendation module 1303.

[0564] The aforementioned determining module 1301 is used to determine at least one cluster of permanent points of the target user and the semantic tags of life scenarios corresponding to each cluster of permanent points based on the user spatiotemporal behavior data of the target user within a predetermined time period.

[0565] The aforementioned determining module 1301 is also used to determine the housing transaction intention information of the aforementioned target user based on the behavioral data of the aforementioned target user on the housing transaction platform;

[0566] The aforementioned generation module 1302 is used to generate user feature information of the target user based on the aforementioned user spatiotemporal behavior data, the semantic tags of life scenarios corresponding to each of the aforementioned permanent point clusters, and the aforementioned housing transaction intent information.

[0567] The aforementioned generation module 1302 is also used to input the aforementioned user feature information and the attribute feature information of each candidate property into the dual-tower model to obtain the first recommendation score for each candidate property.

[0568] The aforementioned recommendation module 1303 is used to recommend properties to the target user based on the first recommendation score of each candidate property.

[0569] The housing recommendation device provided in this application determines at least one cluster of permanent locations of the target user based on the user's spatiotemporal behavior data within a predetermined time period, and marks each cluster of permanent locations with a corresponding semantic tag for a living scenario. Then, the semantic tag for the living scenario corresponding to each cluster of permanent locations is fused with the housing transaction intention information tag, so that the obtained user feature information can reflect the type of living scenario corresponding to the user at different permanent locations. Therefore, when recommending housing based on the user feature information and the attribute feature information of candidate housing, housing that is more in line with the target user's living scenario can be recommended, thereby improving the accuracy of housing recommendation.

[0570] In some embodiments, the user spatiotemporal behavior data includes at least one location information of the target user within the predetermined time period, and each location information includes a time point and the coordinates of the target user's location at that time point.

[0571] The aforementioned determining module 1301 is specifically used for:

[0572] Clustering is performed on at least one location information of the target user within the aforementioned predetermined time period to obtain at least one cluster of permanent points.

[0573] Based on at least one preset semantic judgment rule, the first life scene semantic label contained in the first semantic judgment rule satisfied by the first permanent point cluster is determined as the life scene semantic label corresponding to the first permanent point cluster; the first permanent point cluster is one of the at least one permanent point cluster.

[0574] Each semantic judgment rule corresponds to a time period and a semantic label for a life scenario.

[0575] In other embodiments, the aforementioned recommended module 1303 is specifically used for:

[0576] The user-property interaction graph is aggregated using a graph neural network (GNN) to update the user feature information corresponding to the user node and the property feature information corresponding to the property node in the user-property interaction graph. The user-property interaction graph is generated based on the interaction data between at least one user and at least one candidate property.

[0577] Based on the updated user feature information corresponding to the user node and the property feature information corresponding to the property node in the above user-property interaction graph, a second recommendation score is obtained for each candidate property.

[0578] Based on the first and second recommendation scores of each candidate property, the target recommendation score for each candidate property is obtained;

[0579] Based on the target recommendation score for each candidate property, properties are recommended to the aforementioned target users.

[0580] In some other embodiments, the above-mentioned generation module 1302 is specifically used for:

[0581] The aforementioned user feature information is input into the user embedding network in the aforementioned dual-tower model for feature encoding, and the user feature vector is output.

[0582] The attribute feature information of each candidate property is input into the property embedding network in the above dual-tower model for feature encoding, so as to obtain the property feature vector of each candidate property.

[0583] Based on the user feature vector and the property feature vector of each candidate property, the first recommendation score for each candidate property is output.

[0584] It should be noted that the housing recommendation device can implement all the processes implemented in the above method embodiments and achieve the same beneficial effects. To avoid repetition, it will not be described again here.

[0585] In the case where the functions of the integrated modules described above are implemented in hardware, this application provides a possible structural schematic diagram of the electronic device involved in the above embodiments. For example... Figure 14 As shown, the electronic device 140 includes: a processor 142, a communication interface 143, and a bus 144. Optionally, the electronic device 140 may also include a memory 141.

[0586] Processor 142 may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 142 may be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 142 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0587] Communication interface 143 is used to connect with other devices via a communication network. This communication network can be Ethernet, wireless access network, wireless local area network (WLAN), etc.

[0588] The memory 141 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0589] As one possible implementation, the memory 141 can exist independently of the processor 142. The memory 141 can be connected to the processor 142 via a bus 144 and is used to store instructions or program code. When the processor 142 calls and executes the instructions or program code stored in the memory 141, it can implement the housing recommendation method provided in the embodiments of this application.

[0590] In another possible implementation, the memory 141 can also be integrated with the processor 142.

[0591] Bus 144 can be an Extended Industry Standard Architecture (EISA) bus, etc. Bus 144 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 14 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0592] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the service calling device can be divided into different functional modules to complete all or part of the functions described above.

[0593] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described housing recommendation method embodiment and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0594] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0595] This application embodiment also provides a readable storage medium storing a program or instructions, which, when executed by a computer, implements the property recommendation method provided in the above embodiments. It is understood that all or part of the processes in the above method embodiments can be executed by computer instructions instructing related hardware; the readable storage medium can be any of the foregoing embodiments or memory; the readable storage medium can also be an external storage device of the service invocation device, such as a pluggable hard drive, Smart MediaCard (SMC), Secure Digital (SD) card, flash card, etc., equipped on the service invocation device. Further, the readable storage medium can include both internal storage units of the service invocation device and external storage devices. The readable storage medium is used to store the computer program and other programs and data required by the service invocation device. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0596] This application also provides a computer program product, which is stored in a storage medium and, when executed by a computer, implements the housing recommendation method provided in the above embodiments.

[0597] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0598] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0599] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for recommending housing listings, characterized in that, The property recommendation methods include: Based on the spatiotemporal behavior data of the target user within a predetermined time period, at least one cluster of permanent locations of the target user and semantic tags of life scenarios corresponding to each cluster of permanent locations are determined. Based on the target user's behavioral data on the housing transaction platform, the target user's housing transaction intention information is determined; Based on the user's spatiotemporal behavior data, the semantic tags of life scenarios corresponding to each cluster of permanent locations, and the housing transaction intent information, the user feature information of the target user is generated. The user feature information and the attribute feature information of each candidate property are input into the dual-tower model to obtain the first recommendation score for each candidate property. Based on the first recommendation score of each candidate property, properties are recommended to the target user.

2. The housing recommendation method according to claim 1, characterized in that, The user spatiotemporal behavior data includes at least one location information of the target user within the predetermined time period, and each location information includes a time point and the coordinates of the target user's location at that time point; The process of determining at least one cluster of persistent points for the target user and semantic tags for each life scenario based on the user's spatiotemporal behavior data within a predetermined time period includes: Clustering is performed on at least one location information of the target user within the predetermined time period to obtain the at least one cluster of persistent points; Based on at least one preset semantic judgment rule, the first life scene semantic tag contained in the first semantic judgment rule satisfied by the first permanent point cluster is determined as the life scene semantic tag corresponding to the first permanent point cluster; the first permanent point cluster is one of the at least one permanent point cluster. Each semantic judgment rule corresponds to a time period and a semantic label for a life scenario.

3. The housing recommendation method according to claim 2, characterized in that, Each semantic judgment rule includes at least one of the following: the time period definition corresponding to the semantic judgment rule, the scene description information of the life scenario to which the time period applies, the judgment conditions, and the corresponding life scenario semantic tags.

4. The housing recommendation method according to claim 1, characterized in that, The process of recommending properties to the target user based on the first recommendation score of each candidate property includes: The user-property interaction graph is aggregated using a graph neural network (GNN) to update the user feature information corresponding to the user node and the property feature information corresponding to the property node in the user-property interaction graph. The user-property interaction graph is generated based on the interaction data between at least one user and at least one candidate property. Based on the updated user feature information corresponding to the user node and the property feature information corresponding to the property node in the user-property interaction graph, a second recommendation score is obtained for each candidate property. Based on the first and second recommendation scores of each candidate property, the target recommendation score for each candidate property is obtained; Based on the target recommendation score of each candidate property, properties are recommended to the target user.

5. The housing recommendation method according to claim 1, characterized in that, The step of inputting the user feature information and the attribute feature information of each candidate property into the dual-tower model to obtain the first recommendation score for each candidate property includes: The user feature information is input into the user embedding network in the dual-tower model for feature encoding, and the user feature vector is output. The attribute feature information of each candidate property is input into the property embedding network in the dual-tower model for feature encoding to obtain the property feature vector of each candidate property. Based on the user feature vector and the property feature vector of each candidate property, the first recommendation score for each candidate property is output.

6. A housing recommendation device, characterized in that, The property recommendation device includes: The determination module is used to determine at least one cluster of permanent locations of the target user and the semantic tags of life scenarios corresponding to each cluster of permanent locations based on the user spatiotemporal behavior data of the target user within a predetermined time period. The determining module is further configured to determine the target user's housing transaction intent information based on the target user's behavioral data on the housing transaction platform; The generation module is used to generate user feature information of the target user based on the user's spatiotemporal behavior data, the life scene semantic tags corresponding to each permanent point cluster, and the housing transaction intent information; The generation module is also used to input the user feature information and the attribute feature information of each candidate property into the dual-tower model to obtain the first recommendation score for each candidate property. The recommendation module is used to recommend properties to the target user based on the first recommendation score of each candidate property.

7. The housing recommendation device according to claim 6, characterized in that, The user spatiotemporal behavior data includes at least one location information of the target user within the predetermined time period, and each location information includes a time point and the coordinates of the target user's location at that time point; The determining module is specifically used for: Clustering is performed on at least one location information of the target user within the predetermined time period to obtain the at least one cluster of persistent points; Based on at least one preset semantic judgment rule, the first life scene semantic tag contained in the first semantic judgment rule satisfied by the first permanent point cluster is determined as the life scene semantic tag corresponding to the first permanent point cluster; the first permanent point cluster is one of the at least one permanent point cluster. Each semantic judgment rule corresponds to a time period and a semantic label for a life scenario.

8. The housing recommendation device according to claim 7, characterized in that, Each semantic judgment rule includes at least one of the following: the time period definition corresponding to the semantic judgment rule, the scene description information of the life scenario to which the time period applies, the judgment conditions, and the corresponding life scenario semantic tags.

9. The housing recommendation device according to claim 6, characterized in that, The recommendation module is specifically used for: The user-property interaction graph is aggregated using a graph neural network (GNN) to update the user feature information corresponding to the user node and the property feature information corresponding to the property node in the user-property interaction graph. The user-property interaction graph is generated based on the interaction data between at least one user and at least one candidate property. Based on the updated user feature information corresponding to the user node and the property feature information corresponding to the property node in the user-property interaction graph, a second recommendation score is obtained for each candidate property. Based on the first and second recommendation scores of each candidate property, the target recommendation score for each candidate property is obtained; Based on the target recommendation score of each candidate property, properties are recommended to the target user.

10. The housing recommendation device according to claim 6, characterized in that, The generation module is specifically used for: The user feature information is input into the user embedding network in the dual-tower model for feature encoding, and the user feature vector is output. The attribute feature information of each candidate property is input into the property embedding network in the dual-tower model for feature encoding to obtain the property feature vector of each candidate property. Based on the user feature vector and the property feature vector of each candidate property, the first recommendation score for each candidate property is output.

11. An electronic device, characterized in that, It includes a processor and a memory, the memory storing programs or instructions that can run on the processor, the programs or instructions being executed by the processor to implement the property recommendation method as described in any one of claims 1-5.

12. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a computer, implement the property recommendation method as described in any one of claims 1-5.

13. A computer program product, characterized in that, The computer program product is stored in a storage medium, and when executed by a computer, the computer program product implements the property recommendation method as described in any one of claims 1-5.

Images