Information recommendation method, device, storage medium, and program product

Abstract

Embodiments of the present application provide an information recommendation method and device, a storage medium and a program product. In the embodiments of the present application, a target geographical object associated with a user and attribute data thereof in a target library are introduced, geographical semantic features, distance features and geographical context features are included in a user modeling process, and a data paradigm of only relying on established behavior data and static portraits is broken through; on this basis, steady-state demand potential information and dynamic interest intention information of the user are predicted according to attribute data and historical behavior data of the target geographical object respectively, and the user demand features are described from two dimensions of long-term steady-state demand and short-term dynamic interest; further, the steady-state demand potential information and the dynamic interest intention information are combined with candidate information objects in a target application scene to jointly recommend information, so that the recommended result can not only fit the current interest of the user, but also maintain consistency with the long-term demand of the user, and the recommendation effect is greatly improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an information recommendation method, device, storage medium, and program product. Background Technology

[0002] In the e-commerce sector, personalized recommendation systems based on user profiles and historical behavioral data have become a core infrastructure for improving platform operational efficiency and user experience. Current mainstream recommendation technologies, whether collaborative filtering, content analysis, or more complex deep learning hybrid models, largely depend on the dimensions of the input data, such as dynamic behavioral sequences and static basic profiles.

[0003] Dynamic behavioral sequences, such as user clicks, browsing, searching, and purchase records, constitute the most direct expression of user intent and are the core source for recommendation systems to capture user intent. Static basic profiles, such as attribute information in user registration information or interest tags inferred from behavior, can provide relatively stable prior knowledge for recommendation systems.

[0004] However, while these data dimensions form the foundation of recommender system capabilities, they also solidify their cognitive boundaries. This deep reliance on limited datasets causes recommender performance to hit a difficult-to-break "ceiling," forming a technological barrier to its development.

[0005] However, this deep reliance on fixed data dimensions has also led to the "involution" of recommender systems. Within the existing data paradigm, the optimization space for recommender systems is increasingly narrowing, making it difficult to achieve a leap in recommendation performance. Summary of the Invention

[0006] This application provides an information recommendation method, device, storage medium, and program product, which breaks through the traditional data paradigm used for recommendation and greatly improves the recommendation effect of the recommendation system.

[0007] This application provides an information recommendation method, comprising: acquiring user profile data and historical behavior data, wherein the profile data includes identification information of target geographic objects associated with the user; acquiring attribute data of the target geographic objects from a target library based on the identification information, wherein the target library is obtained by augmenting the geographic object description information in an ontology library with geographic semantic features, distance features, and geographic context features based on geographic object description information from a third-party library, wherein the geographic object description information includes attribute data of the corresponding geographic object; predicting the user's steady-state demand potential information based on the attribute data of the target geographic object, and predicting the user's dynamic interest intention information based on the historical behavior data; and performing joint information recommendation for the user based on the user's steady-state demand potential information and dynamic interest intention information, combined with candidate information objects in the target application scenario.

[0008] This application also provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor is coupled to the memory to execute the computer program to implement the steps in the methods described above.

[0009] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps in the methods described above.

[0010] This application also provides a computer program product, which includes a computer program / instructions that, when executed by a processor, enable the processor to implement the steps described in the above method embodiments.

[0011] In this embodiment, the target geographic object associated with the user and its attribute data in the target database are introduced. Geographic semantic features, distance features, and geographic context features are incorporated into the user modeling process, breaking through the data paradigm that relies solely on predetermined behavioral data and static profiles. Based on this, the steady-state demand potential information and dynamic interest intention information of the user are predicted according to the attribute data and historical behavioral data of the target geographic object, respectively, characterizing the user's demand features from two dimensions: long-term steady-state demand and short-term dynamic interest. Furthermore, the steady-state demand potential information and dynamic interest intention information are combined with candidate information objects in the target application scenario for joint information recommendation, so that the recommendation results can not only match the user's current interests but also maintain consistency with their long-term needs, greatly improving the recommendation effect. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of a first application scenario of a geographic object value assessment method provided in the first embodiment of this application; Figure 2 This is a schematic diagram of a second application scenario of a geographic object value assessment method provided in the first embodiment of this application; Figure 3 This is a flowchart of a geographic object value assessment method provided in the first embodiment of this application; Figure 4 This is a schematic diagram of the fitting curve in a geographic object value assessment method provided in the first embodiment of this application; Figure 5 This is a schematic diagram of a geographic object value assessment device provided in the second embodiment of this application; Figure 6 This is a schematic diagram of an electronic device provided in the third embodiment of this application; Figure 7 This is a flowchart illustrating an information recommendation method provided in an embodiment of this application; Figure 8 is a flowchart illustrating another information recommendation method provided in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] It should be noted that, in the cases involving user information in the embodiments of this application, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to language models or large models) comply with relevant laws and standards.

[0015] Geographic object valuation (such as real estate) refers to the assessment of resource data required for ownership transfer events of geographic objects. It is a key technology for realizing asset digitization, and the accuracy of geographic object valuation is crucial for geographic object transactions and an important prerequisite for the digitization of asset transactions. Currently, there are two main methods for assessing the value of geographic objects: one is manual calculation, which is not only time-consuming and costly, but also prone to significant differences in assessment due to the experience and preferences of different appraisers; the other is to use geographic object valuation models to estimate the value of geographic objects in a digital form. While this method can improve the accuracy of assessment compared to manual calculation, it requires the collection of a large amount of unordered related data, and the collected data varies from user to user. This results in a method that is not only computationally intensive and time-consuming, but also prone to significant discrepancies between the valuation results and the actual transaction results, as well as between the results obtained by different users, leading to excessively long transaction cycles or even failure to complete the transaction. Therefore, how to improve the efficiency of geographic object valuation while reducing computational resources is a pressing issue that needs to be addressed.

[0016] Based on this, this application provides a method for assessing the value of a geographic object. When obtaining the value assessment result of a target geographic object, this method comprehensively considers the preliminary value assessment result determined based on the attribute data of the target geographic object, the statistical value of the historical transaction prices of all similar geographic objects similar to the target geographic object, the predicted value of the current assessment result of the region determined by the fitting curve of the historical transaction prices and time relationship of other geographic objects in the region where the target geographic object is located at different times, the trend change coefficient determined by the ratio of the predicted value to the historical transaction price corresponding to the previous time unit on the fitting curve, and the impact of the missing number of similar geographic objects on the value assessment result of the target geographic object. This improves the accuracy of the value assessment of the target geographic object. Moreover, the data required for the assessment calculation is ordered and deterministic, which reduces the consumption of computing resources, improves the efficiency of the assessment calculation, and enhances the stability of the calculation results.

[0017] The following explains the technical terms used in the embodiments of this application: Robustness refers to the stability and adaptability of a system, model, or method in the face of uncertainty, disturbances, or changes. A robust system can maintain its performance under different conditions without being affected by external factors. Robustness can be applied in many fields, including engineering, statistics, computer science, and artificial intelligence.

[0018] Unsupervised learning is a machine learning paradigm that differs from supervised learning primarily in that it does not use labeled data. In unsupervised learning, algorithms learn from unlabeled data, attempting to discover patterns or structures within the data. Common tasks in unsupervised learning include clustering, dimensionality reduction, and anomaly detection.

[0019] Time series analysis is a dataset that records a series of observations in chronological order. These observations can be any measurable variable, such as stock prices, temperature changes, website traffic, etc. Time series analysis has wide applications in many fields, including but not limited to economics, finance, meteorology, engineering, and social sciences.

[0020] The following combination Figures 1-2 The application scenarios of the geographic object value assessment method provided in the embodiments of this application are described in detail.

[0021] Please see Figure 1 , Figure 1 This is a schematic diagram of a first application scenario of a geographic object value assessment method provided in the first embodiment of this application. In this embodiment, a housing unit is used as an example to describe the value assessment of the housing unit. First, the attribute data of the housing unit is obtained. This attribute data may specifically be information about the community where the housing unit is located, such as the number of buildings in the community, the construction year of the community, and the greening information of the community. Then, housing unit information is obtained, such as the floor where the housing unit is located, the housing unit area, and the historical transaction prices of other housing units in the community. Finally, the geographic information of the housing unit also needs to be obtained, such as the district information of the housing unit, and whether there are supporting facilities such as subway, park, and school around the housing unit. Based on the above attribute data of the housing unit, a preliminary value assessment result of the housing unit is obtained through a preset regression model, which is the first result. Then, clustering information is obtained through a clustering model to obtain similar housing units with similar attribute data to the housing unit to be assessed. Furthermore, the historical transaction prices of all similar housing units are obtained, and the statistical value of all historical transaction prices is calculated as the second result of similarity prediction. The missing value of similar properties is determined based on the number of similar properties and the number of transactions in the area where the property to be evaluated is located. Then, based on the prices in the property's neighborhood, commercial district, zoning area, and city, a time-series model is used to obtain the property's trend change coefficient. Finally, using the missing value as an adjustment coefficient and the product of the second result and the trend change coefficient as the adjustment amount, the first result is adjusted to obtain the final property valuation result.

[0022] In practice, a first supplementary result can be obtained based on the category and attribute data of the property to be evaluated, and a second supplementary result can be obtained based on the city and attribute data of the property to be evaluated. Finally, the valuation result is integrated with the first and second supplementary results to obtain the overall valuation result of the property. Please see [link to relevant documentation]. Figure 2 , Figure 2 This is a schematic diagram of a second application scenario of a geographic object valuation method provided in the first embodiment of this application. In this embodiment, geographic objects can be classified, for example, the type of geographic object can be a residence, apartment, villa, or garage. They can also be classified by the city to which they belong, for example, whether the geographic object is in a first-tier or second-tier city, or in East China or Northwest China. It should be noted that each type corresponds to a sub-model. By inputting the attribute data of the above different types into the corresponding sub-model, additional valuation results for the target geographic object can be obtained. Based on the value valuation results and the additional valuation results, the overall valuation result of the property is obtained. Of course, in specific implementations, the overall valuation result of the property can also be corrected. For example, in the event of sudden market fluctuations, the degree of impact of market fluctuations on the value valuation result can be determined as a correction coefficient. This correction coefficient can be used to correct the above value valuation result or the overall valuation result, thereby making the final estimated result of the property more accurate.

[0023] The geographic object valuation method described in this application mainly comprises three parts: an attribute layer, a model layer, and a strategy layer. The attribute layer is fundamental, capable of segmenting properties based on multiple dimensions such as residential attributes and geographical location to select appropriate valuation models. Then, different types of properties are routed to the most suitable master and sub-models, thereby improving the accuracy and efficiency of valuation. Residential attributes include building year, area, apartment type, and floor, reflecting the basic situation of the property and being important factors in assessing its value. Geographical location mainly considers the neighborhood, district, and even city where the property is located, which has a significant impact on property prices; for example, areas with convenient transportation and bustling commercial areas. The model layer is the core part, consisting of a master valuation model and multiple sub-models. The master valuation model, as the core model, integrates multiple key factors, such as historical transactions, similar properties, and market trends, aiming to provide a preliminary and comprehensive valuation reference. Sub-models carefully consider the unique attributes of specific types of properties. In a complex real estate market, a single model is insufficient to adapt to all situations. Therefore, multiple sub-models are customized based on different residential attributes and geographical locations. For example, the apartment sub-model might specifically consider the unique attributes of apartments, while the city sub-model focuses more on the city tier to which the property belongs. Although each sub-model may be structurally similar, they will differ significantly in their handling of specific factors to ensure more accurate predictions for specific property types. The strategy layer primarily processes the model results, further refining the valuation results output by the model layer to ensure the system's robustness and flexibility. The model's valuation results may be affected by various non-normative factors in practical applications, such as sudden market fluctuations. The strategy layer handles these unexpected factors through various methods (such as weighted averaging and smoothing) to ensure the stability of the valuation results, thereby enhancing robustness. When encountering unexpected situations, such as the addition or removal of amenities around a property, the strategy layer can quickly adjust relevant parameters or adopt contingency strategies to promptly correct the valuation results. This rapid response mechanism ensures that the valuation system can adapt to dynamically changing market environments, improving flexibility.

[0024] The above is a complete analysis of the geographic object valuation method described in the embodiments of this application. The geographic object valuation method described in the embodiments of this application includes: determining a target geographic object; obtaining attribute data of the target geographic object, determining a preliminary valuation result of the target geographic object based on the attribute data, as a first result; determining at least one similar geographic object of the target geographic object, obtaining the historical transaction prices of all similar geographic objects, and calculating the statistical value of all historical transaction prices, as a second result of similarity prediction; obtaining the historical transaction prices of other geographic objects in the area where the target geographic object is located at different times, determining the predicted value of the current valuation result of the area based on the fitting curve determined by the relationship between the historical transaction prices and time, as a third result, calculating the ratio of the third result to the historical transaction price corresponding to the previous time unit on the fitting curve, as a trend change coefficient; determining the missing degree of the similar geographic objects based on the number of similar geographic objects of the target geographic object and the number of transaction objects in the area where the target geographic object is located; using the missing degree as an adjustment coefficient, and the product of the second result and the trend change coefficient as an adjustment amount, adjusting the first result to obtain the valuation result of the target geographic object.

[0025] The geographic object valuation method described in this application comprehensively considers the preliminary valuation result determined based on the attribute data of the target geographic object, the statistical value of the historical transaction price of all similar geographic objects similar to the target geographic object, the predicted value of the current valuation result of the region determined by the fitting curve of the historical transaction price and time relationship of other geographic objects in the region where the target geographic object is located at different times, the trend change coefficient determined by the ratio of the predicted value to the historical transaction price corresponding to the previous time unit on the fitting curve, and the impact of the missing number of similar geographic objects on the valuation result of the target geographic object when obtaining the valuation result of the target geographic object. This improves the accuracy of the valuation of the target geographic object. Moreover, the data required for the valuation calculation is ordered and deterministic, which reduces the consumption of computing resources, improves the efficiency of the valuation calculation, and enhances the stability of the calculation results.

[0026] The present application will be described in detail below with reference to several embodiments and accompanying drawings.

[0027] First Embodiment The first embodiment of this application provides a method for assessing the value of geographic objects, which is described below in conjunction with... Figure 3 The method for assessing the value of this geographical object is explained in detail.

[0028] Step S301: Determine the target geographic object.

[0029] This step is used to identify the target geographic object. Taking a property as an example, the geographic object can be a residence, apartment, villa, townhouse, etc. The target geographic object refers to the geographic object whose value needs to be evaluated.

[0030] Step S302: Obtain the attribute data of the target geographic object, and determine the preliminary value assessment result of the target geographic object based on the attribute data, as the first result.

[0031] This step is used to obtain attribute data of the target geographic object, and to determine the preliminary value assessment result of the target geographic object based on the attribute data, which is taken as the first result.

[0032] In this embodiment, the attribute data includes the feature data of the geographic object, the data of the defined region to which the geographic object belongs, the geographic information of the geographic object, and the surrounding ancillary information data. Specifically, the geographic information of the geographic object is the relationship information between the target geographic object and its surrounding associated geographic objects; the surrounding associated geographic objects are geographic objects that are within a first preset distance range from the target geographic object and have a related influence on users within the target geographic object. The surrounding ancillary information data is the relationship data between the region where the target geographic object is located and its neighboring regions; the neighboring regions are regions within a second preset distance range from the region where the target geographic object is located.

[0033] The characteristic data of a geographic object can be either numerical or textual information. In this embodiment, the target geographic object is the geographic object whose value is to be evaluated. Taking a housing unit as an example, the geographic object can be of various types, such as commercial housing, residences, apartments, villas, townhouses, garages, warehouses, etc. The numerical information can specifically include the relevant attribute information of the housing unit and the community where it is located. Housing unit information can specifically include the building year, area, floor, living room, bedrooms, bathrooms, and orientation. Community information can specifically include the community's greening rate, plot ratio (also known as floor area ratio or volume fraction, which is the ratio of the total above-ground building area to the net land area of ​​a community), and water, electricity, and gas costs. The textual information mainly includes the descriptive information of the housing unit, which can include the housing unit introduction, community introduction, surrounding amenities, and transportation information.

[0034] It should be noted that information such as the building year, area, floor, and apartment type directly reflects the basic condition of the property and is an important factor in assessing its value. The completeness of surrounding amenities, the convenience of transportation, and whether the community is located in a bustling commercial area are all important factors affecting the property's valuation.

[0035] The defined region to which a geographic object belongs refers to the neighborhood where the target geographic object (e.g., a property) is located, or a sub-neighborhood within that neighborhood. For example, some properties belong to Zone 1, Zone 2, or the East Zone, South Zone, etc. Data on the defined region to which a geographic object belongs refers to the characteristic data of the neighborhood where the target geographic object is located, or the characteristic data of the sub-neighborhoods within that neighborhood. The characteristic data of the neighborhood where the geographic object is located can include quantifiable numerical information and descriptive textual information. Data on the defined region to which a geographic object belongs can refer to the data of the neighborhood where the property is located. A neighborhood refers to a large area of ​​residential buildings within a certain area of ​​a city that has a relatively independent living environment and is equipped with a complete set of living service facilities.

[0036] The geographic information of a geographic object can refer to a combination of information about the property and surrounding locations such as subway stations and hospitals, or it can be calculated information about the convenience of supporting entities to the property. In this embodiment, the surrounding associated geographic objects are geographic objects that are within a first preset distance range from the target geographic object and have a related impact on users within the target geographic object. Specifically, the surrounding associated geographic objects can be supporting entity information, which refers to surrounding geographic factors that have a significant impact on property value, such as schools, subways, hospitals, parks, supermarkets, shopping malls, or other important entities. In practice, an entity list can be pre-set to determine the degree of influence of the entities in the entity list on the property, or the property data and entity data can be input into a preset model to predict the degree of influence of the entity on the property, thereby determining whether the entity is an associated entity of the property. It should be noted that the surrounding associated geographic objects refer to the surrounding associated geographic objects of the property. The surrounding associated geographic objects of the property can be entities outside the community where the property is located or entities inside the community where the property is located, such as convenience stores and schools inside the community, or convenience stores and fast food restaurants in commercial buildings. The surrounding area refers not only to the outside of a residential community or commercial building, but also to the inside of the community or commercial building. Specific surrounding information data can include data on the relationships between the target geographic object's residential community and its neighboring communities.

[0037] In this embodiment of the application, a preliminary value assessment result of the target geographic object can be determined based on the attribute data, which serves as the first result. Specifically, when determining the preliminary value assessment result of the target geographic object based on the attribute data, the feature data of the geographic object, the data of the designated area to which the geographic object belongs, the geographic information of the geographic object, and the surrounding ancillary information data are input into a preset geographic object value assessment model to determine the preliminary value assessment result of the target geographic object, i.e., the first result. .

[0038] The geographic object valuation model is a regression model obtained through training. The geographic object valuation model is trained in the following manner: Sample data is determined from a set of geographic object data consisting of multiple geographic objects that have completed valuation. This sample data includes feature data samples of the geographic objects in the set of geographic object data, data samples of the designated area to which the geographic object belongs, geographic information samples of the geographic object, surrounding information samples, and actual result samples of the completed valuation of the geographic objects in the set of geographic object data. Each feature data sample, each designated area data sample, each geographic information sample, each surrounding information sample, and each actual result sample of a geographic object are used as a training sample pair. These training sample pairs are used to train the geographic object valuation model. An initial geographic object value assessment model is established by inputting a sample of the geographic object's characteristic data, a sample of the geographic object's defined region data, a sample of the geographic object's geographic information, and a sample of surrounding ancillary information data into the initial geographic object value assessment model to determine the initial value assessment result of the geographic object. If the loss data between the initial value assessment result and the actual result sample meets the acceptable loss condition, then the initial geographic object value assessment model is determined as the geographic object value assessment model. Otherwise, the model parameters in the initial geographic object value assessment model are adjusted until the loss data between the value assessment result obtained by the adjusted geographic object value assessment model and the actual result sample meets the acceptable loss condition. The adjusted geographic object value assessment model is then determined as the geographic object value assessment model.

[0039] The above describes one training method for a geographic object value assessment model. In practice, the LightGBM regression model can also be used to train the model based on sample housing data. LightGBM is a fast, distributed, high-performance gradient boosting framework based on decision tree algorithms. It is primarily used for ranking, classification, regression, and many other machine learning tasks. LightGBM's main advantages include faster training speed, higher efficiency, lower memory usage, and higher accuracy. To improve the model's ability to fit outlier data, the loss function can be trained using Huber Loss, i.e. in, For actual results, For the predicted results, = , This is a hyperparameter that determines the point at which the loss function transitions from squared error to absolute error.

[0040] The above describes the process of determining the preliminary value assessment result of the target geographic object based on the attribute data of the target geographic object, which serves as the first result. The following describes the process of determining at least one similar geographic object of the target geographic object, obtaining the historical transaction prices of all similar geographic objects, and calculating the statistical value of all historical transaction prices, which serves as the second result of similarity prediction.

[0041] Step S303: Identify at least one similar geographic object of the target geographic object, obtain the historical transaction prices of all similar geographic objects, and calculate the statistical value of all historical transaction prices as the second result of similarity prediction.

[0042] This step is used to identify similar geographic objects to the target geographic object. Similar geographic objects are those that have similar data characteristics to the target geographic object. Obtaining the historical transaction prices of all similar geographic objects and calculating the statistical values ​​of all historical transaction prices can significantly improve the accuracy and reliability of the target geographic object's value estimation results.

[0043] In the field of housing price forecasting, accurately estimating the price of a property is crucial. To improve forecast accuracy, various methods can be used to reference the characteristics of similar properties to obtain their estimated value. One method is to determine the similar geographical objects using predefined rules. These predefined rules include at least: considering all or part of other geographical objects within the designated area to which the target geographical object belongs as similar geographical objects; considering other geographical objects with the same or similar area as the target geographical object within the designated area and / or adjacent designated areas as similar geographical objects; considering other geographical objects with the same or similar floor level as the target geographical object within the designated area and / or adjacent designated areas as similar geographical objects; and considering other geographical objects with the same or adjacent building level as the target geographical object within the designated area to which the target geographical object belongs as similar geographical objects. For example, properties within the same neighborhood: properties within a neighborhood usually have similar geographical locations and environments, therefore, properties within the same neighborhood are given priority for comparison. Properties with similar areas: properties are categorized by area, and properties with areas approximately the same as the target property are typically selected, such as those within ±10%. Listings on similar floors: The floor level affects the price, so compare listings on floors close to your target property, such as within ±2 floors. Listings in similar buildings: Even within the same complex, prices can vary between different buildings. By selecting listings in the same or similar buildings, you can further reduce the error in your estimated results.

[0044] One method involves using clustering to determine similar geographic objects to the target geographic object. Specifically, this includes: identifying other geographic objects in the region where the target geographic object is located or in adjacent regions; acquiring attribute data of these other geographic objects; clustering these other geographic objects based on the attribute data and using attribute categories as clustering features; and determining similar geographic objects to the target geographic object from different clusters. The step of determining similar geographic objects to the target geographic object from different clusters includes: determining the similarity between other geographic objects in different clusters and the target geographic object, and identifying other geographic objects with similarity exceeding a set threshold as similar geographic objects to the target geographic object.

[0045] Clustering is an unsupervised learning technique in data mining that groups properties according to multiple dimensions and estimates prices by analyzing similar groups. Unsupervised clustering is performed using factors such as geographic location, historical transaction prices, and property layout. Multi-path recall is then used to find similar property information. The main process is as follows: 1) Feature selection: To perform clustering, appropriate features need to be selected, such as geographic location (latitude and longitude coordinates), historical transaction prices, and property layout (e.g., number of rooms, area). 2) Data standardization: To ensure the relative influence between different features, the data can be standardized to make each feature of the same magnitude. 3) Clustering algorithm selection: Clustering algorithms such as K-means and DBSCAN are used to cluster the property data, resulting in high similarity among properties in each cluster. 4) Multi-path recall: Clustering is performed from multiple perspectives (geographic location, price, layout). Each clustering method can find similar properties. A comprehensive evaluation of similar properties from different perspectives yields more accurate property information.

[0046] Another approach involves encoding the feature data of geographic objects using a pre-defined model to obtain encoded feature vectors. These feature vectors are then clustered based on pre-defined clustering conditions to obtain a set of similar geographic objects. Specifically, a deep learning model is used to learn high-level features of properties through unsupervised or semi-supervised methods. Clustering and similarity matching are then performed based on the learned representations. The main process is as follows: 1) Data preparation: Data includes textual features (property descriptions), numerical features (area, price, etc.), and image features (property images). 2) Model selection: Deep learning models, such as Autoencoders or Variational Autoencoders (VAEs), are used for unsupervised learning to learn high-level features from the data. 3) Representation learning: The model is used to learn representations of the property features, mapping the property data to a high-dimensional feature space. At this point, each property has a higher-dimensional feature representation. 4) Similar property clustering: Clustering is performed in the high-dimensional feature space, grouping similar properties into one category. Price estimation for the property to be evaluated is then performed using this information. 5) Model optimization: Semi-supervised learning methods can be used to train the model using a portion of data with known labels, thereby improving the model's representation ability.

[0047] In specific implementation, the similarity between the other geographic objects and the target geographic object is calculated as follows: Based on the attribute data of the target geographic object, a vector representation of the attribute data is obtained as a first vector A; a vector representation of the attribute data of any other geographic object is obtained as a second vector B; the cosine similarity between the first vector and the second vector is calculated as the similarity between the other geographic object and the target geographic object. ,Right now When determining similarity based on the location of a geographic object, the similarity between the other geographic objects and the target geographic object is calculated as follows: based on the geographic location information of the object to be evaluated ( ) and the geographic location information of any other geographic object ( The geographical distance between the target geographic object and other geographic objects is determined; this geographical distance is used as the similarity between the other geographic objects and the target geographic object. The geographic location information of the object to be evaluated (…) The geographic distance can be the latitude and longitude data of the target geographic object, and the geographic distance can be Euclidean distance or Manhattan distance. Geographic distance is used to measure the location similarity between other geographic objects and the target geographic object. ,Right now When determining similarity based on the house type of a geographic object, the similarity between the other geographic objects and the target geographic object is calculated as follows: First, the house type of the target geographic object is converted into a numerical feature representation, which serves as the third vector. Convert the house type data of any other geographic object into a numerical feature representation, as the fourth vector. For example, if the target geographic object's house type is two bedrooms, one living room, one kitchen, one bathroom, and south-facing, in practice, the two bedrooms, one living room, one kitchen, one bathroom, and south-facing can be converted into corresponding numerical features for representation, resulting in various vector representations. These vector representations are then concatenated to obtain the high-dimensional vector of the overall house type. Of course, the two bedrooms, one living room, one kitchen, and one bathroom can also be directly represented by vectors (…). ) represents a one-bedroom, one-living room, zero-kitchen, one-bathroom unit represented by a vector ( The expression is represented as follows: Then, the cosine similarity between the third and fourth vectors is calculated as the similarity between the other geographic object and the target geographic object. ,Right now When determining similarity based on the area of ​​geographic objects, the similarity between other geographic objects and the target geographic object is calculated as follows: the areas of other geographic objects are normalized; the normalized result is used as the similarity between the other geographic objects and the target geographic object. For example, after normalizing the areas of the target geographic object and other geographic objects, the absolute values ​​of the normalized results are subtracted to obtain the area similarity. The closer the result is to zero, the closer the property areas are, and the higher the similarity of the properties. The method for normalizing the area is as follows: It should be noted that the above methods determine similarity using attribute data, location information, house type, and area between geographic objects. Alternatively, multiple feature similarities can be combined and a weighted average can be used to obtain the final property similarity score. In practice, the appropriate similarity calculation method can be selected based on actual needs. The following is the comprehensive similarity... The calculation method, in which, For positional similarity, The weight given by positional similarity. For room type similarity, The weight given to the similarity of room types. For area similarity, The weight given to area similarity. It can be obtained through model training: The above describes the process of obtaining the similarity between other geographic objects and the target geographic object, as well as the set of similar geographic objects. In specific implementation, appropriate methods or combinations of several methods can be selected according to actual needs to obtain the set of similar geographic objects, thereby improving the accuracy of housing price prediction. By combining the results of multiple methods for comprehensive evaluation, the accuracy and reliability of price prediction can be significantly improved. After identifying at least one similar geographic object of the target geographic object, the historical transaction prices of all similar geographic objects are obtained, and the statistical value of all historical transaction prices is calculated as the second result of similarity prediction. The calculation of the statistical value of all historical transaction prices includes: determining the similarity that characterizes the degree of association between each similar geographic object and the target geographic object, and using the similarity as a weight to calculate the statistical value of all historical transaction prices by weighted averaging.

[0048] In obtaining a set of similar geographic objects Subsequently, similarity data between each of the similar geographic objects in the set of similar geographic objects and the target geographic object is further obtained. , It can refer to any one of the multiple individual similarities obtained from the above calculations, or it can refer to the comprehensive similarity calculated above, as well as the historical transaction price of each of the similar geographical objects in the set of similar geographical objects. Based on the similarity data between each of the similar geographic objects and the target geographic object. and the historical transaction price of each of the aforementioned similar geographic objects. The statistical value of all historical transaction prices is calculated by weighting all historical transaction prices using the similarity as a weight. ,Right now in, For similarity data, in actual calculations, if the similarity is calculated based on positional similarity, then the similarity... According to the above The value is determined by the similarity score; if the similarity is calculated based on room type similarity, then the similarity score... According to the above The value is determined by the similarity score; if the similarity is calculated based on a comprehensive similarity score, then the similarity score is... According to the above The value can be determined by whether the similarity is calculated based on area or price similarity. Correspondingly, values ​​are assigned based on the area similarity or price similarity mentioned above. The historical transaction prices of similar properties. This is a collection of similar properties. Finally, the statistical value of all historical transaction prices is calculated by either taking the arithmetic mean of all historical transaction prices or using the median of all historical transaction prices.

[0049] Step S304: Obtain the historical transaction prices of other geographical objects in the area where the target geographical object is located at different times. Based on the fitting curve determined by the relationship between the historical transaction prices and time, determine the predicted value of the current evaluation result of the area as the third result. Calculate the ratio of the third result to the historical transaction price corresponding to the previous time unit on the fitting curve as the trend change coefficient.

[0050] When obtaining the trend change coefficient, the process first involves acquiring the historical transaction prices of other geographic objects in the target geographic object's region at different times. The target geographic object's region can be a pre-defined neighborhood, a combined area of ​​surrounding neighborhoods, an area within a pre-defined distance from the target geographic object, the administrative division information of the target geographic object, or the city to which the target geographic object belongs. The specific implementation can be tailored to actual needs. After obtaining the historical transaction prices of other geographic objects in the target geographic object's region at different times, the predicted value of the current assessment result for that region can be determined based on the fitted curve determined by the relationship between the historical transaction prices and time. This predicted value serves as the third result. .

[0051] In practice, housing prices in different regions can be aggregated by month to construct time series data in terms of region, month, and average price, thereby obtaining housing price trend data for that region. Please see [link / reference]. Figure 4 , Figure 4This is a schematic diagram of the fitting curve in a geographic object valuation method provided in the first embodiment of this application. According to the fitting curve, the housing price trend of the region can be obtained. For example, based on historical transaction data (such as average price) before January 2024, the estimated housing price data for March 2024 and April 2024 can be predicted. Figure 4 (In the middle circle). It should be noted that the embodiments of this application use the Prophet model. The basic principle of Prophet is to decompose the time series into a trend term, a seasonal term, and a holiday term, and then use an additive model for prediction based on this. Specifically, it is assumed that the time series consists of the following three parts: trend term... : Describes the long-term trend of the time series, using a piecewise linear model with an autoregressive term and a seasonal term for fitting; the seasonal term : Describes the periodic changes of a time series, using Fourier series for fitting; holiday term : Describes abnormal events (such as holidays) in a time series, represented by a custom binary variable; Error term: Represents variations that the model failed to capture; it is usually assumed to be normally distributed. Adding the trend term, seasonal term, and holiday term yields the predicted value of the time series. Right now In practice, in order to obtain the trend change coefficient In determining the predicted value of the current assessment results for this region. Then, obtain the historical transaction prices corresponding to the previous time unit (e.g., the previous month) on the fitted curve. Trend change coefficient The third result Historical transaction prices corresponding to the previous time unit on the fitted curve The ratio, i.e. It should be noted that the previous time unit on the fitted curve is determined based on the time unit adopted in real time according to the relationship between historical transaction prices and time. For example, if the time unit adopted in determining the fitted curve is year, then the previous time unit is the previous year; if the time unit adopted in determining the fitted curve is month, then the previous time unit is the previous month; if the time unit adopted in determining the fitted curve is quarter, then the previous time unit is the previous quarter.

[0052] Step S305: Determine the missingness of the similar geographic objects based on the number of similar geographic objects of the target geographic object and the number of transactions in the region where the target geographic object is located.

[0053] This step is used to determine the degree of missing similar geographic objects.

[0054] When determining the missing value of similar geographic objects, the following steps are taken: First, obtain the number of similar geographic objects to the target geographic object and the number of transactions in the region where the target geographic object is located. Then, obtain the difference between the number of similar geographic objects and the number of transactions in the region where the target geographic object is located. Finally, obtain the ratio of this difference to the number of transactions in the region where the target geographic object is located, which is used as the missing value of the similar geographic objects. For example, if the number of similar geographic objects to the target geographic object is 15 and the number of transactions in the region where the target geographic object is located is 20, then the missing value is... for: It should be noted that if the number of similar geographic objects to the target geographic object is 30, and the number of transactions in the area where the target geographic object is located is 20, then the missing value will be calculated because the number of similar geographic objects to the target geographic object exceeds the number of transactions in the area where the target geographic object is located. If the number of transactions in the region where the target geographic object is located is 0, then the missing value is 1.

[0055] Step S306: Using the missing degree as an adjustment coefficient and the product of the second result and the trend change coefficient as an adjustment amount, adjust the first result to obtain the value assessment result of the target geographic object.

[0056] This step is used to obtain the value assessment results of the target geographic object. The value assessment results combine the preliminary value assessment results of the target geographic object, the second result of similarity prediction, the trend change coefficient, and the missingness of similar geographic objects, so as to make the final value assessment results more accurate.

[0057] In this embodiment of the application, the step of using the missing value as an adjustment coefficient and the product of the second result and the trend change coefficient as an adjustment amount to adjust the first result to obtain the value assessment result of the target geographic object includes: obtaining the negative missing value corresponding to the first result based on the full value corresponding to the missing value and the missing value; using the negative missing value as the weight of the first result and the missing value as the weight of the second result to obtain the weighted sum of the first result and the second result; using the trend change coefficient as the adjustment factor data of the weighted sum data, and using the trend change coefficient to adjust the weighted sum data to obtain the value assessment result of the target geographic object.

[0058] It should be noted that, based on the first result Second result and trend change coefficient To obtain the value assessment results of the target geographic object. Specifically, this involves determining the degree of missing similar geographic objects. Then, assuming If the range is 0-1, then the first result The corresponding negative missing value is 1- The negative missing value 1- As the first result The weight of the missing value As a second result By determining the weights, the first result can be obtained. Compared with the second result The weighted sum of the data, and then the trend change coefficient. As an adjustment factor for the weighted sum data, the trend change coefficient is used. Adjust the weighted sum data to obtain the value assessment results of the target geographic object. ,Right now It should be noted that when the missing value When the value approaches 0, the housing supply forecast tends to favor the first result. When the degree of absence When the value tends towards 1, the housing supply forecast tends to favor the second result. In other words, when the number of transactions in the area where the target geographic object is located is relatively abundant, the results obtained from the geographic object valuation model can be used; when the number of transactions in the area where the target geographic object is located is relatively sparse, the transaction results of similar geographic objects can be referenced.

[0059] The above describes the process of obtaining the value assessment results of a target geographical object through the single-property prediction module, the similar-property price prediction module, and the housing price trend prediction module. In the similar-property price prediction module, the clustering information section mainly provides information on similar properties and similar communities for new properties and sparsely traded neighborhoods, increasing the overall robustness of the model. In specific implementation, a first additional result can be obtained based on the category data and attribute data of the property to be assessed, and a second additional result can be obtained based on the city data and attribute data of the property to be assessed. Finally, the value assessment result is merged with the first and second additional results to obtain the total assessment result of the property to be assessed, making the final property value prediction result more accurate. Specifically, the following steps are taken: obtaining the category data and city data of the target geographical object; determining the assessment result of the target geographical object as the first additional result based on the category data and attribute data; determining the assessment result of the target geographical object as the second additional result based on the city data and attribute data; and merging the value assessment result with the first and second additional results to obtain the total assessment result of the target geographical object.

[0060] The step of determining the evaluation result of the target geographic object as a first additional result based on the category data and the attribute data includes: determining a sub-model based on the category data; and determining the evaluation result of the target geographic object as a first additional result based on the sub-model obtained from the category data and the attribute data. The step of determining the evaluation result of the target geographic object as a second additional result based on the city data and the attribute data includes: determining a sub-model based on the city data; and determining the evaluation result of the target geographic object as a second additional result based on the sub-model obtained from the city data and the attribute data.

[0061] It should be noted that the category data of the target geographic object can represent the type of the target geographic object, such as residential, apartment, villa, or townhouse. For example, if the type of the target geographic object is apartment, considering the special attributes of the apartment type, an apartment sub-model can be selected. This sub-model is constructed based on historical transaction records of the apartment, using attribute data such as the apartment's geographical location, area, number of rooms, year of construction, supporting facilities, rental price, and utility fees. The training method for the apartment valuation sub-model is similar to that of the geographic object valuation model described above; please refer to the training method for the geographic object valuation model mentioned above, which will not be detailed here.

[0062] The above describes one training method for the apartment value assessment sub-model. In practice, the GBDT model can also be used for training based on sample apartment data. The GBDT model (Gradient Boosting Decision Tree) is a reinforcement learning algorithm commonly used for regression and classification problems. It improves model accuracy by combining multiple weak learners (usually decision trees) into a strong learner. The GBDT model primarily builds the decision tree iteratively, attempting to correct the errors of the previous model with each iteration. The basic process is as follows: 1. Initial Model: First, a simple model (usually the mean) is used to initialize the prediction. For regression problems, the initial output is the mean of the target values ​​of the training data. 2. Residual Calculation: The current model is used to predict the training data, calculating the error (residual) for each sample. The residual is the difference between the true value and the predicted value. 3. New Tree Construction: A new decision tree is built based on the residuals. The new tree aims to minimize the loss function of the current model; specifically, it fits the residuals calculated in the previous step. 4. Update the model: Add the predictions of the new trees to the predictions of the current model. A learning rate (shrinkage) is typically used to control the impact of each tree on the final prediction. 5. Iterate: Repeat steps 2 to 4 until the preset number of trees is reached, or the residuals no longer change significantly.

[0063] In practical implementation, in addition to the apartment valuation sub-model, villa valuation sub-models and garage valuation sub-models can also be constructed. For the villa valuation sub-model, considering the special attributes of villas, a separate model can be built for villa properties. This model takes into account data such as villa location, area, total floors, number of rooms, architectural style, year of construction, plot ratio, and property management fees, and is constructed based on historical transaction records. The villa valuation sub-model can be trained using the same methods as the apartment valuation sub-model, or it can be trained using the GBDT model. For the garage valuation sub-model, it can separately estimate the value of parking spaces for properties containing parking spaces. This estimation comprehensively considers information such as the price of the community where the parking space is located, the size of the parking space, the type of parking space, and the rent.

[0064] It should be noted that the city data of the target geographic object can be obtained through city sub-models and regional sub-models. For city sub-models, considering the different development levels of cities—for example, first-tier cities have more active secondhand housing transactions and higher unit prices, while third- and fourth-tier cities have sparse secondhand housing transactions and lower unit prices—we group and model cities based on their development levels to construct city sub-models. The training method for city sub-models is similar to that of the sub-models mentioned above and will not be detailed here. For regional sub-models, considering the different development levels of different regions—for example, eastern cities have higher housing liquidity than western cities—we construct regional sub-models based on national address divisions, separately for each region. The training method for regional sub-models is similar to that of the sub-models mentioned above and will not be detailed here.

[0065] It's important to note that the geographic object valuation model, as the master valuation model, is the core model. This master model integrates multiple key factors, such as historical transactions, similar properties, and market trends, aiming to provide a preliminary and comprehensive valuation reference. Sub-models carefully consider the unique attributes of specific property types. In the complex real estate market, a single model cannot adapt to all situations. Therefore, multiple sub-models are customized based on different residential attributes and geographical locations. For example, the apartment valuation sub-model may specifically consider the unique attributes of apartments, while the city sub-model focuses more on the city tier to which the property belongs. Although each sub-model may be structurally similar, they will differ significantly in their handling of specific factors to ensure more accurate predictions for specific property types.

[0066] In this embodiment of the application, the step of fusing the value assessment result with the first additional result and the second additional result to obtain the overall assessment result of the target geographic object includes: obtaining a first weight of the value assessment result; obtaining a second weight of the first additional result and the second additional result; and obtaining the overall assessment result of the target geographic object based on the value assessment result, the first weight, the first additional result, the second additional result, and the second weight.

[0067] In practice, the value assessment result can be the main result, and the first and second supplementary results can be used as auxiliary results, based on the main result. Harmony and auxiliary results To obtain the overall evaluation result of the target geographic object. Specifically, based on the type of housing and its city, a logistic regression (LR) model is used based on training data to learn different weights for different sub-models, and finally, the final model prediction result is obtained through a weighted summation. Specifically, this is based on the main result... First weight Auxiliary results Second weight To obtain the main result and the auxiliary results The weighted sum of data is used as the overall evaluation result for the target geographic object. ,Right now It should be noted that, based on the main result Harmony and auxiliary results To obtain the overall evaluation result of the target geographic object. hour, There can be multiple auxiliary results. The sum of .

[0068] The above describes the overall process of obtaining the overall evaluation result of the target geographic object. In specific implementation, after obtaining the value assessment result or the overall evaluation result of the target geographic object, the process further includes: using a dynamic correction strategy to correct the value assessment result or the overall evaluation result of the target geographic object, including: identifying favorable or unfavorable factors affecting the value assessment of the target geographic object; determining the degree of influence of the favorable or unfavorable factors on the current value assessment as a correction coefficient; and using the correction coefficient to correct the value assessment result or the overall evaluation result of the target geographic object.

[0069] It should be noted that after obtaining the overall assessment result of the target geographic object, the valuation result or the overall valuation result of the target geographic object will be revised. Specifically, first, the favorable or unfavorable factors affecting the valuation of the target geographic object are identified. Then, the degree of influence of these favorable or unfavorable factors on the current valuation is determined as a correction coefficient. Finally, the valuation result or the overall valuation result of the target geographic object is revised using the correction coefficient. For example, the surrounding facilities of the geographic object may be temporarily added or removed; it may also be determined based on the characteristic data of the demand side of the geographic object, such as the number of properties already owned by the homebuyer. All of the above situations will affect the valuation result of the property. Through the correction coefficient, the valuation result of the property can better adapt to the dynamically changing market environment.

[0070] The above is a complete description of the geographic object valuation method described in the embodiments of this application. On the one hand, this method solves the limitation of existing technologies in identifying similar properties. Most existing property valuation models rely solely on historical transaction data of the local community to assess property value, neglecting comprehensive analysis of similar properties and similar communities, resulting in significant defects in the generalization performance of the models. For example, even if different communities are geographically close, their prices may differ significantly due to differences in building quality, supporting facilities, and other factors. Therefore, relying solely on data from a single community for valuation may overlook these important external factors. On the other hand, this method solves the problem of existing technologies being unable to adapt to rapidly changing market dynamics. Real estate market prices are affected by a variety of complex factors, such as supply and demand, which are usually difficult to predict and change rapidly. However, many existing valuation models are built based on static data and cannot reflect the latest market dynamics in a timely manner, especially when the market is highly volatile, which may lead to a large deviation between the valuation results and the actual prices. On the other hand, this method addresses the problem of insufficient consideration of geographical heterogeneity. Housing prices can vary significantly between different cities and even different areas within the same city. General-purpose valuation models that do not adequately account for geographical heterogeneity may lead to substantial errors in valuation results. For example, housing prices in city centers can differ greatly from those in suburbs, and some existing models cannot accurately capture such regional differences, thus affecting valuation accuracy. In contrast, the geographical object data processing method described in this application comprehensively considers individual properties, similar properties, and market trends. The model is not only based on housing information within a single community but also incorporates historical transaction data from similar properties and similar communities, thereby enhancing the model's robustness and accuracy. Furthermore, by analyzing the characteristics of housing price fluctuations in different regions, the model can learn and capture housing price trends, thereby improving its ability to predict future market trends.

[0071] The geographic object valuation method described in this application includes: identifying a target geographic object; obtaining attribute data of the target geographic object, and determining a preliminary valuation result of the target geographic object based on the attribute data, as a first result; identifying at least one similar geographic object of the target geographic object, obtaining the historical transaction prices of all similar geographic objects, and calculating the statistical value of all historical transaction prices, as a second result of similarity prediction; obtaining the historical transaction prices of other geographic objects in the area where the target geographic object is located at different times, and determining the predicted value of the current valuation result of the area based on the fitting curve determined by the relationship between the historical transaction prices and time, as a third result; calculating the ratio of the third result to the historical transaction price corresponding to the previous time unit on the fitting curve, as a trend change coefficient; determining the missing degree of the similar geographic objects based on the number of similar geographic objects of the target geographic object and the number of transaction objects in the area where the target geographic object is located; adjusting the first result using the missing degree as an adjustment coefficient and the product of the second result and the trend change coefficient as an adjustment amount, to obtain the valuation result of the target geographic object.

[0072] The geographic object valuation method described in this application comprehensively considers the preliminary valuation result determined based on the attribute data of the target geographic object, the statistical value of the historical transaction price of all similar geographic objects similar to the target geographic object, the predicted value of the current valuation result of the region determined by the fitting curve of the historical transaction price and time relationship of other geographic objects in the region where the target geographic object is located at different times, the trend change coefficient determined by the ratio of the predicted value to the historical transaction price corresponding to the previous time unit on the fitting curve, and the impact of the missing number of similar geographic objects on the valuation result of the target geographic object when obtaining the valuation result of the target geographic object. This improves the accuracy of the valuation of the target geographic object. Moreover, the data required for the valuation calculation is ordered and deterministic, which reduces the consumption of computing resources, improves the efficiency of the valuation calculation, and enhances the stability of the calculation results.

[0073] Second Embodiment In the first embodiment described above, a method for assessing the value of a geographic object is provided. Correspondingly, the second embodiment of this application provides a device for assessing the value of a geographic object. Since the device embodiment is basically similar to the first method embodiment, it is described simply; relevant details can be found in the description of the method embodiment. The device embodiment described below is merely illustrative.

[0074] Please refer to Figure 5 This is a schematic diagram of a geographic object value assessment device provided in the second embodiment of this application.

[0075] The geographic object valuation device 500 includes: Unit 501 is used to determine the target geographic object; The first result evaluation unit 502 is used to obtain attribute data of the target geographic object and determine the preliminary value evaluation result of the target geographic object based on the attribute data, as the first result; The second result acquisition unit 503 is used to identify at least one similar geographic object of the target geographic object, obtain the historical transaction prices of all similar geographic objects, and calculate the statistical value of all historical transaction prices as the second result of similarity prediction. The trend change coefficient determination unit 504 is used to obtain the historical transaction prices of other geographical objects in the area where the target geographical object is located at different times, and determine the predicted value of the current evaluation result of the area based on the fitting curve determined by the relationship between the historical transaction price and time, as the third result. The ratio of the third result to the historical transaction price corresponding to the previous time unit on the fitting curve is calculated as the trend change coefficient. The missing value determination unit 505 is used to determine the missing value of the similar geographical objects based on the number of similar geographical objects of the target geographical object and the number of transaction objects in the area where the target geographical object is located; The value assessment result determination unit 506 is used to adjust the first result by using the missing degree as an adjustment coefficient and the product of the second result and the trend change coefficient as an adjustment amount, so as to obtain the value assessment result of the target geographic object.

[0076] Third Embodiment Corresponding to the method embodiments described above, the third embodiment of this application also provides an electronic device. For example... Figure 6 As shown, Figure 6 This is a schematic diagram of an electronic device provided in the third embodiment of this application. The electronic device includes: at least one processor 601, at least one communication interface 602, at least one memory 603, and at least one communication bus 604. Optionally, the communication interface 602 can be an interface for a communication module, such as the interface for a GSM module. The processor 601 may be a CPU, an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The memory 603 may include high-speed RAM and may also include non-volatile memory, such as at least one disk storage device. The memory 603 stores a program, and the processor 601 calls the program stored in the memory 603 to execute the methods provided in the above embodiments of this application.

[0077] Fourth embodiment Corresponding to the methods described above in this application, the fourth embodiment of this application also provides a computer storage medium. The computer storage medium stores a computer program, which is executed by a processor to perform the methods provided in the above embodiments of this application.

[0078] The method for assessing the value of geographic objects provided in the above embodiments can be applied to various application scenarios, such as information prediction, information recommendation, and various information evaluation scenarios. The following explanation focuses on information recommendation.

[0079] Figure 7 This is a flowchart illustrating an information recommendation method provided in an embodiment of this application. Figure 7 As shown, the method includes the following steps: S701: Obtain user profile data and historical behavior data, wherein the profile data includes target geographic objects associated with the user, the attribute data of the target geographic objects reflects the preliminary value information of the target geographic objects, and the historical behavior data reflects the user's dynamic interest intentions; S702: Based on the attribute data of the target geographic object, determine the similar geographic objects of the target geographic object and their quantity, wherein the attribute data of the target geographic object includes at least the target region to which the target geographic object belongs; S703: Generate the value trend change coefficient of the target area based on the fitted curve of the historical transaction price and time relationship of other geographical objects in the target area; S704: Determine the missingness of the similar geographic objects based on the number of similar geographic objects and the number of transactions in the target area; S705: Using the missing value as an adjustment coefficient, and the product of the statistical value of the historical transaction price of the similar geographical objects and the trend change coefficient as an adjustment amount, the preliminary value information is adjusted to obtain the target value information, which reflects the user's steady-state demand potential. S706: Based on the historical behavior data and the target value information of the target geographic object, perform joint information recommendation for the user from two dimensions: dynamic interest intent and steady-state demand potential.

[0080] In this embodiment, user profile data and historical behavior data are acquired. The profile data includes target geographic objects associated with the user. Taking a property as an example, the target geographic object can be a residence, apartment, villa, or townhouse. Historical behavior data includes, but is not limited to, user clicks, browsing, searches, and purchase records.

[0081] Furthermore, in this embodiment, based on the attribute data of the target geographic object, similar geographic objects and their quantities are determined. The attribute data of the target geographic object at least includes the target region to which the target geographic object belongs. Optionally, the attribute data includes regional attributes and area attributes, with the regional attribute indicating that the target geographic object belongs to the target region.

[0082] In this embodiment, taking a property as the target geographic object, the attribute data of the target geographic object can specifically be the information of the neighborhood where the property is located. Then, clustering information is obtained through a clustering model to identify similar geographic objects with attribute data similar to the target geographic object and their quantities.

[0083] Furthermore, in this embodiment, the missingness of similar geographic objects is determined based on the number of similar geographic objects and the number of transactions in the target area. When determining the missingness data for similar geographic objects, the difference between the number of similar geographic objects of the target geographic object and the number of transactions in the area where the target geographic object is located is calculated. The ratio of this difference to the number of transactions in the area where the target geographic object is located is used as the missingness of similar geographic objects.

[0084] Furthermore, using the missing value as an adjustment coefficient, and the product of the statistical value of historical transaction prices of similar geographic objects and the trend change coefficient as an adjustment amount, the preliminary value information is adjusted to obtain target value information, which reflects the user's steady-state demand potential. Since the target geographic object has a strong correlation with the user and can reflect the user's long-term, steady-state characteristics, such as the user's residential level, economic level, and social circle, this information can be used to characterize the user's steady-state demand potential.

[0085] In this embodiment, based on historical behavioral data and target value information of the target geographic object, joint information recommendation is performed on the user from two dimensions: dynamic interest intention and steady-state demand potential.

[0086] In this embodiment, the user's recent interest change trends are identified in the dimension of dynamic interest intent, and the user's long-term geographical object value preferences are identified in the steady-state dimension. By fusing information from the two dimensions, a joint information recommendation is performed to generate personalized recommendation results.

[0087] In this embodiment, attribute data and historical behavioral data of the target geographic object are jointly modeled to form a collaborative recommendation link from behavior-driven to value-driven. Similar geographic objects and their quantities are determined based on the attribute data of the target geographic object. A value trend change coefficient is generated based on the fitted curve of the historical transaction price versus time relationship within the target area. The missing value of similar objects is then calculated by combining the number of similar objects with the number of transactions in the area. Using the missing value as an adjustment coefficient and the statistical value of the historical transaction price of similar objects and the value trend change coefficient as adjustment amounts, the preliminary value information is adjusted to obtain target value information reflecting the user's steady-state demand potential. The dynamic interest intent represented by historical behavioral data and the target value information are jointly used for information recommendation in two dimensions. This breaks through the traditional data paradigm used for recommendation and greatly improves the recommendation effect of the recommendation system.

[0088] In addition to the information recommendation method described above, this application also provides another information recommendation method that relies on user-associated address objects. Figure 8 This is a flowchart illustrating another information recommendation method provided in an embodiment of this application. For example... Figure 8 As shown, the method includes the following steps: S801: Obtain user profile data and historical behavior data, wherein the profile data includes the identification information of the target geographic objects associated with the user; S802: Based on the identification information, obtain the attribute data of the target geographic object from the target library. The target library is obtained by using the geographic semantic features, distance features and geographic context features of the geographic object description information to perform data augmentation on the geographic object description information in the ontology library using the geographic object description information in the third-party library. The geographic object description information includes the attribute data of the corresponding geographic object. S803: Based on the attribute data of the target geographic object, predict the user's steady-state demand potential information, and based on historical behavior data, predict the user's dynamic interest intention information; S804: Based on the user's steady-state demand potential information and dynamic interest intent information, combined with candidate information objects in the target application scenario, perform joint information recommendation for the user.

[0089] In this embodiment, the profile data is used to describe the user's characteristics, preferences, and attribute information. The obtained profile data includes the identification information of the target geographic objects associated with the user. Taking housing as an example, the target geographic objects may include, but are not limited to: residential geographic objects (e.g., communities, residential areas, buildings); public facility geographic objects (e.g., schools, subway stations, bus stops, hospitals); commercial and service geographic objects (e.g., business districts, shopping centers, commercial buildings, office buildings); and geographic objects divided by coverage area (e.g., street coverage areas, regional centers, districts / counties). In this embodiment, the target geographic objects associated with the user refer to geographic objects that the user is interested in or related to. The target geographic objects have corresponding identification information.

[0090] In this embodiment, historical behavior data refers to a record of user actions on a target platform in chronological order. Target platforms include, but are not limited to, e-commerce platforms, social media platforms, payment platforms, and logistics platforms. Historical behavior data includes, but is not limited to, user clicks, browsing, searches, and purchase records.

[0091] Furthermore, based on the identifier information of the target geographic object, the attribute data of the target geographic object is obtained from the target database. The target database is obtained by augmenting the geographic object description information in the ontology database with geographic semantic features, distance features, and geographic context features based on the geographic object description information, using geographic object description information from a third-party database. The geographic object description information includes the attribute data of the corresponding geographic object. Specifically, the geographic semantic features, distance features, and geographic context features are extracted for any geographic object description information from the semantic aspects of name and address, spatial location, and surrounding geographic environment, respectively.

[0092] In this embodiment, the attribute data of a geographic object includes the geographic object's feature data, the data of the defined region to which the geographic object belongs, the geographic information of the geographic object, and the surrounding ancillary information data. Specifically, the geographic information of the geographic object is the relationship information between the target geographic object and its surrounding associated geographic objects. The surrounding associated geographic objects are geographic objects that are within a first preset distance range from the target geographic object and have a related influence on users within the target geographic object. The surrounding ancillary information data is the relationship data between the region where the target geographic object is located and its neighboring regions. The neighboring regions are regions within a second preset distance range from the region where the target geographic object is located.

[0093] Given the deep reliance of traditional behavioral data systems on dynamic behavioral sequences and static basic profiles, demand perception capabilities are limited by existing data paradigms, making it difficult to break through the cognitive boundaries formed by single behavioral dimensions and fixed profile features. In this embodiment, on the one hand, by introducing attribute data of target geographic objects, a steady-state demand inference path distinct from traditional behavioral data systems is constructed. The attribute data of target geographic objects reflects, to some extent, the user's relatively stable demand or potential capabilities. Generally, the value of a user's neighborhood or residence can reflect the user's consumption level or purchasing power; the higher the value of the neighborhood or residence, the stronger the user's consumption level or purchasing power to some extent. On the other hand, a dynamic demand inference path is constructed using historical user behavior data. Historical user behavior data reflects, to some extent, the user's dynamic or short-term interests, thereby obtaining user demand information at both the "long-term steady state" and "short-term dynamic" time scales. A detailed introduction follows.

[0094] Specifically, in this embodiment, based on the attribute data of the target geographic object, the steady-state demand potential information of the user is predicted, thereby enabling the prediction of the user's long-term demand tendency. Simultaneously, the user's historical behavioral data has time-series characteristics, which can present the user's demand change trend within a recent time window. Based on this, historical behavioral data can be used to predict the user's dynamic interest intention information, thereby identifying the user's demand change trend within a short-term time window.

[0095] Furthermore, based on the user's steady-state demand potential information and dynamic interest intent information, combined with candidate information objects in the target application scenario, joint information recommendations are made to the user. In this embodiment, the target application scenario is not limited. For example, it could be e-commerce marketing, auction marketing, and various information recommendation scenarios.

[0096] In this embodiment, when performing joint information recommendation, steady-state demand potential information and dynamic interest intent information are used as a joint expression of user needs, and these are matched and analyzed with candidate information objects in the target application scenario. Therefore, by fusing steady-state demand potential information and dynamic interest intent information, more accurate information recommendations can be made for users across multiple time scales and multiple demand dimensions.

[0097] In this embodiment, by introducing target geographic objects associated with the user and their attribute data in the target database on the basis of user profile data and historical behavior data, geographic semantic features, distance features, and geographic context features are incorporated into the user modeling process, breaking through the data paradigm that relies solely on predetermined behavior data and static profiles. On this basis, based on the attribute data and historical behavior data of the target geographic objects, the steady-state demand potential information and dynamic interest intention information of the user are predicted, thus characterizing the user's demand features from two dimensions: long-term steady-state demand and short-term dynamic interest. Furthermore, the steady-state demand potential information and dynamic interest intention information are combined with candidate information objects in the target application scenario for joint information recommendation, so that the recommendation results can not only match the user's current interests but also maintain consistency with their long-term needs, greatly improving the recommendation effect.

[0098] In one optional embodiment, predicting a user's steady-state demand potential information based on the attribute data of a target geographic object includes: determining similar geographic objects based on the attribute data of the target geographic object, wherein the attribute data of the target geographic object includes at least the target region to which the target geographic object belongs; and calling a pre-trained first information prediction model to perform information prediction based on the attribute information of the target region, the attribute data of the target geographic object, and the attribute data of the similar geographic objects to obtain the user's steady-state demand potential information. When predicting the user's steady-state demand potential information, the attribute data of the target geographic object and similar geographic objects, as well as the attribute information of the target region, are introduced as input to the pre-trained first information prediction model. This enables the first information prediction model to predict the user's steady-state demand potential information within a richer spatial semantic and environmental context, significantly improving the accuracy and stability of the prediction results. This avoids the solidification of cognitive boundaries caused by over-reliance on historical behavior sequences and static profiles, allowing recommendations to move beyond existing data paradigms. It provides more comprehensive, stable, and reliable steady-state demand potential information for subsequent joint information recommendations, thereby significantly improving the overall recommendation performance and the matching degree with user needs.

[0099] In this embodiment, the attribute data of the target geographic object includes at least the target area to which it belongs. Therefore, the target area can be used as a basic reference to limit the scope of candidate similar objects.

[0100] In this embodiment, the method for determining similar geographic objects is not limited. For example, based on the attribute data of the target geographic object and other geographic objects, clustering processing is performed using an attribute-based clustering algorithm and / or a high-dimensional feature vector-based clustering algorithm to obtain similar geographic objects of the target geographic object. The relevant content regarding the clustering algorithm can be found in the foregoing embodiments and will not be repeated here.

[0101] Furthermore, based on the attribute information of the target area, the attribute data of the target geographic object, and the attribute data of similar geographic objects, a pre-trained first information prediction model is invoked to predict the user's steady-state demand potential information. The first information prediction model can comprehensively utilize the attribute relationships between the target geographic object and similar geographic objects to predict the user's demand tendency over a long-term time scale, thereby obtaining the user's steady-state demand potential information.

[0102] In one optional embodiment, the process of calling a pre-trained first information prediction model to predict information based on the attribute information of the target area, the attribute data of the target geographic object, and the attribute data of similar geographic objects to obtain the user's steady-state demand potential information includes: classifying the attribute information of the target area, the attribute data of the target geographic object, and the attribute data of similar geographic objects to obtain first data of multiple information modalities; inputting the first data of multiple information modalities into the first information prediction model, which includes a first global decoder, a first encoder and a first sub-decoder corresponding to each of the multiple information modalities, and performing the following operations within the first information prediction model: encoding process: inputting the first data of multiple information modalities into the first encoders corresponding to the multiple information modalities for encoding to generate first data vectors corresponding to each of the multiple information modalities; decoding process: inputting the first data vectors corresponding to each of the multiple information modalities into the corresponding first sub-decoders for local decoding, and fusing the first data vectors corresponding to each of the multiple information modalities and inputting them into the first global decoder for global decoding; prediction process: predicting the user's steady-state demand potential information based on the local value information of the target geographic object output by the multiple first sub-decoders and / or the global value information of the target geographic object output by the first global decoder. By performing local decoding on the first data vectors of each information modality in their respective first sub-decoders, local value information reflecting the local features of each information modality is obtained. Simultaneously, by fusing the first data vectors of multiple information modalities and inputting them into the first global decoder for global decoding, global value information for characterizing the comprehensive features of the multimodal model can be obtained. Therefore, the first information prediction model can utilize features from multiple information modalities simultaneously at both the local and global levels, thereby improving the accuracy and stability of predicting users' steady-state demand potential information. The resulting steady-state demand potential information is richer and more reliable, providing high-quality long-term demand data for subsequent joint information recommendation based on steady-state demand potential information and dynamic interest intent information.

[0103] In this embodiment, the data of different information modalities in the attribute information of the target area, the attribute data of the target geographic object, and the attribute data of similar geographic objects are first classified to obtain the first data of multiple information modalities.

[0104] In this embodiment, the first information prediction model includes a first global decoder, a first encoder, and a first sub-decoder corresponding to each of the multiple information modes. Each information mode has a dedicated first encoder and a first sub-decoder.

[0105] Furthermore, the first data of multiple information modalities are input into the first information prediction model. In the first information prediction model, the first data of different classified information modalities are input into each corresponding first encoder, and through the encoding process, a first data vector corresponding to each information modality is generated.

[0106] Furthermore, in this embodiment, the first data vectors corresponding to each of the multiple information modalities are respectively input to the corresponding first decoder for local decoding. The first decoder is responsible for generating the corresponding local decoding result based on the first data vector of each information modality. Additionally, the data vectors corresponding to the multiple modalities are fused to obtain a first fused vector. Then, the first fused vector is input to the global decoder for global decoding to obtain the global decoding result.

[0107] In this embodiment, global decoding and local decoding are relative concepts, both essentially referring to the results obtained after the model decodes the input features. Local decoding refers to the decoding process performed independently by each information modality at the level of a sub-decoder (e.g., the first sub-decoder), used to recover or reconstruct the information features corresponding to that modality; local decoding corresponds to the local decoding result. Global decoding refers to the decoding process performed uniformly by the global decoder after the local encoding results of each information modality have been fused; global decoding yields the global decoding result.

[0108] In this embodiment, the local decoding result output by the first sub-decoder is local value information, and the global decoding result output by the first global decoder is global value information.

[0109] Furthermore, in this embodiment, the prediction process predicts the user's steady-state demand potential information based on the local value information of the target geographic object output by multiple first decoders and / or the global value information of the target geographic object output by the first global decoder.

[0110] In this embodiment, the steady-state demand information of users can be predicted in different ways. Several implementation methods are given below, but are not limited to these.

[0111] In one implementation, the system relies on a first global decoder for generation. In this embodiment, each first encoder encodes first data vectors of different information modalities. These first data vectors are then fused by the first global decoder and globally decoded to generate the final steady-state demand potential information. Specifically, the first global decoder integrates the data from each information modality and outputs the final steady-state demand potential information by learning the relationships between the various information modalities.

[0112] In another implementation, the steady-state demand potential information is generated based on the local decoding results of the first sub-decoder. Each first sub-decoder performs local decoding on its corresponding first data vector, thereby generating local decoding results for multiple information modes. The output of each first sub-decoder reflects the independent contribution of that information mode to the final steady-state demand potential information. Then, the local decoding results are fused using a weighted summation method to obtain the final steady-state demand potential information. The final steady-state demand potential information is the weighted fusion result of the outputs of each first sub-decoder.

[0113] In another alternative implementation, the first sub-decoder and the first global decoder are combined for generation. The first data vectors corresponding to multiple information modes are locally decoded by their respective sub-decoders, and finally fused and globally decoded in the first global decoder to generate the final steady-state demand potential information.

[0114] Furthermore, in this application embodiment, the dynamic interest intention information of users is predicted based on historical behavior data, and the specific implementation of predicting the dynamic interest intention information of users based on historical behavior data is not limited.

[0115] In one optional embodiment, predicting a user's dynamic interest intent information based on historical behavior data includes: classifying the historical behavior data according to the information modalities presented by the historical behavior data to obtain second data of multiple information modalities; inputting the second data of multiple information modalities into a second information prediction model, the second information prediction model including a second global decoder, second encoders and second sub-decoders corresponding to the multiple information modalities respectively, and performing the following operations within the second information prediction model: encoding process: inputting the second data of multiple information modalities into the second encoders corresponding to the multiple information modalities respectively for encoding to generate second data vectors corresponding to each of the multiple information modalities; decoding process: inputting the second data vectors corresponding to each of the multiple information modalities into the corresponding second sub-decoders for local decoding, and fusing the second data vectors corresponding to each of the multiple information modalities and inputting them into the second global decoder for global decoding; prediction process: predicting the user's dynamic interest intent information based on the local interest intent information output by the multiple second sub-decoders and / or the global interest intent information output by the second global decoder.

[0116] By dividing historical behavioral data into multiple information modalities based on the information modalities presented, and utilizing a second information prediction model comprising a second encoder, a second decomposition decoder, and a second global decoder to perform modal encoding, modal local decoding, and cross-modal global decoding respectively, the system can perform structured processing of user historical behavioral data at three dimensions: modal level, local level, and global level. Compared to existing technologies that typically model historical behavior as a single sequence, this embodiment generates a second data vector for each information modality and obtains local interest intention information, then combines the fused second data vector to obtain global interest intention information. This allows for the extraction of dynamic interest intention information with cross-modal correlations while preserving modal differences.

[0117] In this embodiment, the historical behavior data is first classified according to the information modalities presented by the historical behavior data to obtain second data for multiple information modalities. Further, the second data for multiple information modalities is input into a second information prediction model. In the second information prediction model, the classified second data for different information modalities are respectively input into each corresponding second encoder, and through the encoding process, a second data vector corresponding to each information modality is generated.

[0118] Furthermore, in this embodiment, the second data vectors corresponding to each of the multiple information modalities are input to their respective second sub-decoders for local decoding. The second sub-decoders are responsible for generating corresponding local decoding results based on the second data vector of each information modality. Additionally, the data vectors corresponding to the multiple modalities are fused to obtain a second fused vector. This second fused vector is then input to a second global decoder for global decoding to obtain a global decoding result.

[0119] In this embodiment, the local decoding result output by the second sub-decoder is local interest intent information, and the global decoding result output by the second global decoder is global interest intent information.

[0120] Furthermore, in this embodiment, the prediction process predicts the user's dynamic interest intention information based on the local interest intention information of the target geographic object output by multiple second sub-decoders and / or the global interest intention information of the target geographic object output by the second global decoder.

[0121] In this embodiment, the user's dynamic interest and intent information can be predicted in different ways. Several implementation methods are given below, but are not limited to these.

[0122] In one implementation, generation relies on a second global decoder. In this embodiment, each second encoder encodes second data vectors of different information modalities. These second data vectors are then fused by the second global decoder and globally decoded to generate the final dynamic interest intent information. Specifically, the second global decoder integrates the data from each information modality and outputs the final dynamic interest intent information by learning the relationships between the various information modalities.

[0123] In another implementation, the dynamic interest intent information is generated based on the local decoding results of the second sub-decoder. Each second sub-decoder performs local decoding on its corresponding second data vector, thereby generating local decoding results for multiple information modalities. The output of each second sub-decoder reflects the independent contribution of that information modality to the final dynamic interest intent information. Then, the local decoding results are fused using a weighted summation method to obtain the final stable dynamic interest intent information. The final dynamic interest intent information is the weighted fusion result of the outputs of each second sub-decoder.

[0124] In another alternative implementation, the second sub-decoder and the second global decoder are combined for generation. The second data vectors corresponding to multiple information modalities are locally decoded by their respective second sub-decoders, and finally fused and globally decoded in the second global decoder to generate the final dynamic interest intent information.

[0125] Furthermore, in this embodiment, joint information recommendation is performed on the user based on the user's steady-state demand potential information and dynamic interest intention information, combined with candidate information objects in the target application scenario. Optionally, the user's steady-state demand potential information and dynamic interest intention information are input into the information recommendation model, and the following operations are performed in the information recommendation model: using the first attention network in the information recommendation model, the user's steady-state demand potential information and dynamic interest intention information are weighted and fused to generate a joint embedding vector representing the user's current information needs; using the second attention network in the information recommendation model, attention calculation is performed on the joint embedding vector and the information vectors corresponding to the candidate information objects to obtain the matching degree between the joint embedding vector and each candidate information object; using the recommendation module in the information recommendation model, target information is selected from the candidate information based on the matching degree, and the target information is output.

[0126] By inputting users' steady-state demand potential information and dynamic interest intent information into the information recommendation model, and using a first attention network to weightedly fuse the two types of demand information, a joint embedding vector that better reflects the user's current information needs can be generated. Furthermore, a second attention network calculates the matching degree between the joint embedding vector and the information vectors of candidate information objects, enabling the recommendation process to adaptively consider the differences in characteristics among different candidate information objects. Therefore, this embodiment simultaneously considers the user's long-term needs and short-term interest changes in the selection of candidate information objects, significantly improving the accuracy and relevance of the recommendation results. In this embodiment, the first attention network in the information recommendation model is used to perform weighted fusion of user demand information at different time scales. Specifically, the first attention network performs weighted fusion of data from two different time scales based on steady-state demand potential information and dynamic interest intent information to obtain a joint embedding vector that can comprehensively represent user needs. This joint embedding vector comprehensively represents the user's demand performance at both long-term and short-term time scales, achieving a unified representation of the user's current information needs.

[0127] In one optional embodiment, when using a first attention network to weightedly fuse the user's steady-state demand potential information and dynamic interest intention information to generate a joint embedding vector representing the user's current information needs, the process includes: inputting the user's steady-state demand potential information and dynamic interest intention information into the first attention network, and performing the following operations in the first attention network: generating a gating signal based on the dynamic interest intention information, and using the gating signal to modulate the steady-state demand potential information element-wise to obtain an activated steady-state demand potential vector; performing cross-attention calculation on the activated steady-state demand potential vector and the dynamic interest intention information to obtain a context semantic vector; concatenating the activated steady-state demand potential vector, the context semantic vector, and the global aggregated representation of the dynamic interest intention information, and inputting them into multiple parallel fusion subnetworks to obtain multiple candidate fusion vectors; generating routing weights based on the steady-state demand potential information and the dynamic interest intention information, and weighted summing the multiple candidate fusion vectors according to the routing weights to obtain a joint embedding vector.

[0128] Therefore, the first attention network modulates the steady-state demand potential information element-wise using gating signals, highlighting features with higher correlation to dynamic interest intent information. Then, a contextual semantic vector is obtained through cross-attention calculation to express the relationship between the two types of demand information. The activated steady-state demand potential vector, contextual semantic vector, and the global aggregated representation of dynamic interest intent information are input into multiple parallel fusion sub-networks to generate candidate fusion vectors. Finally, the candidate fusion vectors are weighted and summed using routing weights generated based on the two types of demand information. The joint embedding vector obtained through the above processing can more accurately represent the user's current information needs, improving the accuracy of subsequent recommendation matching.

[0129] In this embodiment, a gating signal is generated based on dynamic interest intent information, and this gating signal is used to modulate the steady-state demand potential information element-wise to obtain an activated steady-state demand potential vector. The gating signal is used to adjust the contribution level of each component of the steady-state demand potential information along the feature dimension, so that steady-state demand features with high correlation to the current dynamic interest intent information are amplified during the element-wise modulation process, while irrelevant or weakly correlated features are relatively suppressed. Therefore, the activated steady-state demand potential vector represents the reweighted steady-state demand expression under the current dynamic interest context, providing a more suitable steady-state demand input for subsequent cross-attention calculations with dynamic interest intent information.

[0130] In this embodiment, after obtaining the activated steady-state demand potential vector, the first attention network can further perform cross-attention calculation on the activated steady-state demand potential vector and dynamic interest intent information to obtain a contextual semantic vector. Here, cross-attention calculation can be understood as an attention mechanism used to model the correlation between features from different sources. The contextual semantic vector obtained through the cross-attention mechanism can be regarded as a semantic association enhancement representation of the activated steady-state demand potential vector within a dynamic interest context.

[0131] Furthermore, the first attention network can concatenate the activated steady-state demand potential vector, the aforementioned contextual semantic vector, and the global aggregated representation of dynamic interest intent information, and input the concatenated feature vector into multiple parallel fusion sub-networks. These multiple fusion sub-networks generate corresponding candidate fusion vectors. Given multiple candidate fusion vectors, corresponding routing weights are generated based on the steady-state demand potential information and the dynamic interest intent information. These routing weights indicate the contribution ratio of each candidate fusion vector to the final fusion result. The routing weights can be understood as an adaptive set of weighted coefficients derived from the joint features of the steady-state demand potential information and the dynamic interest intent information. By performing a weighted summation on each candidate fusion vector, the final joint embedding vector is obtained. This joint embedding vector simultaneously integrates long-term steady-state demand, short-term dynamic interest, and the semantic association between them, serving as a unified representation of the user's current information demand state and providing direct input for subsequent candidate information object matching and joint recommendation.

[0132] Furthermore, in this embodiment, a second attention network in the information recommendation model is used to perform attention calculation on the joint embedding vector and the information vectors corresponding to the candidate information objects to obtain the matching degree between the joint embedding vector and each candidate information object. In an optional embodiment, the joint embedding vector is input into the second attention network, and the following operations are performed in the second attention network: semantic parsing is performed on the information vector of each candidate information object to obtain the sub-vector representation of each candidate information object at at least two different semantic granularity levels; based on the joint embedding vector and the sub-vector representation of each candidate information object at each semantic granularity level, the absolute matching degree between each candidate information object and the joint embedding vector is calculated; based on the information vector of each candidate information object, a competitive context representation of each candidate information object relative to other candidate information objects is constructed; the matching degree between the joint embedding vector and the competitive context representation of each candidate information object is calculated to obtain the relative matching degree between each candidate information object and the joint embedding vector considering the competitive relationship; based on the absolute matching degree and relative matching degree between each candidate information object and the joint embedding vector, the matching degree between each candidate information object and the joint embedding vector is generated.

[0133] By inputting the joint embedding vector into the second attention network and performing semantic parsing on the information vectors of candidate information objects, each candidate information object can be represented by a sub-vector at different semantic granularity levels, enhancing the hierarchical nature of semantic expression. The absolute matching degree calculated based on the joint embedding vector and multi-granularity sub-vectors reflects the degree of direct matching between the candidate information object and the user's current information needs. Simultaneously, by constructing a competitive context representation using the information vectors of candidate information objects and calculating the relative matching degree, the relative competitive position of the candidate information object within the candidate set can be reflected. Furthermore, by combining the matching degree generated from the absolute and relative matching degrees, target information can be selected while considering both the degree of direct matching and the relative competitive position within the set, significantly improving the recommendation effect.

[0134] In this optional embodiment, the second attention network can perform semantic parsing on the information vector corresponding to each candidate information object to obtain the sub-vector representation of the candidate information object at at least two different semantic granularity levels. The semantic granularity level can be understood as different levels of abstraction or different content dimensions of the information vector of the candidate information object. Specifically, the semantic granularity level can be constructed as a hierarchical structure from coarse to fine, including a coarse-grained level, a medium-grained level, and a fine-grained level, which are used to represent the category-level semantics, brand or attribute-level semantics, and specific instance-level semantics of the candidate information object, respectively.

[0135] In this embodiment, the absolute matching degree between each candidate information object and the joint embedding vector is calculated based on the joint embedding vector and the sub-vector representation of each candidate information object at each semantic granularity level. The absolute matching degree can be used to characterize the degree of direct matching between a single candidate information object and the joint embedding vector.

[0136] In one optional embodiment, calculating the absolute matching degree between each candidate information object and the joint embedding vector based on the joint embedding vector and the sub-vector representation of each candidate information object at each semantic granularity level includes: for each candidate information object, calculating the matching degree between the joint embedding vector and the sub-vector representation of the candidate information object at each semantic granularity level to obtain the local matching degree between the candidate information object and the joint embedding vector at each semantic granularity level; dynamically generating the fusion weight of the candidate information object at each semantic granularity level through a gating network based on the joint embedding vector and / or the context information of the candidate information object; and weighting and fusing the local matching degree between the candidate information object and the joint embedding vector at each semantic granularity level according to the fusion weight to obtain the absolute matching degree between the candidate information object and the joint embedding vector.

[0137] By calculating the matching degree between the joint embedding vector and the sub-vector representations of the candidate information object at different semantic granularities, multiple local matching degrees of the candidate information object can be obtained at a fine-grained level, thus reflecting its relevance at different semantic granularities. Furthermore, the fusion weights generated based on the joint embedding vector and / or the contextual information of the candidate information object can adaptively adjust the contribution ratio of each semantic granularity level in the final matching calculation, thereby highlighting more representative semantic levels. Finally, the absolute matching degree obtained by weighted fusion of the local matching degrees at each level can more comprehensively and accurately characterize the direct matching degree between the candidate information object and the user's current information needs. Therefore, the absolute matching degree can more accurately distinguish the degree of association between different candidate information objects in the candidate set, avoiding the traditional reliance on a single semantic level and improving the discriminative power and accuracy of the recommendation results.

[0138] In this embodiment, for each candidate information object, the second attention network can calculate the matching degree between the joint embedding vector and the sub-vector representation of the candidate information object at each semantic granularity level, thereby obtaining the local matching degree between the candidate information object and the joint embedding vector at each semantic granularity level. The local matching degree is used to characterize the degree of semantic matching between the semantic sub-representation of the candidate information object and the user's current information needs (represented by the joint embedding vector) at the corresponding semantic granularity level.

[0139] In this embodiment, the second attention network can dynamically generate fusion weights for the candidate information object at various semantic granularity levels through a gating network, based on the joint embedding vector and / or the context information of the candidate information object. The fusion weights are used to adjust the contribution of different semantic granularity levels in the absolute matching degree calculation, and can be implemented as a set of adaptive hierarchical weighting coefficients.

[0140] Having obtained the fusion weights, the second attention network can use these weights to weight and fuse the local matching degrees of the candidate information object with the joint embedding vector at each semantic granularity level, thus obtaining the absolute matching degree between the candidate information object and the joint embedding vector. By using the fusion weights to weight the local matching degrees at each semantic granularity level, the absolute matching degree can comprehensively reflect the matching information of the candidate information object at different semantic granularity levels and adaptively adjust the influence of each semantic level on the final fusion result, so that the absolute matching degree can accurately characterize the direct matching degree between the candidate information object and the joint embedding vector.

[0141] Furthermore, to characterize the relative competitive relationships among candidate information objects within the candidate set, the second attention network can also construct a competitive context representation of each candidate information object relative to other candidate information objects based on the information vector of that candidate information object. This competitive context representation is a differentiated representation obtained by comparing and modeling the current candidate information object based on the features of other candidate information objects, and is used to characterize its relative competitive position.

[0142] In this embodiment, after obtaining the competitive context representation of each candidate information object, the second attention network can calculate the matching degree between the joint embedding vector and the competitive context representation to obtain the relative matching degree between each candidate information object and the joint embedding vector. The relative matching degree is used to measure the degree to which the current candidate information object matches the user's current information needs relative to other candidate information objects, thereby reflecting the relative priority of the candidate information object within the candidate set.

[0143] Furthermore, based on the absolute and relative matching degrees of each candidate information object and the joint embedding vector, a matching degree between each candidate information object and the joint embedding vector is generated. This matching degree comprehensively considers both the direct matching degree between the candidate information object and the joint embedding vector and the relative competitive position of the candidate information object in the candidate set, and is used to comprehensively characterize the overall matching situation between each candidate information object and the joint embedding vector.

[0144] In one optional embodiment, based on the geographic semantic features, distance features, and geographic context features of the geographic object description information, data augmentation is performed on the geographic object description information in the ontology library using geographic object description information from a third-party library, including: S11. For any geographic object description information in the third-party library, recall similar geographic object description information in the ontology library to obtain at least two geographic object description information, wherein the geographic object description information includes at least the name information, address information and spatial location information of the geographic object. S12. Using a geographic semantic feature encoder, calculate the geographic semantic similarity between the at least two geographic object description information based on the name information and address information in the at least two geographic object description information; S13. Using a distance feature encoder, calculate the spatial distance feature between the at least two geographic object description information based on the spatial location information in the at least two geographic object description information; S14. Using a geographic context feature encoder, calculate the geographic context similarity between the description information of the at least two geographic objects and their surrounding geographic environment information. S15. Input the geographic semantic similarity, spatial distance features and geographic context similarity into the classification head network, and output the classification result of whether the description information of the at least two geographic objects points to the same geographic object; S16. Based on the classification results, the geographic object description information in the ontology library is augmented using the geographic object description information belonging to the third-party library from the at least two geographic object description information to obtain the target library.

[0145] In this embodiment, for each geographic object description information in the third-party library, based on the rule matching strategy and / or text similarity matching strategy, at least one similar geographic object description information is recalled in the ontology library to form the above-mentioned at least two geographic object description information.

[0146] In some embodiments, similar geographic object description information is filtered (or recalled) using certain methods as at least two geographic object description information to be processed in this application embodiment. For example, based on any geographic object description information in the ontology library, similar geographic object description information can be recalled from various third-party libraries, and then the any geographic object description information and all the recalled similar geographic object description information are used together as at least two geographic object description information to be processed in this application embodiment. As another example, based on any geographic object description information in the ontology library, similar geographic object description information can be recalled from each third-party library, and then for each third-party library, the any geographic object description information and the similar geographic object description information recalled from that third-party library are used together as at least two geographic object description information to be processed in this application embodiment. For example, for any third-party library, based on any geographic object description information in that third-party library, similar geographic object description information is retrieved from the ontology library. Then, that geographic object description information and all similar geographic object description information retrieved from the ontology library are used together as at least two geographic object description information to be processed in this embodiment. For example, for any third-party library, based on any geographic object description information in that third-party library, similar geographic object description information is retrieved from the ontology library. Then, that geographic object description information and one similar geographic object description information retrieved from the ontology library are used together as at least two geographic object description information to be processed in this embodiment.

[0147] In this embodiment, the geographic object description information includes name information, address information, and spatial location information. To fully utilize the geographic object description information, this embodiment uses a neural network model to extract features from three dimensions: name and address semantics, spatial location information, and surrounding geographic environment. This fully utilizes the features extracted from different information dimensions, and the neural network model uses these extracted features to accurately identify geographic objects.

[0148] It should be noted that the implementation form of the neural network model used in this application embodiment is not limited, and it can be various deep learning-based neural network models. Optionally, the various neural network models used can be deep learning models with relatively small parameter scales or deep learning models with relatively large parameter scales. The large model is merely an example, and this application embodiment does not limit the number of model parameters supported by the deep learning model used, aiming to meet actual needs. The deep learning models involved in this application embodiment are not limited to a specific type, and different models can be selected according to implementation needs. For example, an artificial intelligence-based language model (LM) can be used, and of course, a multimodal model (MM) capable of processing multiple modal information can also be used.

[0149] In this embodiment, the neural network model used includes at least: a geographic semantic feature encoder, a distance feature encoder, a geographic context feature encoder, and a classification head network. These neural network models cooperate and collaborate to determine whether the descriptive information of at least two geographic objects points to the same geographic object. Each model contributes to improving the accuracy of the determination. The role of each model is explained in detail below.

[0150] In this embodiment, a geographic semantic feature encoder is used to calculate the geographic semantic similarity between at least two geographic object descriptions based on the name and address information in the descriptions of at least two geographic objects.

[0151] Among them, the name and address information in the geographic object description information is relatively text-based. Therefore, the name and address information in at least two geographic object description information are input into the geographic semantic feature encoder to calculate the semantic similarity between the at least two geographic object description information.

[0152] For example, taking a residential community as a geographic object, geographic object description information for two communities is obtained from different database sources. For each community's geographic object description information, the corresponding name and address information are extracted. In some embodiments, to facilitate semantic similarity calculation, the extracted name and address information are formatted into a structured input sequence containing preset tags, for example, [CLS] Name: <Name Information> Address: <Address Information> [SEP]. Here, [CLS] indicates the start of the sequence and can also be considered a classification tag, representing the aggregated information of the entire sequence. [SEP] is used to separate different geographic object description information and can be called a segmentation tag. <Name Information> and <Address Information> refer to the name and address information of the geographic object, respectively.

[0153] In this embodiment, geographic object description information A1 is abbreviated as A1, and its corresponding name information is Community A; its corresponding address information is XX Road XX No. . Geographic object description information A2 is abbreviated as A2, and its corresponding name information is Community A, Zone 2; its corresponding address information is YY Road YY No. . Optionally, any of the above geographic object description information can be formatted as: [CLS] Name: <Name Information> Address: <Address Information> [SEP]. For example, [CLS] Name: Community A; Address: XX Road XX No. [SEP]. Wherein, [CLS] and [SEP] are two identifiers. [CLS] is used to indicate the beginning of the input sequence, and its corresponding final hidden state is used as the global semantic representation vector of the entire input sequence pair, which facilitates the output of the global semantic representation vector by the geographic semantic feature encoder for subsequent similarity calculation or classification decision; [SEP] is used to separate the description information of different geographic objects in the input sequence, so that the geographic semantic feature encoder can distinguish the text boundaries of different geographic object description information and effectively model the semantic similarity between different geographic object description information.

[0154] In this embodiment, a geographic semantic feature encoder is used to semantically encode descriptions of at least two geographic objects and obtain the geographic semantic similarity between them. For example, the geographic semantic feature encoder can extract the global semantic representation vector at the [CLS] location to calculate the geographic semantic similarity between at least two geographic object descriptions. Alternatively, semantic representation vectors from all locations can be taken and weighted and fused to calculate the geographic semantic similarity between at least two geographic object descriptions. The geographic semantic similarity between at least two geographic object descriptions deeply integrates the name and address information of the at least two geographic object descriptions and captures their semantic similarity. In this embodiment, a distance feature encoder is used to calculate the spatial distance feature between at least two geographic object descriptions based on their spatial location information. In this embodiment, the spatial location information of the at least two geographic object descriptions is used as input to the distance feature encoder, which calculates the spatial distance feature between the at least two geographic object descriptions.

[0155] In this embodiment, spatial location information refers to information describing the spatial location of a geographic object, such as the latitude and longitude information of the geographic object, but not limited to this. In addition to latitude and longitude information, spatial location information can also be implemented as relative location information reflecting the spatial relationship between the geographic object and surrounding roads, landmarks or other reference objects.

[0156] In an optional embodiment, taking the latitude and longitude of a geographic object as an example, the Haversian distance or spherical cosine distance between the latitude and longitude of any two geographic objects can be calculated, and the spatial distance characteristics between any two geographic objects can be determined based on the calculation result.

[0157] Furthermore, if two address entity descriptions point to the same geographic object in the real world, their surrounding geographic information will highly overlap. This surrounding geographic information can include, for example, points of interest (POIs) such as shops, schools, and bus stops; or road structure information such as road density, road type, and the location of transportation hubs—there are no limitations on this.

[0158] In this embodiment, based on the address information in the description information of each geographic object, a map service interface is invoked to obtain the surrounding environmental information within a preset distance range around the geographic object described in the geographic object description information. The preset distance range can be 300 meters, 500 meters, 1 kilometer, etc., and is not limited thereto; it can be flexibly set according to accuracy requirements or coverage requirements.

[0159] Furthermore, using a geographic context feature encoder, the geographic context similarity between at least two geographic object descriptions is calculated based on the geographic environment information surrounding each of the at least two geographic objects.

[0160] Specifically, through the geographic context feature encoder, not only can the descriptive information of each geographic object be fused with its corresponding geographic environment information to obtain a feature vector that can represent the context of the geographic object, but also feature interaction processing can be performed on at least two geographic context feature vectors to construct a feature representation that reflects the contextual relationship between at least two geographic objects.

[0161] The feature interaction processing can include, but is not limited to, concatenating, subtracting, or combining concatenation and subtraction of at least two geographic context feature vectors. Through feature interaction processing, a feature vector is generated that characterizes the degree of similarity between the geographic objects described by at least two geographic object descriptions in terms of their geographic environment, i.e., geographic context similarity.

[0162] In this embodiment, geographic semantic similarity, spatial distance features, and geographic context similarity are input into the classification head network, which performs classification processing and finally outputs the classification result of whether at least two geographic object descriptions point to the same geographic object.

[0163] In one optional embodiment, the classification head network is implemented as a general classification head network. It can then employ a feature fusion approach, concatenating geographic semantic similarity features, spatial distance features, and geographic context similarity features to form a fused feature vector. This vector is then input into the classification head network for classification, outputting a classification result indicating whether the descriptions of two geographic objects refer to the same geographic object. In this optional embodiment, the classification head network can adopt a multilayer perceptron structure. Using a general classification head network for classification processing offers advantages such as simplicity and high efficiency.

[0164] In another optional embodiment, a hierarchical fusion architecture classification head network is designed, comprising multiple sub-classification head networks and a total classification head network. Geographic semantic similarity features, spatial distance features, and geographic context similarity features can be input into different sub-classification head networks to obtain corresponding initial classification results. Further, the total classification head network fuses and predicts the outputs of the multiple sub-classification head networks to obtain a classification result, which indicates whether the descriptive information of two geographic objects points to the same geographic object. Even more optionally, during the fusion and prediction process of the total classification network's outputs from the multiple sub-classification head networks, the influence of geographic semantic similarity features, spatial distance features, and geographic context similarity features can be comprehensively considered. That is, the geographic semantic similarity features, spatial distance features, and geographic context similarity features, along with the initial classification results output from the multiple sub-classification head networks, are simultaneously input into the total classification head network for fusion and prediction to obtain the final classification result. Among them, multiple sub-classification head networks generate local discrimination results based on different semantic dimensions of geographic object description information. The overall classification head network performs nonlinear fusion and confidence calibration on the local discrimination results and outputs globally consistent classification decisions. This can effectively overcome the defects of single semantic dimensions being susceptible to noise interference or information loss, and significantly improve the accuracy, robustness and scene adaptability of geographic object matching.

[0165] In this embodiment, based on the classification results, data augmentation is performed using geographic object description information from at least two geographic object description information ontology databases belonging to third-party databases to obtain the target database.

[0166] In this embodiment, if the classification result indicates that at least two geographic object descriptions point to the same geographic object, the geographic object descriptions belonging to the third-party library and the geographic object descriptions belonging to the ontology library are associated to obtain the target library. Specifically, the geographic object descriptions from the third-party library are used as supplementary geographic object descriptions for the same geographic object in the ontology library, thus associating the geographic descriptions from the third-party library with those from the ontology library to form the target library. After the association process, the generated target library will contain geographic object descriptions from both the third-party library and the ontology library. This improves the mounting rate and accuracy of the target library. If at least two geographic object descriptions point to different geographic objects, the geographic object descriptions from the third-party library are added to the ontology library. In this way, the geographic object descriptions in the ontology library can be expanded.

[0167] In this embodiment, for at least two geographic object descriptions, a geographic semantic feature encoder, a distance feature encoder, and a geographic context feature encoder are used to encode name and address semantics, spatial location, and surrounding geographic environment to obtain geographic semantic similarity, spatial distance features, and geographic context similarity. These are then input into a classification head network to output a classification result indicating whether the at least two geographic object descriptions point to the same geographic object. Based on the classification result, data augmentation is performed on the geographic description information of the ontology database to obtain the target database. Thus, by introducing a model to learn the relationships between different features, multi-dimensional features such as geographic semantics, spatial location, and surrounding geographic environment are fully utilized. Feature fusion and accurate judgment are achieved in the classification head network, improving the accuracy and robustness of geographic object matching, and increasing the loading rate and accuracy of the target database.

[0168] This application provides a model training method for a first or second information prediction model. Training the model using this method yields either the first or second information prediction model. The training process of the first information prediction model is described below as an example. The training method for the second information prediction model can be referenced below, the difference being the different training sample data and target results.

[0169] One model training method includes the following steps: S21. Obtain training sample data. The training sample data corresponds to the target result in the target application scenario. The target result is the preset value information corresponding to the training sample data. S22. Classify the training sample data according to the information modality presented by the training sample data; S23. Input the training sample data into the initial model. The initial model includes a first global decoder, a first encoder and a first sub-decoder corresponding to multiple information modalities, and the training process includes: S24. Encoding process: Input the training sample data into the first encoder corresponding to the multiple information modalities according to the above classification categories for encoding, and generate the data vectors corresponding to the multiple information modalities respectively. S25. Decoding process: The data vectors corresponding to each of the multiple information modes are input to the corresponding first decoder for local decoding, and the data vectors corresponding to each of the multiple information modes are fused and input to the first global decoder for global decoding. S26. Optimization process: Adjust the parameters of the corresponding first encoder according to the sub-residual between the local decoding results of the multiple first decoders and the target result, and adjust the parameters of the multiple first encoders according to the total residual between the first global decoding result and the target result; iteratively execute the encoding process and the decoding process until the total residual meets the set value to obtain the first information prediction model.

[0170] In this embodiment, the application scenarios of the above-described model training method are not limited. For example, it can be used in scenarios such as image and text generation, information prediction, information recommendation, e-commerce marketing, auction marketing, and various information evaluations. The target result refers to the expected result corresponding to the training sample data, which is the labeled result of the training sample data. The training sample data and the target result will vary depending on the application scenario and model function. For example, in an information recommendation scenario, an information recommendation model can be trained based on the training sample data, and the corresponding target result can be the recommendation result corresponding to the training sample data. Specifically, taking product information recommendation in an e-commerce marketing scenario as an example, the training sample data can be user profile data, historical behavior data, etc., and the corresponding target result can be recommended product information. Therefore, the method of this embodiment can be used to train a product information recommendation model that can recommend product information to users based on user profile data, historical behavior data, etc.

[0171] In this embodiment, the training sample data is a dataset used for model training. Each training sample data set includes multiple data categories, each presenting a different information modality. An information modality refers to the form of the data. In some embodiments of this application, the information modality can be simply referred to as a modality. Classifying different types of data yields information data with multiple modalities.

[0172] Different application scenarios have corresponding target results. In the image-text generation scenario, the training sample data includes two information modalities: image data and text prompts. The model input is an image and its corresponding text description, and the output is a generated image that is semantically consistent with the prompts. In the product recommendation scenario, the training sample data includes user attribute information, user historical behavior data, and product image-text information. The model input is user profiles, historical behavior, and product content features, and the target result is the annotation of user interaction results with the target product, such as behavioral indicators like clicks, favorites, or purchases. In the information evaluation scenario (e.g., geographic object value evaluation), the training sample data includes location, supporting environment, and architectural features. The model input is a multimodal feature set of the geographic object, and the target result is the annotated quantitative result, such as a numerical evaluation value.

[0173] Furthermore, the training sample data is input into the initial model. The initial model includes a first global decoder, and first encoders and first sub-decoders corresponding to multiple information modalities. Each information modality has a dedicated first encoder and first sub-decoder.

[0174] In this embodiment, the different modal data that have been classified are input into each corresponding first encoder, and a data vector for each modality is generated through the encoding process.

[0175] In this embodiment, the data vectors corresponding to each of the multiple modalities are input to their respective first decoders for local decoding. The first decoders are responsible for generating local decoding results based on the encoding vector of each modality. Furthermore, the data vectors corresponding to the multiple modalities are fused to obtain a fused vector. This fused vector is then input to a global decoder for global decoding to obtain a global decoding result.

[0176] In this embodiment, global decoding and local decoding are relative concepts, both essentially referring to the results obtained after the model decodes the input features. Local decoding refers to the decoding process performed independently by each modality at the first decoder level, used to recover or reconstruct the information features corresponding to that modality; local decoding corresponds to the local decoding result. Global decoding refers to the decoding process performed uniformly by the first global decoder after the encoding results of each modality have been fused; global decoding yields the global decoding result.

[0177] In one example, in a text-to-image generation scenario, the local decoding results and the global decoding results can be represented as image data.

[0178] Each information modality corresponds to a first sub-decoder that can perform local decoding on the structural information, texture details, or semantic layout of the image, forming multiple local decoding results; the first global decoder generates a generated image based on the fused fusion vector, and outputs a generated image that is consistent with the semantics of the prompt word.

[0179] In this embodiment, the parameters of the corresponding first encoder are adjusted based on the sub-residuals between the local decoding results and the target result of each of the multiple first decoders. The sub-residual refers to the error between the local decoding result of the first decoder and the target result. The sub-residual is fed back to the corresponding first encoder through backpropagation to adjust the parameters of the first encoder, thereby reducing the error.

[0180] In this embodiment, the parameters of multiple first encoders are adjusted based on the total residual between the global decoding result and the target result. The total residual is the error between the global decoding result and the target result; the parameters of the multiple first encoders are adjusted after backpropagation.

[0181] In this embodiment, the method of adjusting the parameters of the first encoder based on the total residual and the sub-residual is not limited.

[0182] In one optional implementation, the parameters of the corresponding first encoders are first adjusted using sub-residuals, and then optimized using the total residual. Sub-residuals can quickly correct the feature learning of a single modality at the local level, making the encoded representation of each information modality more accurate and stable. After the sub-residual adjustment is completed, all first encoders are uniformly adjusted based on the total residual between the global decoding result and the target result, coordinating the feature distribution among modalities at the overall level. By optimizing the model layer by layer from "intra-modal refinement" to "inter-modal fusion," the overall training efficiency and the consistency of the final output can be improved.

[0183] In another optional implementation, the sub-residuals corresponding to multiple first decoders can be jointly calculated with the total residual corresponding to the first global decoder to generate a comprehensive target residual. This target residual can be fused with the sub-residuals and the total residual according to a preset weight coefficient. The model adjusts the parameters of each first encoder in reverse according to the target residual, thereby taking into account both the accuracy of local modal features and the consistency of global feature representation, and improving the generalization performance of the model in multimodal tasks.

[0184] In another optional implementation, the adjustment order can be dynamically determined based on the relative magnitudes of the sub-residuals and the total residuals during training. For example, when the sub-error of a single mode is large, the sub-residuals are used first to adjust the parameters of the corresponding first encoder; when the total residual is large, the total residual is used first for overall optimization. By evaluating the changing trends of the sub-residuals and the total residuals, the optimization order is adaptively determined, thereby improving the effectiveness of parameter adjustment and the training convergence speed.

[0185] In this embodiment, the above encoding and decoding processes are executed iteratively. The total residual after each training iteration is calculated and compared with a preset value. The training ends when the total residual meets the preset value, thus obtaining the target model.

[0186] In this application embodiment, a hybrid model architecture of multiple encoders, multiple sub-decoders, and a global decoder is proposed. Sample data of different information modalities correspond to different encoders and sub-decoders, transforming the mode of a single model processing all training sample data into a divide-and-conquer mode. The inherent complexity of each encoder and sub-decoder is low, which can reduce the complexity of model training and save resource consumption. Moreover, different encoders and sub-decoders can be trained in parallel, which is beneficial to shorten the model training time. Furthermore, by utilizing the dual decoding mechanism of sub-decoders and global decoders, the encoder parameters can be continuously adjusted through local and global optimization to optimize the model performance, improve the accuracy of inference results, and accelerate the model convergence speed, thereby further improving the model training efficiency and saving model training time.

[0187] It should be noted that the execution subject of each step of the method provided in the above embodiments can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 701 to 703 can be device A; or the execution subject of steps 701 and 702 can be device A, and the execution subject of step 703 can be device B; and so on.

[0188] Furthermore, in some of the processes described in the above embodiments and accompanying drawings, multiple operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or they may be executed in parallel. The operation numbers, such as 701, 702, etc., are merely used to distinguish different operations and do not represent any execution order. Additionally, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first" and "second" in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0189] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 9 As shown, in practice, this electronic device includes a memory 94 and a processor 95.

[0190] Memory 94 is used to store computer programs and can be configured to store various other data to support operation on the electronic device. Examples of this data include instructions for any application or method used to operate on the electronic device, data structures, contact data, phone book data, messages, pictures, videos, etc.

[0191] Processor 95, coupled to memory 94, is used to execute computer programs in memory 94 for: acquiring user profile data and historical behavior data, wherein the profile data includes target geographic objects associated with the user, the attribute data of the target geographic objects reflects the preliminary value information of the target geographic objects, and the historical behavior data reflects the user's dynamic interest intentions; determining similar geographic objects and their quantities based on the attribute data of the target geographic objects, wherein the attribute data of the target geographic objects at least includes the target region to which the target geographic object belongs; generating a value trend change coefficient for the target region based on the fitted curve of the historical transaction price versus time relationship of other geographic objects in the target region; determining the missing degree of similar geographic objects based on the number of similar geographic objects and the number of transactions in the target region; adjusting the preliminary value information using the missing degree as an adjustment coefficient and the product of the statistical value of the historical transaction price of the similar geographic objects and the trend change coefficient as an adjustment amount to obtain target value information, wherein the target value information reflects the user's steady-state demand potential; and performing joint information recommendation for the user from two dimensions: dynamic interest intentions and steady-state demand potential, based on the historical behavior data and the target value information of the target geographic objects.

[0192] For detailed implementation of each of the above operations, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0193] Furthermore, such as Figure 9 As shown, the electronic device also includes other components such as a communication component 96, a display 97, a power supply component 98, and an audio component 99. Figure 9 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 9 The components shown. Additionally... Figure 9 The components within the dashed box are optional, not mandatory, and their specific requirements depend on the product form of the electronic device. The electronic device in this embodiment can be a terminal device such as a desktop computer, laptop computer, smartphone, or IoT device, or a server-side device such as a conventional server, cloud server, or server array. If the electronic device in this embodiment is a terminal device such as a desktop computer, laptop computer, or smartphone, it may include... Figure 9 The components within the dashed box; if the electronic device in this embodiment is implemented as a conventional server, cloud server, or server array, etc., it may be omitted. Figure 9 The component within the dashed box.

[0194] The aforementioned memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0195] The aforementioned communication component is configured to facilitate wired or wireless communication between the device containing the communication component and other devices. The device containing the communication component can access wireless networks based on communication standards, such as 2G, 3G, 4G / LTE, 5G, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0196] The aforementioned display includes a screen, which may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a Touch Panel, the screen can be implemented as a touchscreen to receive input signals from the user. The Touch Panel includes one or more touch sensors to sense touches, swipes, and gestures on the Touch Panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0197] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0198] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0199] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium. Accordingly, this application also provides a computer program product, which includes a computer program or instructions that, when executed by a processor, cause the processor to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above method embodiments.

[0200] It should also be noted that 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 process, method, article, or apparatus. Unless otherwise specified, 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.

[0201] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An information recommendation method, characterized in that, include: Acquire user profile data and historical behavior data, wherein the profile data includes the identification information of target geographic objects associated with the user; According to the identification information, the attribute data of the target geographic object is obtained from the target library. The target library is obtained by using the geographic semantic features, distance features and geographic context features of the geographic object description information to perform data augmentation on the geographic object description information in the ontology library using the geographic object description information in the third-party library. The geographic object description information includes the attribute data of the corresponding geographic object. Based on the attribute data of the target geographic object, predicting the user's steady-state demand potential information includes: determining similar geographic objects of the target geographic object based on the attribute data of the target geographic object, wherein the attribute data of the target geographic object at least includes the target region to which the target geographic object belongs; classifying the attribute information of the target region, the attribute data of the target geographic object, and the attribute data of the similar geographic objects to obtain first data of multiple information modalities; inputting the first data of multiple information modalities into a first information prediction model, wherein the first information prediction model includes a first global decoder, a first encoder and a first sub-decoder corresponding to the multiple information modalities respectively, within the first information prediction model. The system performs the following operations: Encoding process: The first data of the multiple information modalities are respectively input into the first encoder corresponding to the multiple information modalities for encoding, generating the first data vector corresponding to each of the multiple information modalities; Decoding process: The first data vector corresponding to each of the multiple information modalities is respectively input into the corresponding first sub-decoder for local decoding, and the first data vector corresponding to each of the multiple information modalities is fused and then input into the first global decoder for global decoding; Prediction process: Based on the local value information of the target geographic object output by the multiple first sub-decoders and the global value information of the target geographic object output by the first global decoder, the steady-state demand potential information of the user is predicted; Based on the historical behavior data, predict the user's dynamic interest and intent information; Based on the user's steady-state demand potential information and dynamic interest intent information, combined with candidate information objects in the target application scenario, joint information recommendation is performed for the user.

2. The method according to claim 1, characterized in that, Based on the historical behavior data, predict the user's dynamic interest and intent information, including: Based on the information modalities presented by the historical behavior data, the historical behavior data is classified to obtain second data with multiple information modalities; The second data of the multiple information modalities are input into the second information prediction model, which includes a second global decoder, a second encoder and a second sub-decoder corresponding to the multiple information modalities, and the following operations are performed within the second information prediction model: Encoding process: The second data of the multiple information modalities are respectively input into the second encoder corresponding to the multiple information modalities for encoding, thereby generating the second data vector corresponding to each of the multiple information modalities; Decoding process: The second data vectors corresponding to each of the multiple information modes are input to the corresponding second sub-decoders for local decoding, and the second data vectors corresponding to each of the multiple information modes are fused and input to the second global decoder for global decoding; Prediction process: Based on the local interest intent information output by the multiple second sub-decoders and / or the global interest intent information output by the second global decoder, predict the user's dynamic interest intent information.

3. The method according to claim 1, characterized in that, Also includes: Acquire training sample data, wherein the training sample data corresponds to a target result in the target application scenario, and the target result is a preset value information corresponding to the training sample data; The training sample data is classified according to the information modality presented by the training sample data; The training sample data is input into an initial model, which includes a first global decoder, a first encoder and a first sub-decoder corresponding to multiple information modalities, and the training process includes: Encoding process: The training sample data are input into the first encoder corresponding to the multiple information modalities according to the above classification categories for encoding, so as to generate the data vectors corresponding to each of the multiple information modalities; Decoding process: The data vectors corresponding to each of the multiple information modes are input to the corresponding first decoder for local decoding, and the data vectors corresponding to each of the multiple information modes are fused and input to the first global decoder for global decoding; Optimization process: Adjust the parameters of the corresponding first encoder according to the sub-residual between the local decoding results of each of the multiple first decoders and the target result, and adjust the parameters of the multiple first encoders according to the total residual between the first global decoding result and the target result; The encoding and decoding processes are executed iteratively until the total residual satisfies a set value, so as to obtain the first information prediction model.

4. The method according to any one of claims 1-3, characterized in that, Based on the user's steady-state demand potential information and dynamic interest intent information, combined with candidate information objects in the target application scenario, joint information recommendation is performed for the user, including: The user's steady-state demand potential information and dynamic interest intention information are input into the information recommendation model, and the following operations are performed in the information recommendation model: Using the first attention network in the information recommendation model, the steady-state demand potential information and dynamic interest intention information of the user are weighted and fused to generate a joint embedding vector representing the user's current information demand; Using the second attention network in the information recommendation model, attention is calculated on the joint embedding vector and the information vector corresponding to the candidate information object to obtain the matching degree between the joint embedding vector and each candidate information object; Using the recommendation module in the information recommendation model, target information is selected from the candidate information based on the matching degree, and the target information is output.

5. The method according to claim 4, characterized in that, Using a first attention network, the user's steady-state demand potential information and dynamic interest intention information are weighted and fused to generate a joint embedding vector representing the user's current information demand, including: The user's steady-state demand potential information and dynamic interest intention information are input into the first attention network, and the following operations are performed in the first attention network: A gating signal is generated based on the dynamic interest intent information, and the gating signal is used to modulate the steady-state demand potential information element by element to obtain the activated steady-state demand potential vector. Cross-attention calculation is performed on the activated steady-state demand potential vector and the dynamic interest intent information to obtain the context semantic vector; The activated steady-state demand potential vector, the context semantic vector, and the global aggregated representation of the dynamic interest intent information are concatenated and then input into multiple parallel fusion sub-networks to obtain multiple candidate fusion vectors. Based on the steady-state demand potential information and the dynamic interest intent information, a routing weight is generated, and the multiple candidate fusion vectors are weighted and summed according to the routing weight to obtain the joint embedding vector.

6. The method according to claim 4, characterized in that, Using the second attention network in the information recommendation model, attention is calculated on the joint embedding vector and the information vectors corresponding to the candidate information objects to obtain the matching degree between the joint embedding vector and each candidate information object, including: The joint embedding vector is input into the second attention network, and the following operations are performed in the second attention network: Semantic parsing is performed on the information vector of each candidate information object to obtain the sub-vector representation of each candidate information object at at least two different semantic granularity levels; Based on the joint embedding vector and the sub-vector representation of each candidate information object at each semantic granularity level, calculate the absolute matching degree between each candidate information object and the joint embedding vector; Based on the information vector of each candidate information object, construct a competitive context representation of each candidate information object relative to other candidate information objects; The matching degree of the joint embedding vector and the competitive context representation of each candidate information object is calculated to obtain the relative matching degree of each candidate information object with the joint embedding vector under the consideration of the competitive relationship; the competitive context representation is used to obtain the differentiated representation after comparing and modeling the current candidate information object based on the features of other candidate information objects, and is used to characterize its relative competitive position; The matching degree between each candidate information object and the joint embedding vector is generated based on the absolute and relative matching degree between each candidate information object and the joint embedding vector.

7. The method according to claim 6, characterized in that, Based on the joint embedding vector and the sub-vector representations of each candidate information object at various semantic granularity levels, the absolute matching degree between each candidate information object and the joint embedding vector is calculated, including: For each candidate information object, the matching degree of the joint embedding vector with the sub-vector representation of the candidate information object at each semantic granularity level is calculated to obtain the local matching degree of the candidate information object with the joint embedding vector at each semantic granularity level. Based on the joint embedding vector and / or the context information of the candidate information object, the fusion weights of the candidate information object at each semantic granularity level are dynamically generated through a gating network. The local matching degree between the candidate information object and the joint embedding vector at each semantic granularity level is weighted and fused according to the fusion weight to obtain the absolute matching degree between the candidate information object and the joint embedding vector.

8. The method according to any one of claims 1-3 or 5-6, characterized in that, Based on geographic semantic features, distance features, and geographic context features of geographic object description information, data augmentation is performed on geographic object description information in the ontology library using geographic object description information from third-party libraries, including: For any geographic object description information in the third-party library, similar geographic object description information is retrieved in the ontology library to obtain at least two geographic object description information, wherein the geographic object description information includes at least the name information, address information and spatial location information of the geographic object; Using a geographic semantic feature encoder, the geographic semantic similarity between the at least two geographic object descriptions is calculated based on the name and address information in the descriptions of the at least two geographic objects. Using a distance feature encoder, spatial distance features between the at least two geographic object descriptions are calculated based on spatial location information in the descriptions of the at least two geographic objects. Using a geographic context feature encoder, the geographic context similarity between the description information of the at least two geographic objects is calculated based on the description information of the at least two geographic objects and the geographic environment information of their respective surroundings. The geographic semantic similarity, spatial distance features, and geographic context similarity are input into the classification head network, and the classification result is output to determine whether the description information of the at least two geographic objects points to the same geographic object. Based on the classification results, the geographic object description information in the ontology library is augmented using the geographic object description information belonging to a third-party library from the at least two geographic object description information to obtain the target library.

9. An electronic device, characterized in that, include: Processor and memory; The memory stores a computer program, and after the processor runs the computer program, it performs the method described in any one of claims 1-8.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, performs the method described in any one of claims 1-8.

11. A computer program product, characterized in that, include: Computer instructions, when executed by a processor, perform the method described in any one of claims 1-8.

Images