A house source task intelligent allocation system and method and a storage medium

Abstract

The application discloses a kind of house source task intelligent distribution system, the present application is extracted to the feature vector of house source by the deep feature extraction of multiple information of target house source;Meanwhile, the personnel feature vector is generated by combining the basic information, historical sales information and real-time task load of real estate agent;Then, the spatial distance feature of house source and agent is introduced again;Then, the task matching feature vector of target house source and each real estate agent is generated based on the foregoing features;Finally, the matching score is obtained by inputting the task matching feature vector into the task matching model, that is, the agent who is best at handling the house source can be accurately identified according to the matching score, and finally the task is assigned;Therefore, the present application realizes the accurate quantification of house source and agent adaptation, significantly improves the distribution scientificity, and effectively avoids the problems of large workload, low efficiency, poor real-time performance and inconsistent distribution in manual distribution, thereby improving the overall operation efficiency.

Description

Technical Field

[0001] This invention belongs to the field of housing management technology, specifically relating to a housing task intelligent allocation system, method and storage medium. Background Technology

[0002] In the real estate brokerage service sector, the effective management and allocation of property listings is a key factor in improving transaction efficiency. In the traditional model, listings are typically allocated manually by store managers or operations staff based on their personal experience, and then assigned to specific agents for follow-up. However, this allocation method has the following drawbacks: On the one hand, manual allocation often relies on subjective judgment and lacks quantitative analysis of the property's own characteristics (such as unit type, price, location, decoration condition, owner's intention, etc.). It also fails to systematically consider the agent's historical performance, areas of expertise, service area and other capabilities, resulting in highly random and low-accuracy matching results. On the other hand, with the gradual participation of community property management in property listing and the popularization of district-based and community-based agent operation models, the number of properties and the scale of agents are growing in tandem. The workload of manual allocation has increased significantly, which not only reduces allocation efficiency, but also makes it difficult to guarantee real-time and consistency. It is easy for high-quality properties to be occupied by agents with low matching degree, thereby causing problems such as extended transaction cycle and backlog of listed properties.

[0003] Therefore, given the aforementioned shortcomings, how to provide an intelligent housing task allocation system that can solve the problems of strong subjectivity and low efficiency of traditional technologies has become an urgent issue to be addressed. Summary of the Invention

[0004] The purpose of this invention is to provide a housing task intelligent allocation system, method and storage medium to solve the problems of low accuracy of matching results due to strong subjectivity in allocation and low efficiency of manual allocation in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a smart housing task allocation system is provided, including: The information entry terminal is used to obtain the property information of the target property and the personnel information of several real estate agents. The property information includes the basic attribute information of the target property, the property popularity information, the house image and the property description text, and the personnel information includes the basic information of the personnel, the personnel's historical property sales information and task load information. A housing allocation server is used to extract features from basic attribute information and housing popularity information in housing listings, obtaining basic attribute features and housing popularity time-series features, respectively. It also performs fusion feature extraction on house images and housing description text in the housing listing information to obtain text-image fusion features. Based on the basic attribute features, the text-image fusion features, and the housing popularity time-series features, a housing feature vector is generated. Based on the basic information of the personnel, the personnel's historical housing sales information and task load information in the personnel information, a personnel feature vector corresponding to each real estate agent is generated. The property allocation server is used to establish spatial distance feature vectors between each real estate agent and the target property, and to generate task matching feature vectors between each real estate agent and the target property based on the property feature vectors, the feature vectors of each person, and the spatial distance feature vectors. The property allocation server is also used to input the task matching feature vectors of each task into a pre-trained task matching model to obtain the task matching score between the target property and each real estate agent, and to determine the optimal agent for the target property based on the task matching score, so as to send the property listing task of the target property to the optimal agent.

[0006] Based on the aforementioned disclosures, this invention performs deep feature extraction on various information of the target property (basic attributes, popularity time series, house images, and descriptive text), and then fuses the extracted features to generate a property feature vector. Simultaneously, it combines the real estate agent's basic information, historical sales information, and real-time task load to generate a personnel feature vector. Next, it introduces the spatial distance feature between the target property and the real estate agent. Then, based on the property feature vector, personnel feature vector, and spatial distance feature, it generates task matching feature vectors between the target property and each real estate agent. Finally, it inputs the task matching feature vectors into a pre-trained task matching model to obtain a matching score. Based on the matching score, it accurately identifies the agent most skilled in handling the property and ultimately assigns the task. Therefore, this invention achieves precise quantitative matching of properties and agents, significantly improving the scientific nature of the allocation and effectively avoiding the problems of large workload, low efficiency, poor real-time performance, and inconsistent allocation associated with manual allocation, thereby improving overall operational efficiency. Therefore, this invention is highly suitable for large-scale application and promotion.

[0007] In one possible design, the housing allocation server includes a feature extraction unit, wherein, when performing fusion feature extraction on the house image and housing description text in the housing information, the feature extraction unit is configured as follows: A pre-trained BERT model is used to perform semantic mapping on the property description text to obtain a text semantic vector; The house image is divided into several image blocks, and the several image blocks are input into a multi-layer image feature extraction network to generate a first image feature, a second image feature, a third image feature, and a fourth image feature, respectively. The feature dimensions of the four image features are different from each other. The fourth image feature is input into a multilayer perceptron network to obtain a global image feature vector; The text semantic vector is transformed to obtain four transformed text vectors. The feature dimensions of the four transformed text vectors are the same as the feature dimensions of the four image features, and they correspond one-to-one. The four transformed text vectors are concatenated with their corresponding image features to obtain four initial concatenated vectors. These four initial concatenated vectors are then input into the first feature mapping network for feature mapping processing to obtain four mapped concatenated vectors. Perform feature concatenation on the four mapping concatenation vectors to obtain the initial text image fusion feature vector; The initial text image fusion feature vector is input into the second feature mapping network and the third feature mapping network respectively to obtain the first text image mapping feature vector and the second text image mapping feature vector; The first text image mapping feature vector is multiplied by the global image feature vector, and the result of the dot product is weighted and summed with the second text image mapping feature vector to obtain the text image fusion feature.

[0008] In one possible design, the multi-layer image feature extraction network includes: a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a fourth feature extraction layer. The first feature extraction layer includes: a linear embedding layer and a first feature extraction block. The second, third, and fourth feature extraction layers each include a downsampling layer and a second feature extraction block. The number of second feature extraction blocks in the second, third, and fourth feature extraction layers are 1, 3, and 1, respectively. The linear embedding layer is used to perform data mapping processing on each image patch to obtain several mapping vectors, and to generate an initial feature map based on these mapping vectors. The first feature extraction block is used to extract local and global features from the initial feature map to obtain the first image features; The downsampling layer in the second feature extraction layer is used to downsample the first image features to obtain the first downsampled features, and input the first downsampled features into the corresponding second feature extraction block to output the second image features; The downsampling layer in the third feature extraction layer is used to downsample the second image features to obtain the second downsampled features, and then input the second downsampled features into the corresponding three second feature extraction blocks in sequence to output the third image features; The downsampling layer in the fourth feature extraction layer is used to downsample the third image features to obtain the third downsampled features, and then input the third downsampled features into the corresponding second feature extraction block to obtain the fourth image features.

[0009] In one possible design, the network structures of the first feature extraction block and the second feature extraction block are the same, and the first feature extraction block includes: a first LN layer, a windowed self-attention layer, a first residual layer, a second LN layer, a first multilayer perceptron layer, a second residual layer, a third LN layer, a moving window self-attention layer, a third residual layer, a fourth LN layer, a second multilayer perceptron layer, and a fourth residual layer. The first LN layer is used to perform layer normalization on the initial feature map to obtain the first normalized feature. A windowed self-attention layer is used to extract local features from the first normalized features to obtain local image features, and input the local image features into the first residual layer so that the first residual layer performs residual connection between the local image features and the initial feature map to obtain residual fusion features; The second LN layer is used to perform layer normalization on the residual fusion features to obtain normalized fusion features; The first multilayer perceptron is used to perform nonlinear transformation processing on the normalized fused features to obtain the first nonlinear transformed features, and output the first nonlinear transformed features to the second residual layer so that the second residual layer can perform residual connection between the first nonlinear transformed features and the residual fused features to obtain local output features. The third LN layer is used to perform layer normalization on the local output features to obtain the second normalized features. The second normalized features are then input into the moving window self-attention layer so that the moving window self-attention layer can extract cross-window features from the second normalized features to obtain cross-window associated image features. The third residual layer is used to perform residual connection between the cross-window associated image features and the local output features to obtain global fusion features, and input the global fusion features into the fourth LN layer so that the fourth LN layer can perform layer normalization processing on the global fusion features to obtain global normalized features. The second multilayer perceptron layer is used to perform nonlinear transformation processing on the global normalized features to obtain the second nonlinear transformation features, and input the second nonlinear transformation features into the fourth residual layer so that the fourth residual layer performs residual connection between the second nonlinear transformation features and the global fusion features to obtain the first image features.

[0010] In one possible design, the housing allocation server includes: a feature extraction unit, wherein the feature extraction unit is used to calculate the cosine similarity between each person feature vector and the housing feature vector, and to multiply each person feature vector by the housing feature vector element by element to obtain the initial task feature vector corresponding to each real estate agent. The feature extraction unit is also used to perform feature concatenation on the cosine similarity between each person feature vector and the property feature vector, the initial task feature vector corresponding to each real estate agent, and each spatial distance feature vector, so as to obtain the task matching feature vector between each real estate agent and the target property after feature concatenation.

[0011] In one possible design, the housing allocation server includes: a model training unit, wherein the model training unit is used to acquire several historical housing allocation records, wherein any historical housing allocation record includes a sample task matching feature vector between the allocated housing and the allocated real estate agent, and the label data corresponding to the any historical housing allocation record. The model training unit is used to perform parameter optimization on the LightGBM model using a parameter optimization algorithm to obtain the optimal model parameters, and then use the optimal model parameters to update the LightGBM model to obtain the optimal LightGBM model. The model training unit is also used to train the optimal LightGBM model by taking each training data as input and the task matching score between the assigned housing and the assigned real estate agent corresponding to each training data as output, so as to obtain the task matching model after training.

[0012] In one possible design, when using a parameter optimization algorithm to optimize the parameters of the LightGBM model, the model training unit is configured as follows: Obtain the individual population at the t-th iteration, where when t is 1, the individual population at the t-th iteration is the initial population, and each initial individual in the initial population corresponds to a set of model parameters of the LightGBM model. Based on the model parameters corresponding to each individual in the individual population, a LightGBM model corresponding to each individual is generated, and a certain amount of data in the training data is used to train the LightGBM model corresponding to each individual. Based on the output of each LightGBM model, the loss function of each LightGBM model is calculated, so as to calculate the fitness of each individual based on the loss function of each LightGBM model. Based on the fitness of each individual, the globally optimal individual at the t-th iteration is determined; Determine if the iteration stopping condition is met; If not, calculate the probability of each individual choosing the position update strategy at the t-th iteration; Based on the probability of each individual's location update strategy selection, the location update strategy of each individual is determined. The location update strategy of any individual can be a random search strategy, a local encirclement update strategy, a local development strategy, or a neighborhood mutation strategy. Based on the position update strategy for each individual, the position of each individual is updated to obtain the updated individuals, and the updated individuals are used to form the individual population at the (t+1)th iteration. Increment t by 1 and reacquire the individual population at the t-th iteration until the iteration stopping condition is met. Then, take the model parameters corresponding to the globally optimal individual at the iteration stopping condition as the optimal model parameters.

[0013] In one possible design, the model training unit is used to calculate the position update coefficients at the t-th iteration; For any individual, the model training unit is used to determine whether the probability of the individual's location update strategy selection is less than a probability threshold, and when it is determined that the probability of the individual's location update strategy selection is less than the probability threshold, it determines whether the absolute value of the location update coefficient is greater than or equal to a preset threshold. The model training unit is also used to determine that the position update strategy of any individual is a random search strategy when the absolute value of the position update coefficient is greater than or equal to a preset threshold, or to determine that the position update strategy of any individual is a local enclosing update strategy when the absolute value of the position update coefficient is less than a preset threshold. Specifically, when the model training unit determines that the probability of selecting a location update strategy for any individual is greater than or equal to the probability threshold, it generates a selection random number. When the selection random number is less than the probability threshold, it determines that the location update strategy for any individual is a neighborhood mutation strategy, or when the selection random number is greater than or equal to the probability threshold, it determines that the location update strategy for any individual is a local exploitation strategy.

[0014] Secondly, a method for intelligent allocation of housing tasks is provided, including: Obtain the property information of the target property and the personnel information of several real estate agents. The property information includes the basic attribute information of the target property, the property popularity information, the house image and the property description text, and the personnel information includes the basic information of the personnel, the personnel's historical property sales information and task load information. Feature extraction is performed on the basic attribute information and housing popularity information in the housing information to obtain basic attribute features and housing popularity time series features, respectively. Then, fusion feature extraction is performed on the house image and housing description text in the housing information to obtain text image fusion features, so as to generate housing feature vector based on the basic attribute features, the text image fusion features and the housing popularity time series features. Based on the basic information of the personnel, the personnel's historical housing sales information and task load information in the personnel information, a personnel feature vector corresponding to each real estate agent is generated. Establish spatial distance feature vectors between each real estate agent and the target property, and generate task matching feature vectors between each real estate agent and the target property based on the property feature vectors, the feature vectors of each agent, and the spatial distance feature vectors. Each task matching feature vector is input into a pre-trained task matching model to obtain the task matching score between the target property and each real estate agent. Based on each task matching score, the optimal agent for the target property is determined so that the property listing task for the target property can be sent to the optimal agent.

[0015] Thirdly, a housing task intelligent allocation device is provided. Taking the system as an electronic device as an example, it includes a memory, a processor, and a transceiver that are connected in sequence. The memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer programs and execute the housing task intelligent allocation method as described in the second aspect.

[0016] Fourthly, a storage medium is provided, on which instructions are stored, which, when executed on a computer, perform the intelligent allocation method for housing resources as described in the second aspect.

[0017] Fifthly, a computer program product containing instructions is provided, which, when executed on a computer, causes the computer to perform the intelligent allocation method for housing resources as described in the second aspect.

[0018] Beneficial effects: (1) This invention extracts deep features from various information of the target property (basic attributes, popularity time series, house images and descriptive text), and then fuses the extracted features to generate a property feature vector; at the same time, it combines the basic information of real estate agents, historical sales information and real-time task load to generate personnel feature vectors; then, it introduces the spatial distance features between the target property and the real estate agent; next, it generates task matching feature vectors between the target property and each real estate agent based on the property feature vector, personnel feature vector and spatial distance features; finally, it inputs the task matching feature vectors into a pre-trained task matching model to obtain a matching score, thereby accurately identifying the agent who is best at handling the property based on the matching score, and finally assigning tasks; thus, this invention achieves accurate quantitative matching of property and agent, significantly improves the scientific nature of the allocation, and effectively avoids the problems of large workload, low efficiency, poor real-time performance and inconsistent allocation in manual allocation, thereby improving the overall operational efficiency; therefore, this invention is very suitable for large-scale application and promotion. Attached Figure Description

[0019] Figure 1 This is a structural diagram of the intelligent housing task allocation system provided in an embodiment of the present invention; Figure 2 A flowchart for fusing features from house images and property description text provided in this embodiment of the invention; Figure 3 This is a structural diagram of a multilayer image feature extraction network provided in an embodiment of the present invention; Figure 4 This is a structural diagram of the first feature extraction block provided in an embodiment of the present invention; Figure 5 A structural diagram of the first feature mapping network provided for an embodiment of the invention; Figure 6 A flowchart illustrating the steps of the intelligent allocation method for housing resources provided in an embodiment of the present invention. Detailed Implementation

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0021] It should be understood that although the terms first, second, etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit, without departing from the scope of the exemplary embodiments of the invention.

[0022] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0023] Example: See Figure 1 As shown, the intelligent property task allocation system provided in this embodiment extracts multi-dimensional features of properties by having property management and / or real estate agents input property information. Simultaneously, based on the personnel information of real estate agents, it constructs a database of agents specific to a particular area and community. Based on the extracted multi-dimensional features of the properties and the personnel features of the agents in the agent database, and combined with a preset allocation model, it performs task matching calculations on the property features and agent features to achieve optimal matching between the properties and agents in the same area or community. The system automatically allocates property tasks to the agents who are most skilled and have the highest probability of closing deals, thereby improving the property transaction rate and shortening the listing cycle.

[0024] This invention is adapted to a business model where property management or agents input property listings and agents operate within designated areas. Specifically, this model is a new type of real estate brokerage collaboration model that uses the property management or agent side as the entry point for property listings, geographical areas as the operating units, and data intelligence as the allocation engine. Property management personnel or agents fill in basic property information (floor type, area, floor, orientation, etc.) via a mini-program or PC, upload real-life photos, and optionally fill in the owner's expected price, selling / renting intention, etc., thereby ensuring the authenticity of the property information (the property management side has the owner's identity and property ownership information, which can be cross-verified to prevent fake listings from the source). Then, several areas are divided based on urban administrative streets, residential communities, or natural commercial districts. Each area has an area manager responsible for the management of agents and resource coordination within the area. For large residential communities, 1-3 "resident agents" can be set up to specifically handle the viewing, negotiation, and transaction of all properties in that community. Smaller communities are managed by area agents. The system achieves comprehensive coverage by recording each agent's historical transaction records, preferred apartment types, average transaction cycle, customer ratings, response speed, and other dynamic characteristics to create capability tags. Once information is entered, the system extracts property features (location, apartment type, price, decoration, etc.) and generates agent profiles. Based on a pre-defined matching model, it calculates the suitability of each property with an agent and automatically pushes the property to the agent with the highest score. After completing viewings, negotiations, and signing, the agent sends the transaction results (transaction price, transaction cycle, customer feedback) back to the system, creating a data loop. Thus, this system, with the property management company as the trust hub, localized agents as the service providers, and an intelligent allocation algorithm as the core engine, constructs a new real estate service ecosystem characterized by "authentic properties, professional agents, efficient allocation, and shared benefits." This effectively solves the problems of traditional manual allocation, such as high subjectivity, low matching accuracy, and low efficiency, achieving intelligent and precise allocation of property listings.

[0025] Optionally, the aforementioned system may include, but is not limited to, an information entry terminal and a property allocation server. The information entry terminal is used to enter relevant information about properties and real estate agents, while the property allocation server uses the information entered by the information entry terminal to perform quantitative matching of properties and real estate agents, thereby matching the best real estate agent for each property.

[0026] The entire system's working process is as follows: An information entry terminal is used to acquire property information of a target property and personnel information of several real estate agents. In this embodiment, the information entry terminal can be divided into a property management end and an agent end, meaning that both property management and agents can enter property information. This allows for flexible entry of property information by property management and / or agents according to different usage scenarios. Of course, personnel information can be entered by the agent end. Furthermore, the property information includes basic attribute information of the target property, property popularity information, house images, and property description text, and the personnel information includes basic personnel information, personnel's historical property sales information, and task load information.

[0027] In this embodiment, the basic attribute information may include, but is not limited to, the following: area, unit type (number of rooms / living rooms, etc.), floor, orientation, decoration level, and property management fee; similarly, the property popularity information may include the number of listings, viewings, and price adjustments in the same community in recent days, as well as the number of price adjustments for the target property; furthermore, the property images may include, but are not limited to, real-life images of the target property, while the property description text contains detailed descriptions of the target property, such as an introduction to the property, an introduction to the community and surrounding environment, transaction conditions, etc.

[0028] Furthermore, basic personnel information may include, but is not limited to, years of service, total historical transaction volume, and average transaction cycle. Historical property sales information may include: transaction rate in recent days (e.g., in the last 30 days), response time (from order placement to first contact), customer rating, and percentage of transactions in the same area. Finally, task load information may include, but is not limited to, the number of tasks currently held and the number of properties to be viewed (i.e., how many customers have already made viewing appointments).

[0029] Thus, by inputting property information and real estate agent information through the information input terminal, an agent database can be established based on the agent information. This facilitates the subsequent allocation of the optimal real estate agent to each target property based on the agent database. After the information input is completed, multi-dimensional features of the property and agents can be extracted on the property allocation server to generate property feature vectors and agent feature vectors. These vectors can then be used to match properties with real estate agents using a pre-trained task matching model.

[0030] The multi-dimensional feature extraction and task matching process is as follows: The housing allocation server is used to extract features from the basic attribute information and housing popularity information in the housing information, to obtain basic attribute features and housing popularity time series features, respectively. It also performs fusion feature extraction on the house images and housing description text in the housing information to obtain text image fusion features, so as to generate housing feature vectors based on the basic attribute features, the text image fusion features and the housing popularity time series features.

[0031] In specific implementation, a housing allocation server may include, but is not limited to, a feature extraction unit and a task allocation unit, and the aforementioned feature extraction work is completed by the feature extraction unit.

[0032] In practical implementation, the feature extraction unit can directly extract numerical information from the basic attribute information as feature values, such as area, number of rooms and living rooms in the apartment type, floor, property management fees, and other information with clear numerical values, which can be directly used as feature values ​​in the basic attribute features. Among them, orientation can be converted into feature values ​​using one-hot encoding, while decoration level can be represented by numbers from 0 to 3. In this way, by combining the aforementioned features into a vector, the basic attribute features can be obtained.

[0033] Furthermore, for housing popularity information, the feature extraction unit can encode it and input it into the LSTM network. Then, the hidden state of the last step of the network is used as the temporal feature of housing popularity.

[0034] Thus, after extracting the features of basic attribute information and housing popularity information, the fusion feature extraction of housing images and housing description text can be carried out. Since text and images belong to different modalities, multimodal fusion is required to overcome the information limitations of a single modality, thereby extracting housing features more accurately and facilitating more accurate task matching in the future.

[0035] In this embodiment, see Figure 2 As shown, when performing fusion feature extraction on house images and house description text in property listings, the feature extraction unit is configured to perform the following operations: Step 1: Using a pre-trained BERT model, semantic mapping is performed on the property description text to obtain a text semantic vector. In this embodiment, the property description text can be input into the BERT-base-uncased model, and then the [CLS] vector is taken and reduced to 128 dimensions through a fully connected layer to obtain the text semantic vector. The [CLS] symbol is inserted at the beginning of the text to represent the semantic information of the entire text. During model training, this symbol learns the semantic information of the entire text and generates a corresponding output vector. This vector can be used as the semantic representation of the entire text. Therefore, the CLS vector represents the semantic information of the entire input sequence.

[0036] After completing the semantic mapping of the property description text, image features can be extracted, as shown in the following process.

[0037] Step 2: Divide the house image into several image blocks and input the image blocks into a multi-layer image feature extraction network to generate a first image feature, a second image feature, a third image feature, and a fourth image feature, respectively. The feature dimensions of the four image features are different from each other. In this embodiment, the size of any image block can be set to 4×4. Of course, the size of the image block can be set according to actual use. This embodiment is not limited to the above example.

[0038] Meanwhile, this embodiment constructs a multi-layer image feature extraction network to extract image features at different scales, so that they can be concatenated with text semantic vectors at different scales to achieve multi-scale feature fusion.

[0039] Optionally, one of the network structures of a multi-layer image feature extraction network is disclosed below: See Figure 3 As shown, a multi-layer image feature extraction network may include, but is not limited to, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a fourth feature extraction layer. The first feature extraction layer includes a linear embedding layer and a first feature extraction block, and the second, third, and fourth feature extraction layers each include a downsampling layer and a second feature extraction block. In this embodiment, the network structures of the first and second feature extraction blocks are the same, and the number of second feature extraction blocks in the second, third, and fourth feature extraction layers is 1, 3, and 1, respectively.

[0040] The working process of each of the aforementioned feature extraction layers is as follows: First, the linear embedding layer performs data mapping on each image patch to obtain several mapping vectors. Based on these mapping vectors, an initial feature map is generated. That is, the linear embedding layer performs a linear transformation on each image patch, converting each patch into a mapping vector. Then, these mapping vectors are arranged into a two-dimensional grid according to their positions (rows and columns) in the image, thereby generating the initial feature map, which is then output to the first feature extraction block for feature extraction. The first feature extraction block is used to extract local and global features from the initial feature map to obtain the first image features. In specific implementation, the first feature extraction block adopts an attention mechanism to perform local and global feature extraction.

[0041] Optionally, the following uses the first feature extraction block as an example to illustrate its network structure; see [link to relevant documentation]. Figure 4As shown, the first feature extraction block may include, but is not limited to, a first LN layer, a windowed self-attention layer, a first residual layer, a second LN layer, a first multilayer perceptron layer, a second residual layer, a third LN layer, a moving window self-attention layer, a third residual layer, a fourth LN layer, a second multilayer perceptron layer, and a fourth residual layer.

[0042] Thus, after the initial feature map enters the first feature extraction block, it can first undergo layer normalization processing in the first LN layer, i.e., the first LN layer, which is used to perform layer normalization processing on the initial feature map to obtain the first normalized feature; then, the windowed self-attention layer is used to extract local features from the first normalized feature to obtain local image features.

[0043] It should be noted that the windowed self-attention layer divides the first normalized features into non-overlapping local windows, calculates self-attention independently within each window, calculates the attention weights between each patch within the window (capturing the local dependencies between patches within the window), and aggregates them by weight to obtain the output, thereby realizing the aggregation of local features and obtaining local image features.

[0044] After obtaining the local image features, they are input into the first residual layer. The first residual layer performs a residual concatenation between the local image features and the initial feature map to obtain residual fusion features. Next, these features are input into the second LN layer for layer normalization processing. Specifically, the second LN layer performs layer normalization on the residual fusion features to obtain normalized fusion features. Then, the first multilayer perceptron performs a nonlinear transformation on the normalized fusion features to obtain the first nonlinear transformation features. Finally, the first nonlinear transformation features are output to the second residual layer, which then performs a residual concatenation between the first nonlinear transformation features and the residual fusion features to obtain local output features. In this way, the first multilayer perceptron can be used to perform nonlinear transformations and channel-dimensional information interaction on the normalized fusion features, enhancing the model's expressive power.

[0045] After obtaining the local output features, cross-window features can be extracted and combined with the local features to form global features. The process is as follows: The third LN layer is used to perform layer normalization on the local output features to obtain the second normalized features. The second normalized features are then input into the moving window self-attention layer so that the moving window self-attention layer can extract cross-window features from the second normalized features to obtain cross-window associated image features.

[0046] In practical implementation, the moving window self-attention layer introduces window displacement (i.e., sliding the window over the input data) between consecutive layers, causing the originally non-overlapping windows to shift, thereby allowing information exchange across windows and expanding the model's receptive field. In this way, the moving window self-attention layer can achieve cross-window connections, allowing the model to capture global dependencies in the image. Of course, windowed self-attention layers and moving window self-attention layers are commonly used attention-based networks, and their principles will not be elaborated here.

[0047] After obtaining the cross-window associated image features, a third residual layer can be used to perform residual connections between the input and output. This third residual layer connects the cross-window associated image features and the local output features to obtain global fusion features. These global fusion features are then input to a fourth LN layer, which performs layer normalization on them to obtain global normalized features. Next, a second multilayer perceptron is used to perform a nonlinear transformation on the global features. This second multilayer perceptron layer performs a nonlinear transformation on the global normalized features to obtain second nonlinear transformed features (again, nonlinear transformation and channel-dimensional information interaction are performed on the global normalized features to further enhance the model's expressive power). Finally, the second nonlinear transformed features are input to the fourth residual layer, which performs a residual connection between the second nonlinear transformed features and the global fusion features to obtain the first image features.

[0048] Thus, the first feature extraction block uses windowed self-attention layers and moving window self-attention layers to effectively extract local and global features of the image and reduce computational complexity. That is, the alternation of the two can extract fine-grained features with low computational cost of local windows, and can achieve hierarchical global modeling through window movement, thereby achieving a balance between efficiency and expressive power.

[0049] After completing the first level of image feature extraction, the image can be input into the second feature extraction layer for further feature extraction. The specific working process is as follows: The downsampling layer in the second feature extraction layer is used to downsample the first image features to obtain the first downsampled features. The first downsampled features are then input into the corresponding second feature extraction block to output the second image features. In specific applications, the downsampling operation in the second feature extraction layer is as follows: adjacent 2×2 blocks (a total of 4) in the first image features are merged into one larger block, that is, the 4 feature vectors in each 2×2 region are concatenated along the channel dimension to complete the downsampling of the first image features. Then, the first downsampled features are input into the second feature extraction block to continue local and global feature extraction to obtain the second image features. The working process of the second feature extraction block can be referred to that of the first feature extraction block, and will not be repeated here in this embodiment.

[0050] Similarly, after obtaining the second image features, they can be input into the third feature extraction layer for feature extraction at the next scale. The process is as follows: the downsampling layer in the third feature extraction layer is used to downsample the second image features to obtain the second downsampled features, and the second downsampled features are sequentially input into the corresponding three second feature extraction blocks to output the third image features. In this embodiment, the downsampling operation of the second image features is also to concatenate the four feature vectors in each 2×2 region along the channel dimension, and the working process of any second feature extraction block in this layer is the same as that of the first feature extraction block, which will not be described again here.

[0051] After obtaining the third image feature, it is input to the last feature extraction layer to obtain the fourth image feature. That is, the downsampling layer in the fourth feature extraction layer is used to downsample the third image feature to obtain the third downsampled feature, and the third downsampled feature is input to the corresponding second feature extraction block to obtain the fourth image feature. The working process of this layer can be referred to the second feature extraction layer, which will not be repeated in this embodiment.

[0052] Thus, through the aforementioned four feature extraction layers, multi-scale image feature extraction can be achieved, thereby providing rich information for subsequent fine-grained fusion with text; then, based on the fourth image feature, a global image feature vector can be generated, as shown in the third step below.

[0053] The third step is to input the fourth image feature into a multilayer perceptron network to obtain a global image feature vector. In this embodiment, the multilayer perceptron network is an MLP network, which flattens or globally pools the high-dimensional, multi-channel feature map output by the fourth feature extraction layer and then converts it into a fixed-length vector through a fully connected layer to obtain a global representation of the entire image. This is the key transformation from local feature map to global semantic vector (which plays the role of "preserving the overall information of the original image"), so that the image can be combined with the subsequent initial text image fusion features in the same vector space.

[0054] After generating the global image feature vector, text and image features can be concatenated at different scales, as shown in steps four and five below.

[0055] Step 4: Perform dimensional transformation on the text semantic vector to obtain four transformed text vectors. The feature dimensions of the four transformed text vectors are the same as the feature dimensions of the four image features and correspond one-to-one. In this embodiment, a tensor broadcasting mechanism can be used, but is not limited to, to perform dimensional transformation on the text semantic vector, thereby generating transformed text vectors that correspond one-to-one with the aforementioned four image feature dimensions. Then, text and image features with the same dimensions can be concatenated, and the process can be, but is not limited to, as shown in Step 5 below.

[0056] Step 5: Concatenate the four transformed text vectors with their corresponding image features to obtain four initial concatenated vectors (connecting image features and text vectors along the channel dimension, allowing the network to see both "raw visual information" and "text description"). Then, input these four initial concatenated vectors into the first feature mapping network for feature mapping processing to obtain four mapped concatenated vectors. For specific applications, see [link to relevant documentation]. Figure 5 As shown, the first feature mapping network may include, but is not limited to, a first BN layer, a first ReLU layer, a first fully connected layer, a second ReLU layer, and a second fully connected layer connected in sequence. Thus, each initial concatenated vector first enters the first BN layer for normalization, then enters the first ReLU layer and is activated using the corresponding ReLU function, thereby increasing the network's nonlinearity and improving the model's expressive power. Next, the nonlinearly processed features enter the first fully connected layer for linear transformation. Finally, ReLU activation and linear transformation are performed again to obtain the mapped concatenated vector. In this way, through the aforementioned first feature mapping network, the concatenated features can be normalized, nonlinearly activated, and linearly transformed, thereby extracting the joint features of text and image at this scale, thus enhancing the expressive power of the features.

[0057] After completing the mapping of the splicing features, the mapped features at the four scales can be spliced ​​to obtain the initial text image fusion feature vector, as shown in step six below.

[0058] Step 6: Perform feature concatenation on the four mapped concatenated vectors to obtain the initial text-image fusion feature vector. In this embodiment, the feature vectors after mapping at the four scales are concatenated along the channel dimension to form a multi-scale fusion feature vector. This preserves all information from local to global, providing a complete basis for subsequent global combination. After completing the initial text-image fusion, a secondary mapping is required, the process of which is as follows: Step 7: Input the initial text-image fusion feature vector into the second feature mapping network and the third feature mapping network respectively to obtain the first text-image mapping feature vector and the second text-image mapping feature vector. In this embodiment, the second feature mapping network may include, but is not limited to, a second BN layer, a third ReLU layer and a third fully connected layer (the third fully connected layer is followed by a Sigmoid activation function) connected in sequence, while the network structure of the third feature mapping network is the same as that of the first feature mapping network. The output of the second feature mapping network is used to control the degree of preservation of the original image information, while the output of the third feature mapping network is used to control the text-image fusion information.

[0059] After completing the secondary feature mapping, the final feature fusion can be performed, as shown in step 8 below: Step 8: Perform a dot product operation on the first text image mapping feature vector and the global image feature vector, and then perform a weighted summation on the dot product result and the second text image mapping feature vector to obtain the text image fusion feature after weighted summation. In this embodiment, the dot product operation yields the original image features that are retained. The text image fusion feature is obtained by weighted summation on the text image fusion information.

[0060] Thus, as explained above, the feature extraction unit in this embodiment uses a multi-layer image feature extraction network and multiple feature mapping networks to achieve feature fusion between the property description text and the house image, thereby obtaining text-image fusion features.

[0061] Finally, the basic attribute features, text-image fusion features, and property popularity time series features can be combined into a vector, which will then serve as the property feature vector.

[0062] After extracting the multimodal features of the target property's listing information, the feature extraction of the personnel information of each real estate agent can be performed. The process is as follows: The housing allocation server (i.e., its internal feature extraction unit) is used to generate a personnel feature vector corresponding to each real estate agent based on the basic personnel information, historical housing sales information, and task load information in the personnel information. In this embodiment, the feature extraction process of personnel information can refer to the aforementioned basic attribute information, that is, for textual data, encoding (one-hot encoding) is performed, while for numerical data, it is used directly. Then, the various features are combined into a vector to obtain the personnel feature vector.

[0063] After extracting the property feature vector and personnel feature vector, this embodiment also introduces the spatial distance feature between personnel and property, namely: the property allocation server, which is used to establish the spatial distance feature vector between each real estate agent and the target property, and generate the task matching feature vector between each real estate agent and the target property based on the property feature vector, the personnel feature vector, and the spatial distance feature vector.

[0064] In practical applications, the spatial distance feature vector is represented as: ,in, This represents the spatial distance characteristic value between real estate agent A and target property B. Specifically, when real estate agent A and target property B are located in the same neighborhood... The value is 1, when real estate agent A and target property B are located in different neighborhoods within the same area. The value is 0.6 when real estate agent A and target property B are in different areas. The value is 0.2; optionally, the area can be set as a street.

[0065] Thus, after determining the spatial distance feature vectors between each real estate agent and the target property, the property feature vectors and agent feature vectors can be combined to generate the task matching vectors between each real estate agent and the target property. The process is as follows: The feature extraction unit is used to calculate the cosine similarity between each person's feature vector and the property's feature vector, and to multiply each person's feature vector element-wise with the property's feature vector to obtain the initial task feature vector corresponding to each real estate agent; then, it is used to concatenate the cosine similarity between each person's feature vector and the property's feature vector, the initial task feature vector corresponding to each real estate agent, and each spatial distance feature vector to obtain the task matching feature vector between each real estate agent and the target property after feature concatenation.

[0066] Optionally, the task matching vector between real estate agent a and the target property can be represented as: [ ⊙ , , ],in, , These represent the personnel feature vector of the real estate agent and the property feature vector of the target property, respectively.

[0067] After obtaining the task matching feature vectors between each real estate agent and the target property, a pre-trained task matching model can be used to perform task matching. The process is as follows: The task allocation unit in the property allocation server is used to input the matching feature vectors of each task into a pre-trained task matching model to obtain the task matching score between the target property and each real estate agent, and based on the task matching score, determine the optimal agent for the target property so that the property listing task of the target property can be sent to the optimal agent.

[0068] In this embodiment, after any task matching feature vector is input into the task matching model, it will output a score between [0,1]. This score is used as the task matching score between the real estate agent and the target property corresponding to the task matching feature vector. In this way, the matching degree between the target property and different real estate agents can be evaluated based on each task matching score. That is, the higher the task matching score, the higher the matching degree between the target property and the real estate agent. Based on this, the real estate agent with the highest task matching score can be regarded as the optimal agent for the target property, and the property task of the target property can be sent to the terminal (i.e., agent terminal) corresponding to the aforementioned optimal agent.

[0069] It should be noted that when the real estate agent with the highest task matching score reaches the maximum number of tasks, the real estate agent with the second highest task matching score will be selected as the optimal agent, and the allocation of property listings will be based on this.

[0070] In addition, the task matching model mentioned above can be, but is not limited to, the trained LightGBM model; of course, other neural network models can also be used, and this embodiment is not limited to the example mentioned above.

[0071] Therefore, through the detailed description of the intelligent task allocation system for housing listings, this invention extracts deep features from various information of the target housing listings and then fuses these features to generate a housing listing feature vector. Simultaneously, it combines the basic information, historical sales information, and real-time task load of real estate agents to generate a personnel feature vector. Next, it introduces the spatial distance features between the target housing listings and real estate agents. Then, based on the housing listing feature vector, personnel feature vector, and spatial distance features, it generates task matching feature vectors between the target housing listings and each real estate agent. Finally, it inputs the task matching feature vectors into a pre-trained task matching model to obtain a matching score. Based on the matching score, it accurately identifies the agent most skilled in handling the housing listing and ultimately allocates the task. Thus, this invention achieves precise quantitative matching of housing listings and agents, significantly improving the scientific nature of the allocation and effectively avoiding the problems of large workload, low efficiency, poor real-time performance, and inconsistent allocation associated with manual allocation, thereby improving overall operational efficiency. Therefore, this invention is highly suitable for large-scale application and promotion.

[0072] In one possible design, the second aspect of this embodiment provides one of the training methods for the task allocation model in the first aspect of the embodiment.

[0073] Optionally, the example housing allocation server may further include: a model training unit; wherein the model training unit is used to acquire several historical housing allocation records, wherein any historical housing allocation record includes a sample task matching feature vector between the allocated housing and the allocated real estate agent, and label data corresponding to the any historical housing allocation record, and the label data is 1 or 0, wherein a label data of 1 indicates that the allocated housing was sold within a preset period, and a label data of 0 indicates that the allocated housing was not sold within the preset period.

[0074] Next, the model training unit is used to perform parameter optimization on the LightGBM model using a parameter optimization algorithm to obtain the optimal model parameters, and then uses the optimal model parameters to update the LightGBM model to obtain the optimal LightGBM model. Finally, the model training unit is also used to train the optimal LightGBM model with each training data as input and the task matching score between the assigned housing and the assigned real estate agent corresponding to each training data as output, so that the task matching model is obtained after training.

[0075] Optionally, when performing parameter optimization on the LightGBM model using a parameter optimization algorithm, the model training unit is configured to perform the following operations: Step 1: Obtain the individual population at the t-th iteration, where when t is 1, the individual population at the t-th iteration is the initial population, and each initial individual in the initial population corresponds to a set of model parameters of the LightGBM model; in this embodiment, the model parameters may include the learning rate, the number of trees, etc.; thus, by initializing the aforementioned model parameters multiple times, multiple initial individuals can be obtained, and then multiple initial individuals can be used to form the initial population.

[0076] After obtaining the individual population at the t-th iteration, the aforementioned training data can be used to calculate the individual fitness, as shown in step 2 below.

[0077] Step 2: Based on the model parameters corresponding to each individual in the individual population, generate a LightGBM model corresponding to each individual, and train the LightGBM model corresponding to each individual using some data in the training data. Based on the output of each LightGBM model, calculate the loss function of each LightGBM model, so as to calculate the fitness of each individual based on the loss function of each LightGBM model.

[0078] In this embodiment, as explained above, each individual corresponds to a set of model parameters. Therefore, using the model parameters corresponding to each individual, a LightGBM model corresponding to each individual can be constructed. Then, the training data is divided proportionally (e.g., in a 5:5 ratio), and half of the training data is used to train the LightGBM model for each individual, thereby obtaining the output of the model corresponding to each individual. Next, the loss function of the model (e.g., the cross-entropy loss function) is calculated using the labels of the input data and the model output. Finally, the reciprocal of the loss function can be used as the fitness of the individual. Thus, the higher the fitness of an individual, the better the model performance, that is, the better its corresponding model parameters.

[0079] After calculating the fitness of each individual, the globally optimal individual for the current iteration can be determined, as shown in step 3 below.

[0080] Step 3: Based on the fitness of each individual, determine the globally optimal individual at the t-th iteration.

[0081] In this embodiment, the maximum fitness of each individual at the t-th iteration is first selected. Then, it is determined whether the maximum fitness is greater than the fitness of the globally optimal individual at the (t-1)-th iteration. If it is greater, the individual corresponding to the maximum fitness is taken as the globally optimal individual at the t-th iteration; otherwise, the globally optimal individual at the (t-1)-th iteration is taken as the globally optimal individual at the t-th iteration. Of course, when t is 1, the globally optimal individual at the t-th iteration is the individual with the highest fitness in that iteration. After determining the globally optimal individual, it is determined whether the iteration stopping condition is met, as shown in step 4 below.

[0082] Step 4: Determine if the iteration stopping condition is met; in specific implementation, the iteration stopping condition may be, but is not limited to, reaching the maximum number of iterations at t or the fitness of the globally optimal individual at the t-th iteration being greater than or equal to the fitness threshold; if the aforementioned iteration stopping condition is not met, then population evolution is required.

[0083] This embodiment provides a parameter optimization algorithm with multiple evolutionary strategies. In each iteration, it generates a corresponding strategy selection probability for each individual and then prompts the individual to select different position update strategies for position update based on its own strategy selection probability, thereby ensuring the diversity of the population and simultaneously achieving both local and global exploration.

[0084] The population evolution process is shown in steps 5 to 7 below.

[0085] Step 5. If not, calculate the position update strategy selection probability of each individual at the t-th iteration; in specific applications, for any individual, the model training unit is used to select the maximum fitness from the fitness of each individual at the t-th iteration; then, it is used to calculate the position update strategy selection probability corresponding to any individual based on the maximum fitness and the fitness of any individual.

[0086] Optionally, for example but not limited to, the following formula can be used to calculate the probability of location update strategy selection for any individual.

[0087] ; In the formula, This represents the probability of choosing a position update strategy for any individual. Let represent the maximum fitness among all individuals at the t-th iteration. This represents the fitness of any of the individuals. These represent the maximum and minimum selection probabilities, respectively.

[0088] As can be seen from the above formula, the selection probability provided in this embodiment determines whether the current iteration should perform a global search or local exploration based on the fitness of the individual itself. That is, when the fitness of an individual is close to that of the individual with the current best fitness, its transition probability is close to the minimum value, making it easier to enter the guidance of global exploration; conversely, when the fitness of an individual differs greatly from that of the current best fitness, its transition probability is close to the maximum value, making it easier to enter the guidance of local exploration. Based on this, individuals with lower fitness values ​​are more likely to perform global exploration during the search process, while individuals with higher fitness values ​​tend to perform local exploration, thereby avoiding the problem of getting trapped in local optima.

[0089] After calculating the position update selection strategy for each individual in the current iteration based on the aforementioned formula, the position update strategy for each individual can be determined based on this, as shown in step 6 below.

[0090] Step 6: Determine the position update strategy for each individual based on the probability of the position update strategy selection. The position update strategy for any individual can be a random search strategy, a local encirclement update strategy, a local exploitation strategy, or a neighborhood mutation strategy.

[0091] In specific implementation, the model training unit first calculates the position update coefficient at the t-th iteration; then, for any individual, the model training unit determines whether the probability of selecting the position update strategy for any individual is less than a probability threshold (which may be, but is not limited to, 0.5), and when it is determined that the probability of selecting the position update strategy for any individual is less than the probability threshold, it determines whether the absolute value of the position update coefficient is greater than or equal to a preset threshold (which may be, but is not limited to, 1); wherein, the model training unit is also used to determine that the position update strategy for any individual is a random search strategy when it is determined that the absolute value of the position update coefficient is greater than or equal to the preset threshold, or to determine that the position update strategy for any individual is a local enclosing update strategy when it is determined that the absolute value of the position update coefficient is less than the preset threshold.

[0092] It should be noted that when the model training unit determines that the probability of selecting the position update strategy for any individual is greater than or equal to the probability threshold, it generates a selection random number (the selection random number takes a value between 0 and 1), and when the selection random number is less than the probability threshold, it determines that the position update strategy for any individual is a neighborhood mutation strategy, or when the selection random number is greater than or equal to the probability threshold, it determines that the position update strategy for any individual is a local exploitation strategy.

[0093] Optionally, the formula for calculating the location update coefficient is: ; In the formula, Here, T represents the position update coefficient, and T represents the maximum number of iterations. It is a second random number between [0,1].

[0094] Thus, after determining the location update strategy for each individual based on the aforementioned method, the individual's location can be updated accordingly, as shown in step 7 below.

[0095] Step 7: Based on the position update strategy of each individual, update the position of each individual to obtain the updated individuals, and use the updated individuals to form the individual population at the (t+1)th iteration.

[0096] In specific implementation, when the position update strategy corresponding to the i-th individual is a local enclosing update strategy, the position of the i-th individual can be updated in the following ways, but is not limited to:

[0097] First, the population center of the individual population at the t-th iteration is determined; then, the distance between each individual and the population center at the t-th iteration is calculated; next, based on the distance between each individual and the population center, the local encirclement step size is calculated; finally, the position update coefficient, the local encirclement step size, and the globally optimal individual can be used to update the position of the i-th individual, thereby obtaining the updated i-th individual.

[0098] Optionally, for example, but not limited to, the distance between the i-th individual and the population center can be calculated using the following formula.

[0099] ; In the formula, This represents the distance between the i-th individual and the population center. Let j be the j-th element in the position vector of the i-th individual. This represents the j-th element in the vector corresponding to the population center. Let be the total dimension of the position vector of the i-th individual; in this embodiment, the position vector corresponding to the population center is the mean vector of the position vectors of all individuals at the t-th iteration.

[0100] Thus, based on the aforementioned formula, after calculating the distance between the i-th individual and the population center, the local encirclement step length of the individual can be calculated; for example, but not limited to, the aforementioned local encirclement step length can be calculated according to the following formula.

[0101] ; In the formula, Indicates the local bounding step size. This represents the i-th individual. This represents the distance between the i-th individual and the population center. Indicates population size.

[0102] After calculating the local encirclement step size based on the aforementioned formula, the position of the i-th individual can be updated.

[0103] Optionally, for example, but not limited to, the position of the i-th individual can be updated according to the following formula; ; In the formula, For the updated i-th individual, This represents the i-th individual. This represents the globally optimal individual at the t-th iteration. For position update coefficients, The first random number between [0,1] This indicates the local bounding step size.

[0104] It should be noted that the reason for adopting the aforementioned local encirclement strategy for position updates is that in the later stages of traditional optimization algorithms, individuals will continuously move towards the global optimum, leading to a decrease in population diversity and premature convergence of the algorithm. Therefore, to solve this problem, this embodiment introduces an encirclement step size based on the distance between the individual and the population center. That is, when the distance between individuals is close (i.e., the population density is high), the distance between the individual and the population center is smaller. In this case, the local encirclement step size will increase. In this way, more concentrated individuals can take a larger step size for updates, which can increase diversity and explore unknown areas more likely, thereby increasing the optimization ability of the algorithm.

[0105] Conversely, when the distance between individuals is large (i.e., the population density is low and the individuals are more dispersed), the distance between the individuals and the population center is greater. In this case, the local encirclement step size will become smaller, so that individuals can update with a small step size, thereby ensuring more refined exploration in the local area.

[0106] Therefore, the introduction of local encirclement step size solves the problems of population diversity decline and premature convergence during iteration.

[0107] Furthermore, when the position update strategy for the i-th individual is a random search strategy, the position of the i-th individual is updated using the following formula, which is to perform a global search.

[0108] ; In the formula, For the updated i-th individual, This represents the i-th individual. The first random number between [0,1] Let be an individual randomly selected from the population at the t-th iteration. This is the position update coefficient.

[0109] Furthermore, when the location update strategy corresponding to the i-th individual is a local development strategy, the location of the i-th individual is updated using the following formula.

[0110] ; In the formula, For the updated i-th individual, This represents the i-th individual. This represents the globally optimal individual at the t-th iteration. For local development coefficient, It is a third random number between [-1, 1].

[0111] Based on the above formula, it can be seen that the local development strategy moves around the globally optimal individual along a spiral trajectory, which can perform a more refined search near the optimal solution, thereby improving the convergence accuracy of the algorithm; at the same time, the third random number allows the individual to move at different distances from the optimal solution, which has a wider local coverage than simply shrinking in a straight line towards the optimal solution.

[0112] Finally, when the position update strategy for the i-th individual is the neighborhood mutation strategy, the position of the i-th individual is updated in the following manner.

[0113] First, from the individual population at the t-th iteration, determine the neighboring individuals of the i-th individual; in this embodiment, the (i-1)-th individual can be taken as the neighboring individual of the i-th individual; when i is 1, the (i+1)-th individual is taken as the neighboring individual.

[0114] Then, the neighborhood mutation weight at the t-th iteration is calculated; optionally, the formula for calculating the neighborhood mutation weight is: .

[0115] Finally, after calculating the neighborhood mutation weights, the position of the i-th individual can be updated using the neighborhood individuals and the neighborhood mutation weights to obtain the updated i-th individual.

[0116] Optionally, for example, but not limited to, the following formula can be used to update the position of the i-th individual.

[0117] ; In the formula, For the updated i-th individual, Represents the i-th individual, The fourth random number between [0,1] Represents the neighborhood individuals, This represents the neighborhood variation weight.

[0118] As is known from the above, this strategy considers the influence of the current individual and its neighboring individuals on the location information of the next generation individual. When the selected random number is less than 0.5, this mutation is adopted; otherwise, a local development strategy is adopted to increase the probability of the algorithm escaping local optima.

[0119] Therefore, by using the aforementioned four position update strategies to update each individual at the t-th iteration, the individual population at the (t+1)-th iteration can be obtained; then, the aforementioned population update process can be repeated until the iteration stopping condition is met; the loop process is shown in step 8 below.

[0120] Step 8: Increment t by 1 and reacquire the individual population at the t-th iteration until the iteration stopping condition is met. Then, take the model parameters corresponding to the globally optimal individual at the iteration stopping condition as the optimal model parameters.

[0121] In this embodiment, when the iteration stopping condition is met, the iteration can be stopped. Then, the model parameters corresponding to the globally optimal individual when the iteration stopping condition is met are used as the optimal parameters to update the model and obtain the optimal LightGBM model.

[0122] After obtaining the optimal LightGBM model, all training data can be used to train the model. That is, LightGBM is trained using the full training data plus the optimal hyperparameters (optimal model parameters). After training, the model is saved and deployed to obtain the task matching model. At this point, all data is fully utilized, improving model stability and prediction accuracy.

[0123] Finally, in practical use, you only need to input the task matching feature vectors of each real estate agent and the target property into the model to obtain the task matching score of each real estate agent and the target property. Based on this, the intelligent and scientific allocation of property tasks can be completed.

[0124] like Figure 6 As shown, the third aspect of this embodiment provides a method for intelligent allocation of housing tasks, wherein the method is executed based on the intelligent allocation system for housing tasks described in the first aspect of the embodiment, and the operation steps of the method may be, but are not limited to, the steps S1 to S5 below.

[0125] S1. Obtain the property information of the target property and the personnel information of several real estate agents, wherein the property information includes the basic attribute information of the target property, the property popularity information, the house image and the property description text, and the personnel information includes the basic information of the personnel, the personnel's historical property sales information and task load information.

[0126] S2. Extract features from the basic attribute information and housing popularity information in the housing information to obtain basic attribute features and housing popularity time series features, respectively. Extract fusion features from the house images and housing description text in the housing information to obtain text image fusion features, so as to generate a housing feature vector based on the basic attribute features, the text image fusion features and the housing popularity time series features.

[0127] S3. Based on the basic information of the personnel, the personnel's historical housing sales information and the task load information in the personnel information, generate the personnel feature vector corresponding to each real estate agent.

[0128] S4. Establish spatial distance feature vectors between each real estate agent and the target property, and generate task matching feature vectors between each real estate agent and the target property based on the property feature vectors, the feature vectors of each agent, and the spatial distance feature vectors.

[0129] S5. Input the feature vectors of each task matching into the pre-trained task matching model to obtain the task matching score between the target property and each real estate agent, and determine the optimal agent for the target property based on the task matching score, so as to send the property listing task of the target property to the optimal agent.

[0130] The working process, working details and technical effects of the method provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0131] The fourth aspect of this embodiment provides a housing task intelligent allocation device. Taking the device as an electronic device as an example, it includes: a memory, a processor, and a transceiver that are connected in sequence. The memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the housing task intelligent allocation method as described in the third aspect of this embodiment.

[0132] For specific examples, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or first-in-last-out (FILO) memory, etc. Specifically, the processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor may be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), or PLA (Programmable Logic Array). The processor may also include a main processor and a coprocessor. The main processor, also known as the CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state.

[0133] In some embodiments, the processor may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. For example, the processor may not be limited to microprocessors of the STM32F105 series, reduced instruction set computer (RISC) microprocessors, x86 architecture processors, or processors with integrated neural network processing units (NPUs). The transceiver may be, but is not limited to, a Wi-Fi transceiver, a Bluetooth transceiver, a General Packet Radio Service (GPRS) transceiver, a ZigBee transceiver (a low-power LAN protocol based on the IEEE 802.15.4 standard), a 3G transceiver, a 4G transceiver, and / or a 5G transceiver. Furthermore, the device may also include, but is not limited to, a power module, a display screen, and other necessary components.

[0134] The working process, working details and technical effects of the electronic device provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0135] The fifth aspect of this embodiment provides a storage medium that stores instructions containing the intelligent allocation method for housing tasks as described in the third aspect of the embodiment. That is, the storage medium stores instructions that, when executed on a computer, perform the intelligent allocation method for housing tasks as described in the third aspect of the embodiment.

[0136] The storage medium refers to a carrier for storing data, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or memory sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0137] The working process, working details and technical effects of the storage medium provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0138] The sixth aspect of this embodiment provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform the intelligent allocation method for housing resources as described in the third aspect of this embodiment, wherein the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0139] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A house source task intelligent allocation system, characterized in that, include: The information entry terminal is used to obtain the property information of the target property and the personnel information of several real estate agents. The property information includes the basic attribute information of the target property, the property popularity information, the house image and the property description text, and the personnel information includes the basic information of the personnel, the personnel's historical property sales information and task load information. The housing allocation server is used to extract features from the basic attribute information and housing popularity information in the housing information, to obtain basic attribute features and housing popularity time series features, respectively. It also extracts fusion features from the house images and housing description text in the housing information to obtain text image fusion features, so as to generate housing feature vectors based on the basic attribute features, the text image fusion features and the housing popularity time series features. as well as Based on the basic information of the personnel, the personnel's historical housing sales information and task load information in the personnel information, a personnel feature vector corresponding to each real estate agent is generated. The property allocation server is used to establish spatial distance feature vectors between each real estate agent and the target property, and to generate task matching feature vectors between each real estate agent and the target property based on the property feature vectors, the feature vectors of each person, and the spatial distance feature vectors. The property allocation server is also used to input the task matching feature vectors of each task into a pre-trained task matching model to obtain the task matching score between the target property and each real estate agent, and to determine the optimal agent for the target property based on the task matching score, so as to send the property listing task of the target property to the optimal agent.

2. The system of claim 1, wherein, The housing allocation server includes a feature extraction unit, wherein, when performing fusion feature extraction on the house image and house description text in the housing information, the feature extraction unit is configured as follows: A pre-trained BERT model is used to perform semantic mapping on the property description text to obtain a text semantic vector; The house image is divided into several image blocks, and the several image blocks are input into a multi-layer image feature extraction network to generate a first image feature, a second image feature, a third image feature, and a fourth image feature, respectively. The feature dimensions of the four image features are different from each other. The fourth image feature is input into a multilayer perceptron network to obtain a global image feature vector; The text semantic vector is transformed to obtain four transformed text vectors. The feature dimensions of the four transformed text vectors are the same as the feature dimensions of the four image features, and they correspond one-to-one. The four transformed text vectors are concatenated with their corresponding image features to obtain four initial concatenated vectors. These four initial concatenated vectors are then input into the first feature mapping network for feature mapping processing to obtain four mapped concatenated vectors. Perform feature concatenation on the four mapping concatenation vectors to obtain the initial text image fusion feature vector; The initial text image fusion feature vector is input into the second feature mapping network and the third feature mapping network respectively to obtain the first text image mapping feature vector and the second text image mapping feature vector; The first text image mapping feature vector is multiplied by the global image feature vector, and the result of the dot product is weighted and summed with the second text image mapping feature vector to obtain the text image fusion feature.

3. The intelligent housing task allocation system according to claim 2, characterized in that, The multi-layer image feature extraction network includes: a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a fourth feature extraction layer. The first feature extraction layer includes: a linear embedding layer and a first feature extraction block. The second, third, and fourth feature extraction layers each include a downsampling layer and a second feature extraction block. The number of second feature extraction blocks in the second, third, and fourth feature extraction layers are 1, 3, and 1, respectively. The linear embedding layer is used to perform data mapping processing on each image patch to obtain several mapping vectors, and to generate an initial feature map based on these mapping vectors. The first feature extraction block is used to extract local and global features from the initial feature map to obtain the first image features; The downsampling layer in the second feature extraction layer is used to downsample the first image features to obtain the first downsampled features, and input the first downsampled features into the corresponding second feature extraction block to output the second image features; The downsampling layer in the third feature extraction layer is used to downsample the second image features to obtain the second downsampled features, and then input the second downsampled features into the corresponding three second feature extraction blocks in sequence to output the third image features; The downsampling layer in the fourth feature extraction layer is used to downsample the third image features to obtain the third downsampled features, and then input the third downsampled features into the corresponding second feature extraction block to obtain the fourth image features.

4. The intelligent housing task allocation system according to claim 3, characterized in that, The first feature extraction block and the second feature extraction block have the same network structure, and the first feature extraction block includes: a first LN layer, a windowed self-attention layer, a first residual layer, a second LN layer, a first multilayer perceptron layer, a second residual layer, a third LN layer, a moving window self-attention layer, a third residual layer, a fourth LN layer, a second multilayer perceptron layer, and a fourth residual layer; The first LN layer is used to perform layer normalization on the initial feature map to obtain the first normalized feature. A windowed self-attention layer is used to extract local features from the first normalized features to obtain local image features, and input the local image features into the first residual layer so that the first residual layer performs residual connection between the local image features and the initial feature map to obtain residual fusion features; The second LN layer is used to perform layer normalization on the residual fusion features to obtain normalized fusion features; The first multilayer perceptron is used to perform nonlinear transformation processing on the normalized fused features to obtain the first nonlinear transformed features, and output the first nonlinear transformed features to the second residual layer so that the second residual layer can perform residual connection between the first nonlinear transformed features and the residual fused features to obtain local output features. The third LN layer is used to perform layer normalization on the local output features to obtain the second normalized features. The second normalized features are then input into the moving window self-attention layer so that the moving window self-attention layer can extract cross-window features from the second normalized features to obtain cross-window associated image features. The third residual layer is used to perform residual connection between the cross-window associated image features and the local output features to obtain global fusion features, and input the global fusion features into the fourth LN layer so that the fourth LN layer can perform layer normalization processing on the global fusion features to obtain global normalized features. The second multilayer perceptron layer is used to perform nonlinear transformation processing on the global normalized features to obtain the second nonlinear transformation features, and input the second nonlinear transformation features into the fourth residual layer so that the fourth residual layer performs residual connection between the second nonlinear transformation features and the global fusion features to obtain the first image features.

5. The intelligent housing task allocation system according to claim 1, characterized in that, The housing allocation server includes a feature extraction unit, which is used to calculate the cosine similarity between each person's feature vector and the housing feature vector, and to multiply each person's feature vector by the housing feature vector element by element to obtain the initial task feature vector corresponding to each real estate agent. The feature extraction unit is also used to perform feature concatenation on the cosine similarity between each person feature vector and the property feature vector, the initial task feature vector corresponding to each real estate agent, and each spatial distance feature vector, so as to obtain the task matching feature vector between each real estate agent and the target property after feature concatenation.

6. The intelligent housing task allocation system according to claim 1, characterized in that, The housing allocation server includes: a model training unit, wherein the model training unit is used to acquire several historical housing allocation records, wherein any historical housing allocation record includes a sample task matching feature vector between the allocated housing and the allocated real estate agent, and the label data corresponding to the any historical housing allocation record; The model training unit is used to perform parameter optimization on the LightGBM model using a parameter optimization algorithm to obtain the optimal model parameters, and then use the optimal model parameters to update the LightGBM model to obtain the optimal LightGBM model. The model training unit is also used to train the optimal LightGBM model by taking each training data as input and the task matching score between the assigned housing and the assigned real estate agent corresponding to each training data as output, so as to obtain the task matching model after training.

7. The intelligent housing task allocation system according to claim 6, characterized in that, When using a parameter optimization algorithm to optimize the parameters of the LightGBM model, the model training unit is configured as follows: Obtain the individual population at the t-th iteration, where when t is 1, the individual population at the t-th iteration is the initial population, and each initial individual in the initial population corresponds to a set of model parameters of the LightGBM model. Based on the model parameters corresponding to each individual in the individual population, a LightGBM model corresponding to each individual is generated, and a certain amount of data in the training data is used to train the LightGBM model corresponding to each individual. Based on the output of each LightGBM model, the loss function of each LightGBM model is calculated, so as to calculate the fitness of each individual based on the loss function of each LightGBM model. Based on the fitness of each individual, the globally optimal individual at the t-th iteration is determined; Determine if the iteration stopping condition is met; If not, calculate the probability of each individual choosing the position update strategy at the t-th iteration; Based on the probability of each individual's location update strategy selection, the location update strategy of each individual is determined. The location update strategy of any individual can be a random search strategy, a local encirclement update strategy, a local development strategy, or a neighborhood mutation strategy. Based on the position update strategy for each individual, the position of each individual is updated to obtain the updated individuals, and the updated individuals are used to form the individual population at the (t+1)th iteration. Increment t by 1 and reacquire the individual population at the t-th iteration until the iteration stopping condition is met. Then, take the model parameters corresponding to the globally optimal individual at the iteration stopping condition as the optimal model parameters.

8. The intelligent housing task allocation system according to claim 7, characterized in that, The model training unit is used to calculate the position update coefficients at the t-th iteration. For any individual, the model training unit is used to determine whether the probability of the individual's location update strategy selection is less than a probability threshold, and when it is determined that the probability of the individual's location update strategy selection is less than the probability threshold, it determines whether the absolute value of the location update coefficient is greater than or equal to a preset threshold. The model training unit is also used to determine that the position update strategy of any individual is a random search strategy when the absolute value of the position update coefficient is greater than or equal to a preset threshold, or to determine that the position update strategy of any individual is a local enclosing update strategy when the absolute value of the position update coefficient is less than a preset threshold. Specifically, when the model training unit determines that the probability of selecting a location update strategy for any individual is greater than or equal to the probability threshold, it generates a selection random number. When the selection random number is less than the probability threshold, it determines that the location update strategy for any individual is a neighborhood mutation strategy, or when the selection random number is greater than or equal to the probability threshold, it determines that the location update strategy for any individual is a local exploitation strategy.

9. A method for intelligent allocation of housing resources tasks, characterized in that, include: Obtain the property information of the target property and the personnel information of several real estate agents. The property information includes the basic attribute information of the target property, the property popularity information, the house image and the property description text, and the personnel information includes the basic information of the personnel, the personnel's historical property sales information and task load information. Feature extraction is performed on the basic attribute information and housing popularity information in the housing information to obtain basic attribute features and housing popularity time series features, respectively. Then, fusion feature extraction is performed on the house image and housing description text in the housing information to obtain text image fusion features, so as to generate housing feature vector based on the basic attribute features, the text image fusion features and the housing popularity time series features. Based on the basic information of the personnel, the personnel's historical housing sales information and task load information in the personnel information, a personnel feature vector corresponding to each real estate agent is generated. Establish spatial distance feature vectors between each real estate agent and the target property, and generate task matching feature vectors between each real estate agent and the target property based on the property feature vectors, the feature vectors of each agent, and the spatial distance feature vectors. Each task matching feature vector is input into a pre-trained task matching model to obtain the task matching score between the target property and each real estate agent. Based on each task matching score, the optimal agent for the target property is determined so that the property listing task for the target property can be sent to the optimal agent.

10. A storage medium, characterized in that, The storage medium stores instructions that, when executed on a computer, perform the intelligent allocation method for housing resources as described in claim 9.

Images