Building function automatic identification method, device and medium based on POI fusion inference

Abstract

The application provides a building function automatic identification method, device and medium based on POI fusion inference. The method comprises the following steps: adopting a bidirectional long short-term memory network and a multi-head self-attention mechanism to encode POI text and acquire deep semantic features; performing collaborative naming entity recognition and function classification through a double-task learning model to obtain entity boundaries and preliminary classification probabilities of each POI; and aggregating POIs into a spatial proximity group based on geographic coordinates and Geohash coding, and fusing the preliminary classification results in the group by using a weighted voting strategy to determine the final function label of the building. The application effectively overcomes text ambiguity and spatial conflicts by combining deep semantic analysis and spatial consistency constraints, and significantly improves the accuracy and robustness of building function identification.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of automatic building function identification technology, and in particular to a method, device and medium for automatic building function identification based on POI fusion inference. Background Technology

[0002] Buildings, as fundamental units of urban spatial structure and important carriers of socio-economic activities, require accurate functional identification for urban cognition, spatial planning, and management decisions. Traditional building function identification methods mainly rely on remote sensing image data, which, while providing information on building morphology and spatial distribution, suffer from insufficient semantic information and poor timeliness.

[0003] With the widespread application of crowdsourced geographic data, especially Points of Interest (POI) data, which features timely updates, rich semantics, and dense spatial coverage, a new data foundation and technical path have been provided for the automatic identification of building functions. Existing POI-based building function identification methods mainly adopt two technical approaches: one is based on statistical analysis and traditional machine learning methods, such as kernel density analysis, buffer association, and classification models such as support vector machines and random forests. Although these methods can establish spatial associations between POIs and buildings, their classification performance is limited by the quality of feature engineering, making it difficult to fully explore the deep semantics of POI text and limiting their adaptability when dealing with complex functional scenarios. The other approach is based on deep learning methods, such as convolutional neural networks, recurrent neural networks, and their variants, which can automatically extract semantic features from POI text. However, in practical applications, they still face problems such as ambiguity in Chinese word segmentation, insufficient long-distance dependency modeling, and insufficient utilization of bidirectional context. Moreover, most methods do not fully consider the impact of geospatial proximity on function inference, leading to ambiguity in classification results when there are "same address, different name" or spatial context conflicts.

[0004] Therefore, existing technologies struggle to effectively handle the semantic complexity of POI texts while fully utilizing geospatial context information for functional inference. There is an urgent need for an automatic building function identification method that can integrate deep semantic analysis and spatial consistency constraints to improve the accuracy and robustness of identification. Summary of the Invention

[0005] This invention provides a method, device, and medium for automatic building function identification based on POI fusion inference, aiming to solve the problems of limited accuracy in building function identification caused by semantic ambiguity, insufficient context modeling, and insufficient utilization of spatial context information in existing methods when processing POI text.

[0006] According to an embodiment of the present invention, an automatic building function identification method based on POI fusion inference is provided, comprising: S1. A deep learning model based on bidirectional long short-term memory network and multi-head self-attention mechanism is used to encode POI text and obtain deep semantic features. S2. Input the deep semantic features into the dual-task learning model to obtain the entity boundary and preliminary classification probability of each POI; S3. Based on the preliminary classification probability and geographical coordinates of each POI, spatial grid aggregation is performed using Geohash encoding to form spatially adjacent POI groups, and spatial consistency voting is performed based on the preliminary classification probability of each POI in the group to determine the final functional label of the corresponding building. The dual-task learning model introduces a learnable balancing factor to construct a weighted loss function.

[0007] According to an embodiment of the present invention, an electronic device is provided, comprising: Processor; and, A memory is configured to store computer-executable instructions, which, when executed, cause the processor to perform the aforementioned automatic building function identification method based on POI fusion inference.

[0008] According to an embodiment of the present invention, a storage medium is provided for storing computer-executable instructions, which, when executed, implement the above-described automatic building function identification method based on POI fusion inference.

[0009] By integrating bidirectional semantic modeling, dual-task collaborative learning, and a spatial consistency voting mechanism, the accuracy and robustness of POI-based building function identification are significantly improved. It effectively overcomes word segmentation ambiguity and semantic vagueness in Chinese POI text, achieving joint optimization of entity recognition and function classification. Furthermore, it utilizes geographical proximity to perform spatial verification and disambiguation of classification results, thereby outputting more reliable and spatially consistent building function labels in complex urban scenarios. This makes it suitable for practical applications such as refined urban governance and planning evaluation. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of the automatic building function identification method based on POI fusion inference according to an embodiment of the present invention; Figure 2This is a diagram illustrating the overall technical architecture of the automatic building function identification method based on POI fusion inference according to an embodiment of the present invention. Detailed Implementation

[0012] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the embodiments of this invention.

[0013] Buildings, as the static framework of urban spatial structure and the dynamic carrier of human socio-economic activities, play a crucial role in understanding complex urban systems and optimizing land resource allocation through the precise distribution of their functions. With the explosive growth of crowdsourced geographic data, intelligently analyzing the physical form, spatial relationships, and socio-economic functions of buildings has become a key step in the advancement of urban information science towards deep intelligent processing. This not only provides an indispensable spatial foundation for building urban digital twins and implementing refined management but also offers a new data perspective for assessing urban vitality and guiding sustainable planning.

[0014] Early building extraction and functional classification primarily relied on manually defined geometric rules or basic GIS spatial statistics (such as kernel density analysis and buffer analysis). Wei et al. inferred building functions by analyzing the spatial relationship between buildings and Points of Interest (POIs) and calculating the frequency-density ratio of POI types. Su et al. proposed combining triangulation and buffers to establish associations between POIs and buildings, and using a Z-score statistical model for functional inference. The classification capabilities of these methods are limited by the completeness of rules and data quality, making it difficult to handle complex urban mixed-function scenarios. In contrast, crowdsourced geographic data can reveal urban dynamics by mining massive amounts of real-time user-generated information. Among these, POI data can reflect the latest urban functional layout with a high update frequency, and its wide spatial distribution effectively compensates for the uneven regional coverage of traditional data.

[0015] In recent years, the development of classification models such as Support Vector Machines, Random Forests, and Logistic Regression has promoted the advancement of building classification based on Points of Interest (POI) data. Wang et al. used a random forest model to optimize POI and building data to improve the classification accuracy and generalization ability of urban functional area identification. Chen et al. used multi-source network geographic data and unsupervised machine learning algorithms to identify residential buildings and differentiate six non-residential functions based on POI type and plot area ratio, achieving a wide-ranging building function classification. Ramalingam et al. adopted a hierarchical framework combining unsupervised clustering based on OSM road constraints with supervised learning, integrating building geometric features and multi-class POI kernel density values ​​to classify urban buildings functionally. Zhang et al. used Sina Weibo check-in data to analyze the spatiotemporal patterns of human movement and activity behavior, achieving refined POI classification through various machine learning algorithms. However, traditional classification methods only employ shallow structures and heavily rely on feature selection, resulting in less than ideal building classification results.

[0016] Deep learning-based methods primarily utilize neural network models to automatically classify buildings from large-scale unlabeled data. Niu et al. used masked region convolutional neural networks and quadratic density methods to identify buildings, supplementing the verification with POI data. Bao et al. combined ultra-high resolution imagery and POI information, using pre-trained multi-scale CNNs to extract the physical and social functional semantic information of buildings. However, CNNs, by extracting local features within a fixed receptive field, limit the representation of sequence dependencies in POI text. Recurrent neural networks (RNNs) can effectively capture the sequence features of POI data to uncover the interactions between contextual information. Kong et al. proposed a graph neural network-based building function classification method, integrating multi-source data and extracting four types of features related to building function. Zhou et al. fused satellite imagery, street view, and POI data, using a multi-label, multimodal Transformer network for hierarchical building use classification. Huang et al. aggregated POI embeddings using LSTM and attention mechanisms and calculated their importance within the region, using a multilayer perceptron (MLP) to generate regional functional distributions. However, the one-way information transmission mechanism used by the above methods cannot make coordinated use of two-way contextual information, and it is difficult to cope with the complex situation of blurred word segmentation boundaries and semantic ambiguity in Chinese, thus restricting the ability to accurately capture and express the overall semantic logic.

[0017] To address the aforementioned issues, this invention introduces a dual-task collaborative approach into building classification, proposing a building function classification method that combines POI dual-task learning with Geohash spatial voting. Figure 2The following is a diagram illustrating the overall architecture of the automatic building function identification method based on POI fusion inference according to an embodiment of the present invention. The main tasks are as follows: (1) Design a POI feature extraction framework that combines bidirectional long short-term memory network and multi-head self-attention mechanism. By capturing the bidirectional long-distance dependence of the sequence through BiLSTM, the multi-head self-attention mechanism is introduced to adaptively extract local and global deep semantic feature representations from the crowdsourced POI data based on multiple subspaces. This effectively alleviates the semantic ambiguity problem caused by insufficient use of context information and lays the foundation for subsequent accurate classification.

[0018] (2) A balance coefficient is introduced to establish a dual-task collaborative model for entity recognition and functional classification, realizing building extraction and initial functional classification. By dynamically weighting and adjusting the contribution of the two task losses to the overall optimization objective, the model can automatically balance the attention to entity boundary recognition and functional semantic discrimination according to the training stage while sharing the underlying features, thereby achieving end-to-end joint optimization and performance improvement.

[0019] (3) A Geohash spatial semantic disambiguation module is constructed. Geohash encoding is used to discretize the continuous geographic space into a standard grid, and buildings and their POIs in the same grid and neighboring grids are grouped into spatially adjacent POI groups. By statistically analyzing the classification confidence of POIs of different functional categories within the group, a weighted voting strategy is used to verify and optimize the preliminary classification results. Spatial consistency constraints are used to effectively resolve single-point classification ambiguities, thereby outputting building function classification results with higher global consistency. The details are explained below.

[0020] Method Implementation Examples According to embodiments of the present invention, an automatic building function identification method based on POI fusion inference is provided. Figure 1 This is a flowchart of the automatic building function identification method based on POI fusion inference according to an embodiment of the present invention. Figure 1 As shown, the automatic building function identification method based on POI fusion inference in this embodiment of the invention specifically includes: S1. A deep learning model based on bidirectional long short-term memory network and multi-head self-attention mechanism is used to encode POI text and obtain deep semantic features. POI texts typically contain multi-granular information at the character, phrase, and sequence levels. The model maps each character or word into a high-dimensional vector, capturing the complex phrase structure and potential functional keywords in Chinese POI texts, such as "hospital," "company," and "shopping mall."

[0021] Specifically, the method of using a deep learning model based on a bidirectional long short-term memory network and a multi-head self-attention mechanism to encode POI text includes: S11. Use a pre-trained language model to encode the POI text name to generate initial word vectors; In the sequence feature modeling process, a BERT (Bidirectional Encoder Representations from Transformers) pre-trained language model is first used to encode the text name to generate context-sensitive word vectors. Specifically, a Chinese pre-trained BERT model (such as bert-base-chinese) can be used. The character sequence of the POI name is input into BERT to obtain the context vector representation of each character, forming the initial word vector sequence. To adapt to the characteristics of POI text, BERT can be lightly fine-tuned to enhance its semantic modeling ability for geographic entity names.

[0022] S12. Input the initial word vectors into a bidirectional long short-term memory network, extract the forward and reverse sequence dependency features respectively, and concatenate them to obtain a bidirectional context feature representation; Specifically, this embodiment of the invention introduces BiLSTM (Bidirectional Long Short-Term Memory) and a multi-head self-attention mechanism. BiLSTM is used to bidirectionally model the input encoded sequence, capturing the multi-level dependencies of the text. When processing the input sequence, BiLSTM concatenates the forward and backward hidden states of each word to generate a bidirectional feature representation. Let the dimension of the unidirectional hidden state be... , No. The bidirectional output of the layer can be represented as: (1); in and These represent the forward and reverse LSTM directions, respectively, at the [number]th [time]. The hidden state of a location, when concatenated, constitutes the bidirectional feature of that location, i.e., the bidirectional contextual feature representation.

[0023] S13. Input the bidirectional contextual feature representation into the multi-head self-attention mechanism, and generate global semantic interaction features by merging the attention weights in multiple subspaces in parallel.

[0024] The feature matrix output by BiLSTM is input into a multi-head self-attention mechanism to learn and focus on similarity weights between different positions in the overall sequence, allowing the model to focus on different feature dimensions in parallel from multiple subspaces. The query vector is obtained by multiplying each element in the input sequence with the learned weight matrix. Key vector Sum value vector It is divided into multiple heads (i.e., multiple subspaces), and the calculation process is as follows: (2); set up , , The weight matrix for each head, when concatenated with the outputs of all attention heads through a linear transformation, can be represented as: (3); (4); in, This is the weight matrix used for the linear mapping.

[0025] Subsequently, multi-scale deep feature fusion is performed on the features output by the multi-head self-attention mechanism. Semantic information under different receptive fields is extracted through convolutional kernels of different scales. Combined with residual structures, the gradient vanishing problem in deep network training is effectively alleviated, enhancing the model's ability to model complex semantic structures. Specifically: Perform multi-scale one-dimensional convolution operations on the features to extract local semantic features under different receptive fields; make The feature matrix output by multi-head self-attention is used to obtain the local feature matrix through a one-dimensional convolution operation. : (5); in, This is the set of sizes for the convolution kernels.

[0026] The local semantic features extracted at each scale are concatenated and mapped along the channel dimension to form the final deep semantic features, specifically: The output features at each scale are concatenated along the channel dimension and linearly mapped to the same feature dimension to obtain the final fused feature matrix. That is, the final deep semantic features: (6); In this way, the model can extract useful local features at different scales, further enhancing its ability to model complex semantic structures. The output of this multi-scale fusion module can be added to the original output of multi-head self-attention through residual connections to preserve the global information of the original sequence and form a more informative deep semantic feature representation.

[0027] S2. Input the deep semantic features into the dual-task learning model to obtain the entity boundary and preliminary classification probability of each POI; There is a close coupling and mutual driving relationship between building entity extraction and functional classification: entity identification provides structural clues for understanding function, while functional semantics helps to confirm entity boundaries. To leverage this characteristic, this invention constructs a dual-task learning model for named entity recognition and functional classification. In the named entity recognition task, position-level emission scores are calculated based on deep semantic features, and a conditional random field layer is introduced to model the dependency between adjacent labels, outputting entity boundaries and type labels. In the functional classification task, deep semantic features are pooled and fused before being input into the classifier, outputting the functional category and preliminary classification probability of each POI. The dual-task model shares the deep semantic features extracted by S1 as a common input, but uses two independent branches at the top level to handle NER and classification tasks respectively, achieving a balance between parameter sharing and task specificity.

[0028] Furthermore, by introducing trainable adaptive balancing coefficients and dynamically weighting the loss functions of the two tasks, joint optimization is achieved. In this embodiment of the invention, the two tasks are mutually regularized while sharing underlying features, thereby improving the accuracy of entity boundary recognition and enhancing the robustness of functional semantic discrimination. Overall, this reduces misjudgments caused by uncertainties and semantic ambiguities in Chinese POI text segmentation. Specifically: In the NER task, position-level emission scores are calculated based on sequence representations generated by BiLSTM, and a learnable label transition matrix is ​​introduced into the CRF layer to explicitly model the dependency between adjacent labels, thereby achieving accurate entity boundary localization and type labeling. Let the length of the input text be... The number of entity categories is In the The location-shared feature is represented as The category score vector for this position can be represented as: (7); in, and These are the weight matrix and the bias term, respectively.

[0029] The dual-task learning model uses the cross-entropy loss function to measure the difference between the predicted probability distribution and the true label distribution. The class probability distribution at this position can be represented as: (8); in, Indicates the first The label category for each location.

[0030] To model the dependencies between tags, a Conditional Random Field (CRF) layer is further introduced to jointly score the entire sequence using emission scores and neighbor transitions, improving the consistency of entity boundaries and decoding robustness. Let the tag transition matrix be... The overall sequence score can be expressed as: (9); in, This is the predicted sequence.

[0031] The model is trained and optimized by maximizing the log-likelihood of the true label sequence: (10); in, Represents the actual label sequence. Let be the set of all possible sequences.

[0032] In the functional classification task, local semantic features extracted by multi-head self-attention are input into Softmax to complete functional category division. Specifically, max-pooling features and average-pooling features extracted from the shared feature matrix are fused using MLP to obtain global features. The softmax function is used to convert the model's predicted scores into probability values ​​(i.e., the confidence score for each class), and its classification probability distribution is shown below. It can be represented as: (11); in, This is the weight matrix. This is the bias function.

[0033] The cross-entropy loss function can be expressed as: (12); in, For the number of functional categories, This represents the one-hot encoding of the true category. Functional categories can be defined according to actual applications, such as commercial, residential, office, medical, educational, transportation, and leisure. The initial classification probability vector output by the classifier reflects the confidence level of the POI belonging to each category, providing a basis for subsequent spatial voting.

[0034] Considering the differences in objective form between NER and functional classification tasks, directly adding their loss functions may lead to training instability or gradient conflicts. Therefore, this invention aims to design a weighted joint optimization objective loss function by introducing a balancing factor to balance the relative contributions of the two tasks, thereby achieving stable training on shared representations and obtaining synergistic performance improvements.

[0035] During training, the total loss function for and The weighted sum of the two losses can be expressed as: (13); in, This is a balancing factor used to adjust the relative contributions of the two tasks to the overall training.

[0036] S3. Based on the preliminary classification probability and geographical coordinates of each POI, spatial grid aggregation is performed using Geohash encoding to form spatially adjacent POI groups, and spatial consistency voting is performed based on the preliminary classification probability of each POI in the group to determine the final functional label of the corresponding building. In building function classification tasks, geospatial information is crucial. The functional context formed by a building and its surrounding Points of Interest (POIs) has a strong spatial dependency, and accurately understanding the local spatial context is key to eliminating functional ambiguity. Based on Tobler's First Law, which states that geographical features are spatially related and closely related features are more closely connected, existing methods, even after resolving semantic ambiguity in the text, still suffer from spatial ambiguity such as conflicting descriptions of POIs with the same address but different names for the same building, and contradictory functional categories of neighboring POIs.

[0037] Therefore, step S3 of this embodiment of the invention is used to solve this type of spatial ambiguity problem. The core is to convert continuous latitude and longitude coordinates into discrete grid identifiers with spatial order preservation, thereby transforming geographical proximity into a computable feature, as follows: The spatial grid aggregation using Geohash encoding specifically involves: By employing Geohash encoding with specific precision, continuous geographic space is discretized into a standard grid. In this embodiment, a Geohash encoding of length 9 is selected, discretizing the space into a grid with a side length of approximately 4.77 meters. This precision design achieves a dual purpose: it can distinguish different entities at the building level, and it ensures that all POIs within the same building are aggregated into the same spatial unit, providing an ideal grid foundation for subsequent conflict resolution and weighted voting based on spatial consistency. In practical applications, the Geohash precision level (e.g., level 7-10) can be dynamically adjusted according to the characteristics of urban areas (e.g., high-density areas in the city center versus low-density areas in the suburbs) to balance spatial resolution and computational efficiency.

[0038] Points of Interest (POIs) with similar geographical locations often possess similar functional attributes. Therefore, this invention uses Geohash encoding to discretize spatially continuous latitude and longitude coordinates into a hierarchical spatial grid, leveraging the principle of spatial locality to achieve efficient spatial querying and aggregation. POIs whose latitude and longitude coordinates fall within the same Geohash grid are aggregated into a group of spatially neighboring POIs. To enhance the spatial context, aggregation can be extended to the current grid's eight direct neighbor grids, forming a local spatial neighborhood, thereby better capturing functional consistency within the microscopic spatial environment.

[0039] For each POI's latitude and longitude coordinates, Geohash encoding identifies spatially adjacent POI groups through spatial mapping, ensuring that geographically close POIs have the same or similar encoding prefixes. Location The Geohash encoding function is defined as follows: (14); in, and They represent the first The longitude and latitude positions of the second bisection Indicates bit concatenation operation. Indicates bit-interleaved combination. For Base32 encoding functions, This is the coding precision level (corresponding to a spatial resolution of approximately 4.77m × 4.77m).

[0040] The spatial clustering degree of POIs within a group to reflect its spatial consistency can be expressed as: (15); in, For the first A spatial group, This is a function for calculating the length of the longest common prefix. The most frequent Geohash code within the group. This represents the total encoding length.

[0041] The spatial voting weights of each group are obtained through normalization, and then combined with the classification confidence to perform consensus voting to obtain the classification result. (16); (17); in, For a collection of functional categories, For indicator functions, for The original prediction category, For position The classification confidence probability is calculated. This weighted voting mechanism comprehensively considers the classification confidence of each POI and the aggregation weight of its local spatial group, which can effectively suppress noise from outlier POIs, improve the reliability of group decision-making, and ultimately output building function labels with consistent spatial semantics.

[0042] The embodiments of the present invention have the following beneficial effects: By integrating bidirectional semantic modeling, dual-task collaborative learning, and a spatial consistency voting mechanism, the accuracy and robustness of POI-based building function identification are significantly improved. It effectively overcomes word segmentation ambiguity and semantic vagueness in Chinese POI text, achieving joint optimization of entity recognition and function classification. Furthermore, it utilizes geographical proximity to perform spatial verification and disambiguation of classification results, thereby outputting more reliable and spatially consistent building function labels in complex urban scenarios. This makes it suitable for practical applications such as refined urban governance and planning evaluation.

[0043] Device Example 1 According to an embodiment of the present invention, an electronic device is provided, comprising: Processor; and, A memory is configured to store computer-executable instructions that, when executed, cause the processor to perform the steps of the method embodiments described above.

[0044] Device Example 2 According to an embodiment of the present invention, a storage medium is provided for storing computer-executable instructions, which, when executed, implement the steps of the method embodiments described above.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatic building function identification based on POI fusion inference, characterized in that... include: S1. A deep learning model based on bidirectional long short-term memory network and multi-head self-attention mechanism is used to encode POI text and obtain deep semantic features. S2. Input the deep semantic features into the dual-task learning model to obtain the entity boundary and preliminary classification probability of each POI; S3. Based on the preliminary classification probability and geographical coordinates of each POI, spatial grid aggregation is performed using Geohash encoding to form spatially adjacent POI groups. Based on a weighted voting strategy, spatial consistency voting is performed on the preliminary classification probability of each POI in the group to determine the final functional label of the corresponding building. The dual-task learning model introduces a learnable balancing factor to construct a weighted loss function.

2. The method according to claim 1, characterized in that, The method employs a deep learning model based on a bidirectional long short-term memory network and a multi-head self-attention mechanism to encode POI text, specifically including: Initial word vectors are generated by encoding POI text names using a pre-trained language model; The initial word vectors are input into a bidirectional long short-term memory network, and the forward and reverse sequence dependency features are extracted and concatenated to obtain a bidirectional contextual feature representation. The bidirectional contextual feature representation is input into a multi-head self-attention mechanism, and global semantic interaction features are generated by fusion through parallel computation of attention weights in multiple subspaces.

3. The method according to claim 2, characterized in that, After generating global semantic interaction features through fusion, the process also includes a multi-scale deep feature fusion operation on the features output by the multi-head self-attention mechanism, specifically as follows: Perform multi-scale one-dimensional convolution operations on the features to extract local semantic features under different receptive fields; The local semantic features extracted at each scale are spliced ​​and mapped in the channel dimension to form the final deep semantic features.

4. The method according to claim 1, characterized in that, The dual-task learning model includes: named entity recognition task and functional classification task; In the named entity recognition task, position-level emission scores are calculated based on deep semantic features, and a conditional random field layer is introduced to model the dependency between adjacent labels, outputting entity boundaries and type labels; In the functional classification task, deep semantic features are pooled and fused before being input into the classifier, which outputs the functional category of each POI and its preliminary classification probability.

5. The method according to claim 1, characterized in that, The learnable balancing factor constructs a weighted loss function for the named entity recognition task loss. Loss of Functional Classification Tasks Through the aforementioned balance factor Weighted summation yields: ； in, This is a balancing factor used to adjust the relative contributions of the two tasks to the overall training.

6. The method according to claim 1, characterized in that, The spatial grid aggregation using Geohash encoding specifically involves: Using Geohash encoding with a specific precision, continuous geospatial space is discretized into a standard grid; POIs whose latitude and longitude coordinates fall within the same Geohash grid are aggregated into a group of spatially adjacent POIs.

7. The method according to claim 6, characterized in that, In the weighted voting strategy, the spatial voting weight of each POI is calculated based on the spatial clustering degree of its spatial neighboring POI group. The spatial clustering degree is determined according to the common prefix length of the Geohash codes of all POIs in the group.

8. The method according to claim 7, characterized in that, The calculation of the spatial voting weights includes: normalizing the length of the common prefix of the Geohash encoding to obtain the normalized weight value of each POI.

9. An electronic device, comprising: processor; as well as, A memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the automatic building function identification method based on POI fusion inference as described in any one of claims 1-8.

10. A storage medium for storing computer-executable instructions, which, when executed, implement the automatic building function identification method based on POI fusion inference as described in any one of claims 1-8.

Images