Chinese address matching method and system based on word segmentation and multiple similarity judgment

Abstract

The embodiment of the invention discloses a Chinese address matching method and system based on word segmentation and multiple similarity judgment, and the method comprises the following steps: S1, address preprocessing: collecting address data, removing irrelevant information in the address data, and unifying the format; s2, word segmentation and element extraction: extracting elements in the address data based on a Chinese word segmentation technology; s3, feature vector generation: aggregating elements, and converting the elements into high-dimensional feature representation; s4, multi-similarity judgment: carrying out comprehensive judgment by adopting heterogeneous similarity measurement and a self-adaptive fusion mechanism in a multi-dimensional feature space; and S5, intelligent judgment and output: according to a judgment result, outputting final judgment of the same address or different addresses. According to the method, the address data with different sources and different expression modes can be analyzed and compared according to a unified mode, and the accuracy and the processing efficiency of the address data are improved.

Description

Technical Field

[0001] This invention relates to the field of address data processing technology, and in particular to a Chinese address matching method and system based on word segmentation and multiple similarity determination. Background Technology

[0002] With the development of the internet and information technology, address data is increasingly used in logistics, finance, e-commerce, government affairs, and other fields. However, existing methods of address data collection and processing have the following shortcomings: 1. Diverse address representations: The same address may be represented in multiple ways, lacking a unified standard, which makes data comparison and integration difficult.

[0003] 2. Non-standard structure: Some address data was not entered in a uniform format, and there were problems such as missing, redundant or disordered order, which affected subsequent processing.

[0004] 3. Limitations of traditional comparison methods: Existing methods mostly use simple string matching, which cannot effectively identify addresses with the same semantics but different expressions, resulting in low accuracy.

[0005] 4. Lack of intelligent processing: There is a lack of word segmentation and element extraction mechanisms tailored to the characteristics of Chinese addresses, making it difficult to achieve efficient automated processing.

[0006] 5. Poor business adaptability: Existing solutions are difficult to flexibly meet the needs of address standardization and similarity determination in different business scenarios, and lack scalability.

[0007] Therefore, there is an urgent need for a method that can automatically, accurately, and efficiently standardize and identify the similarity of Chinese addresses in order to improve data quality and business processing efficiency. Summary of the Invention

[0008] The technical problem to be solved by the embodiments of the present invention is to provide a Chinese address matching method and system based on word segmentation and multiple similarity determination, so as to improve data quality and business processing efficiency.

[0009] To address the aforementioned technical problems, this invention proposes a Chinese address matching method based on word segmentation and multiple similarity determination, comprising the following steps: S1. Address preprocessing: Collect address data, remove irrelevant information from the address data and standardize the format; S2. Word segmentation and element extraction: Extracting elements from address data based on Chinese word segmentation technology; S3. Feature Vector Generation: Convert the set of features into a high-dimensional feature representation; S4. Multiple similarity determination: A comprehensive determination is made by using heterogeneous similarity measurement in a multi-dimensional feature space and an adaptive fusion mechanism. S5. Intelligent Judgment and Output: Based on the judgment result, output the final judgment of whether the addresses are the same or different.

[0010] Accordingly, embodiments of the present invention also provide a Chinese address matching system based on word segmentation and multiple similarity determination, including: Address preprocessing module: Collects address data, removes irrelevant information from the address data, and standardizes the format; Word segmentation and feature extraction module: Based on Chinese word segmentation technology, extract features from address data; Feature vector generation module: Converts the feature set into a high-dimensional feature representation; Multiple similarity determination module: It adopts a heterogeneous similarity measurement in a multi-dimensional feature space and an adaptive fusion mechanism for comprehensive determination; Intelligent determination and output module: Based on the determination result, output the final determination of whether the address is the same or different.

[0011] The beneficial effects of this invention are as follows: 1. This invention improves the efficiency of address data processing: Through multi-level preprocessing and automated word segmentation and element extraction, this invention significantly improves the automation level of Chinese address standardization and comparison, and reduces manual intervention and processing costs.

[0012] 2. This invention enhances address data compatibility: This invention supports Chinese address parsing and co-address determination in multiple expression formats, avoiding adaptation problems caused by changes in address format.

[0013] 3. This invention improves the accuracy of address data: This invention adopts a multi-similarity fusion judgment mechanism, which effectively improves the recognition accuracy of the same address under different expressions, and reduces data redundancy and business misjudgment.

[0014] 4. This invention supports data scalability: This invention adapts to the scalability requirements of different business scenarios and data quality through element hierarchical extraction and dynamic threshold mechanism.

[0015] 5. This invention ensures data consistency: This invention integrates multimodal features and anomaly weighting mechanisms to ensure the consistency and reliability of address data during storage, distribution and processing.

[0016] 6. This invention supports cross-platform compatibility: This invention can be implemented in multiple programming languages, making it easy to integrate into different systems and achieve cross-platform data processing and exchange. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the processing flow of the Chinese address matching method based on word segmentation and multiple similarity determination according to an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following further describes the present invention in detail with reference to the drawings and specific embodiments.

[0019] In the embodiments of the present invention, if there are directional indications (such as up, down, left, right, front, back...), they are only used to explain the relative position relationship, movement conditions, etc. between components in a certain specific posture (as shown in the drawings). If the specific posture changes, the directional indications will also change accordingly.

[0020] In addition, in the present invention, the descriptions such as "first" and "second" are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one such feature.

[0021] The Chinese address matching method based on word segmentation and multiple similarity determination in the embodiments of the present invention includes steps S1 to S5. The processing flow is as follows Figure 1 .

[0022] S1. Address preprocessing: Collect address data, remove irrelevant information (such as administrative divisions like provinces, cities, and districts, irrelevant words, special symbols, etc.) in the address data and unify the format, and at the same time identify and standardize digital information. For example, convert Chinese numerals to Arabic numerals: "No. One hundred and twenty-three" → "No. 123"; parse the digital range: "No. 1 - 3" → ["No. 1", "No. 2", "No. 3"]; normalize the digital format: "No. 123" → "No. 123".

[0023] S2. Word segmentation and element extraction: Based on Chinese word segmentation technology (such as jieba and other Chinese word segmentation tools), extract the elements (keywords such as place names, roads, and house numbers) in the address data. Among them, digital elements are independently marked: digital information such as house numbers and floors is regarded as a special element category; map synonyms and aliases: establish an address synonym library, such as "Tsinghua University" ↔ "Tsinghua", "People's Hospital" ↔ ["People's Medical", "City People's Hospital"], "Zhongshan Road" ↔ ["Zhongshan Avenue", "Zhongshan Street"], "Panyu" ↔ ["Panyu", "Fanyu"], "Silicon Valley" ↔ ["Turtle Valley"]; label the semantics of elements: label the semantic category and confidence for each element.

[0024] Through word segmentation and element extraction, the present invention structurally processes key information such as place names, roads, and house numbers in Chinese addresses, eliminates redundancy and ambiguity, and reduces the parsing difficulty and cost.

[0025] S3. Feature Vector Generation: Convert the set of features into a high-dimensional feature representation.

[0026] Step S3 includes sub-steps S31 to S35.

[0027] S31. Hierarchical Encoding of Elements: The results of word segmentation and element extraction are hierarchically classified according to semantic categories (such as place names, roads, buildings, house numbers, directions, etc.). The element set of each category is encoded separately to form multiple sets of sub-features.

[0028] S32. Sets and Sparse Vectorization: For each feature category, construct a complete vocabulary, mapping the feature set of each address to a sparse binary vector on the complete set. If word frequency information is needed, it can be expanded into a counting vector or a TF-IDF weighted vector.

[0029] S33. Sequence Feature Encoding: In addition to set features, the element sequence after word segmentation is retained to obtain sequence features, with special processing for numerical sequences. These are used for subsequent sequence similarity calculations (such as edit distance and LCS). Positional encoding, n-gram segmentation, and other methods can be further employed to enhance the expressive power of element order and local structure. This step typically generates a fixed-length sequence representation. For example, the last hidden state of the RNN, or pooling of all hidden states (such as average pooling), can be used as a fixed-length vector for the sequence.

[0030] S34. Statistical and Semantic Feature Supplementation: Calculate the statistical features of each element category (such as coverage, word frequency distribution, part-of-speech distribution, etc.) as auxiliary feature vectors. If pre-trained word vectors (such as Word2Vec, BERT, etc.) are introduced, element terms can be embedded to obtain dense semantic vector representations, improving robustness to synonym variants.

[0031] Semantic similarity is calculated based on the cosine similarity of BERT vectors, which serves as a semantic vector (in addition to the semantic vector of the elements obtained, the semantic vector of the entire address can also be calculated).

[0032] When an address contains spatial coordinates, spatial relationship features are calculated. That is, if the address has geographic coordinates, spatial features are extracted. The steps are as follows: If latitude and longitude coordinates are available, the coordinate values ​​(longitude and latitude) can be directly used as features. Alternatively, the geographic grid code (such as Geohash) of the coordinates can be calculated, and the grid code can be embedded to obtain a vector representation. If there are no direct coordinates but multiple levels of administrative divisions, the center point of the administrative divisions can be used as approximate coordinates. Spatial hierarchy features of the area where the address is located can also be calculated, such as administrative division codes (province, city, and district codes), and these can be coded as category features.

[0033] S35. Multimodal feature concatenation: The obtained sparse binary vector, sequence features, and statistical features are concatenated into a high-dimensional composite feature vector, providing a rich feature foundation for subsequent multi-similarity fusion judgment.

[0034] S4. Multiple similarity determination: A comprehensive determination is made by using heterogeneous similarity measurement in a multi-dimensional feature space and an adaptive fusion mechanism.

[0035] Step S4 includes the following sub-steps: S41. Weighted Jaccard similarity calculation: Let the element categories of addresses A and B be i (such as place name, road, building, house number, etc.), and let the term set under each category be G. i and G i First, calculate the basic Jaccard similarity for each category: ; Then, weights are assigned to each category. w i (Weights can be pre-set based on the importance of the elements; for example, house numbers have a higher weight, while general road names have a lower weight.) The weighted Jaccard similarity is: ; S42, Block Edit Distance Metric: The word segmentation sequence is divided into segments based on element blocks (e.g., province, city, district, street, house number, etc.), and each block is defined as... B i and B i For each block, calculate the Levenshtein distance D. i Then normalize to similarity: ; Then, weights are assigned to each block. α i (Similarly, important blocks have higher weights), the weighted block edit distance similarity is: ; S43, Longest Common Subsequence (LCS) Similarity: Let the element sequences after segmenting addresses A and B be denoted as A=[a1,a2,...,a...]. m ] and B=[b1,b2,...,b n m and n are the total number of elements in the element sequences at addresses A and B, respectively. The length L of the longest common subsequence is calculated using a dynamic programming algorithm and then normalized to a similarity score. ; S44, Abnormal Weighting Mechanism: Let E be the set of elements of an address, and E be the set of key elements. key ⊂E (such as roads, house numbers, etc.); For each key element e∈E key If it is missing (i.e. not in the address), then a missing penalty factor δ is calculated. e (For example, if one key element is missing, δe = 0.8; if two are missing, δe = 0.6, etc.); the anomaly correction factor δ is calculated using an exponential decay function: ; Where λ is the attenuation coefficient, II is the indicator function (1 if missing, 0 otherwise), and p e It is the weight of the element (important elements have greater weight). Multiply the anomaly correction factor by the fusion similarity score: ; S45. Dynamic Threshold Determination: Based on the historical data distribution, an adaptive distribution algorithm is used to adjust the dynamic threshold θ. The dynamic threshold θ can be obtained through statistical learning based on features such as historical data distribution and address type. For example, the distribution of similarity scores can be calculated based on historical matching data, and then the optimal threshold can be selected based on the precision-recall curve (PR curve) or ROC curve. Alternatively, a Bayesian method can be used, assuming that the similarity scores follow a certain distribution (such as a normal distribution), and then adjusting the threshold based on prior probabilities and current data.

[0036] S46. Fusion Decision-Making: Weighted Jaccard similarity J weighted Block edit distance similarity S block LCS similarity S LCS Concatenate them into a single feature vector f; The system uses a multilayer perceptron model (MLP) or a weighted linear model to determine whether the addresses are the same. Weighted linear model: S fusion =w T f+ b ; Where w is the weight vector and b is the bias; Multilayer perceptron model: ; Where W1 and W2 are weight matrices, b1 and b2 are biases, and σ is the sigmoid function (outputting similarity between 0 and 1). Determined based on the dynamic threshold θ: .

[0037] This invention employs multiple algorithms, including Jaccard similarity and edit distance, to intelligently compare address data under different representation methods, thereby improving the accuracy of co-address identification.

[0038] S5. Intelligent Judgment and Output: Based on the judgment result, output the final judgment of whether the addresses are the same or different. This invention can set multiple thresholds to comprehensively determine whether they are the same address. This invention supports batch automatic processing and result output, reducing manual intervention and achieving efficient cleaning and deduplication of address data.

[0039] This invention can flexibly adapt to the address standardization and comparison requirements under different business scenarios, and has good scalability and versatility. This invention can be implemented in different programming languages ​​and is suitable for various application scenarios and system integration needs.

[0040] The Chinese address matching system based on word segmentation and multiple similarity determination according to embodiments of the present invention includes: Address preprocessing module: Collects address data, removes irrelevant information from the address data, and standardizes the format; Word segmentation and feature extraction module: Based on Chinese word segmentation technology, extract features from address data; Feature vector generation module: Converts the feature set into a high-dimensional feature representation; Multiple similarity determination module: It adopts a heterogeneous similarity measurement in a multi-dimensional feature space and an adaptive fusion mechanism for comprehensive determination; Intelligent determination and output module: Based on the determination result, output the final determination of whether the address is the same or different.

[0041] The address preprocessing module of this invention outputs standardized addresses for use by the word segmentation and feature extraction module. The feature set output by the word segmentation and feature extraction module is passed to the feature vector generation module. The output of the feature vector generation module serves as the input to the multiple similarity determination module. The multiple similarity determination module integrates various similarity results and submits them to the intelligent determination and output module for final decision-making.

[0042] As one implementation method, the feature vector generation module obtains a high-dimensional feature representation according to the following steps: Hierarchical encoding of elements: The results of word segmentation and element extraction are hierarchically classified according to semantic categories, and the element set of each category is encoded separately to form multiple sets of sub-features; Set and sparse vectorization: For each feature category, construct a complete vocabulary and map the feature set of each address to a sparse binary vector on the complete set; Sequence feature encoding: In addition to set features, the element sequence after word segmentation is retained to obtain sequence features; Statistical and semantic feature supplementation: Calculate the statistical features of each element category as auxiliary feature vectors; Multimodal feature concatenation: The obtained sparse binary vector, sequence features, and statistical features are concatenated into a high-dimensional composite feature vector.

[0043] As one implementation method, the feature vector generation module also calculates semantic similarity based on the cosine similarity of BERT vectors as a semantic vector. When the address contains spatial coordinates, it calculates spatial relationship features and concatenates the obtained sparse binary vector, sequence features, statistical features, semantic vector, and spatial relationship features into a high-dimensional composite feature vector.

[0044] As one implementation method, the multiple similarity determination module performs a comprehensive determination based on the following steps: Weighted Jaccard similarity calculation: Let the element categories of addresses A and B be i, and the term sets under each category be G. i and G i First, calculate the basic Jaccard similarity for each category: ; Then, weights are assigned to each category. w i The weighted Jaccard similarity is: ; Block edit distance metric: The word segmentation sequence is divided into segments based on element blocks, with each block being denoted as . B i and B i For each block, calculate the Levenshtein distance D. i Then normalize to similarity: ; Then, weights are assigned to each block. α i The weighted block edit distance similarity is: ; Longest common subsequence similarity: Let the element sequences after segmenting addresses A and B be denoted as A=[a1,a2,...,a...]. m ] and B=[b1,b2,...,b n ]; The length L of the longest common subsequence is calculated using a dynamic programming algorithm and then normalized to a similarity score. ; Abnormal weighting mechanism: Let E be the set of elements of an address, and E be the set of key elements. key ⊂E; For each key element e∈E key If missing, a missing penalty factor δ is calculated. e The anomaly correction factor δ is calculated using an exponential decay function: ; Where λ is the attenuation coefficient, II is the indicator function, and p e This is the weight of the element; Multiply the anomaly correction factor by the fusion similarity score: ; Dynamic threshold determination: Based on the distribution of historical data, the dynamic threshold θ is adjusted using a distribution adaptive algorithm; Integrated Decision Making: Weighted Jaccard similarity J weighted Block edit distance similarity S block LCS similarity S LCS Concatenate them into a single feature vector f; The model is fused using a multilayer perceptron model or a weighted linear model to determine whether they are the same address. Weighted linear model: S fusion =w T f+ b ; Where w is the weight vector and b is the bias; Multilayer perceptron model: ; Where W1 and W2 are weight matrices, b1 and b2 are biases, and σ is the sigmoid function; Determined based on the dynamic threshold θ: .

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A Chinese address matching method based on word segmentation and multiple similarity determination, characterized in that, Includes the following steps: S1. Address preprocessing: Collect address data, remove irrelevant information from the address data and standardize the format; S2. Word segmentation and element extraction: Extracting elements from address data based on Chinese word segmentation technology; S3. Feature Vector Generation: Convert the set of features into a high-dimensional feature representation; S4. Multiple similarity determination: A comprehensive determination is made by using heterogeneous similarity measurement in a multi-dimensional feature space and an adaptive fusion mechanism. S5. Intelligent Judgment and Output: Based on the judgment result, output the final judgment of whether the addresses are the same or different.

2. The Chinese address matching method based on word segmentation and multiple similarity determination as described in claim 1, characterized in that, Step S3 includes the following sub-steps: S31. Hierarchical encoding of elements: The results of word segmentation and element extraction are hierarchically classified according to semantic categories. The element set of each category is encoded separately to form multiple sets of sub-features. S32. Sets and Sparse Vectorization: For each feature category, construct a complete vocabulary and map the feature set of each address to a sparse binary vector on the complete set. S33. Sequence Feature Encoding: In addition to set features, retain the element sequence after word segmentation to obtain sequence features; S34. Statistical and Semantic Feature Supplement: Calculate the statistical features of each element category as auxiliary feature vectors; S35. Multimodal feature concatenation: The obtained sparse binary vector, sequence features, and statistical features are concatenated into a high-dimensional composite feature vector.

3. The Chinese address matching method based on word segmentation and multiple similarity determination as described in claim 2, characterized in that, In step S3, semantic similarity is also calculated based on the cosine similarity of the BERT vectors as a semantic vector. When the address contains spatial coordinates, spatial relationship features are calculated, and the obtained sparse binary vector, sequence features, statistical features, semantic vector, and spatial relationship features are concatenated into a high-dimensional composite feature vector.

4. The Chinese address matching method based on word segmentation and multiple similarity determination as described in claim 1, characterized in that, Step S4 includes the following sub-steps: S41. Weighted Jaccard similarity calculation: Let the element categories of addresses A and B be i, and the term sets under each category be G. i and G i First, calculate the basic Jaccard similarity for each category: ； Then, weights are assigned to each category. w i The weighted Jaccard similarity is: ； S42, Block Edit Distance Metric: The word segmentation sequence is divided into segments based on element blocks, with each block being denoted as . B i and B i For each block, calculate the Levenshtein distance D. i Then normalize to similarity: ； Then, weights are assigned to each block. α i The weighted block edit distance similarity is: ； S43, Longest Common Subsequence Similarity: Let the element sequences after segmenting addresses A and B be denoted as A=[a1,a2,...,a...]. m ] and B=[b1,b2,...,b n ]; The length L of the longest common subsequence is calculated using a dynamic programming algorithm and then normalized to a similarity score. ； S44, Abnormal Weighting Mechanism: Let E be the set of elements of an address, and E be the set of key elements. key ⊂E; For each key element e∈E key If missing, a missing penalty factor δ is calculated. e The anomaly correction factor δ is calculated using an exponential decay function: ； Where λ is the attenuation coefficient, II is the indicator function, and p e This is the weight of the element; Multiply the anomaly correction factor by the fusion similarity score: ； S45. Dynamic threshold determination: Based on the distribution of historical data, the dynamic threshold θ is adjusted using a distribution adaptive algorithm. S46. Fusion Decision-Making: Weighted Jaccard similarity J weighted Block edit distance similarity S block LCS similarity S LCS Concatenate them into a single feature vector f; The model is fused using a multilayer perceptron model or a weighted linear model to determine whether they are the same address. Weighted linear model: S fusion =w T f+ b ; Where w is the weight vector and b is the bias; Multilayer perceptron model: ； Where W1 and W2 are weight matrices, b1 and b2 are biases, and σ is the sigmoid function; Determined based on the dynamic threshold θ: 。 5. A Chinese address matching system based on word segmentation and multiple similarity determination, characterized in that, include: Address preprocessing module: Collects address data, removes irrelevant information from the address data, and standardizes the format; Word segmentation and feature extraction module: Based on Chinese word segmentation technology, extract features from address data; Feature vector generation module: Converts the feature set into a high-dimensional feature representation; Multiple similarity determination module: It adopts a heterogeneous similarity measurement in a multi-dimensional feature space and an adaptive fusion mechanism for comprehensive determination; Intelligent determination and output module: Based on the determination result, output the final determination of whether the address is the same or different.

6. The Chinese address matching system based on word segmentation and multiple similarity determination as described in claim 5, characterized in that, The feature vector generation module obtains high-dimensional feature representations according to the following steps: Hierarchical encoding of elements: The results of word segmentation and element extraction are hierarchically classified according to semantic categories, and the element set of each category is encoded separately to form multiple sets of sub-features; Set and sparse vectorization: For each feature category, construct a complete vocabulary and map the feature set of each address to a sparse binary vector on the complete set; Sequence feature encoding: In addition to set features, the element sequence after word segmentation is retained to obtain sequence features; Statistical and semantic feature supplementation: Calculate the statistical features of each element category as auxiliary feature vectors; Multimodal feature concatenation: The obtained sparse binary vector, sequence features, and statistical features are concatenated into a high-dimensional composite feature vector.

7. The Chinese address matching system based on word segmentation and multiple similarity determination as described in claim 6, characterized in that, The feature vector generation module also calculates semantic similarity based on the cosine similarity of BERT vectors, which serves as the semantic vector. When the address contains spatial coordinates, it calculates spatial relationship features and concatenates the obtained sparse binary vector, sequence features, statistical features, semantic vector, and spatial relationship features into a high-dimensional composite feature vector.

8. The Chinese address matching system based on word segmentation and multiple similarity determination as described in claim 5, characterized in that, The multiple similarity determination module performs a comprehensive determination based on the following steps: Weighted Jaccard similarity calculation: Let the element categories of addresses A and B be i, and the term sets under each category be G. i and G i First, calculate the basic Jaccard similarity for each category: ； Then, weights are assigned to each category. w i The weighted Jaccard similarity is: ； Block edit distance metric: The word segmentation sequence is divided into segments based on element blocks, with each block being denoted as . B i and B i For each block, calculate the Levenshtein distance D. i Then normalize to similarity: ； Then, weights are assigned to each block. α i The weighted block edit distance similarity is: ； Longest common subsequence similarity: Let the element sequences after segmenting addresses A and B be denoted as A=[a1,a2,...,a...]. m ] and B=[b1,b2,...,b n ]; The length L of the longest common subsequence is calculated using a dynamic programming algorithm and then normalized to a similarity score. ； Abnormal weighting mechanism: Let E be the set of elements of an address, and E be the set of key elements. key ⊂E; For each key element e∈E key If missing, a missing penalty factor δ is calculated. e The anomaly correction factor δ is calculated using an exponential decay function: ； Where λ is the attenuation coefficient, II is the indicator function, and p e This is the weight of the element; Multiply the anomaly correction factor by the fusion similarity score: ； Dynamic threshold determination: Based on the distribution of historical data, the dynamic threshold θ is adjusted using a distribution adaptive algorithm; Integrated Decision Making: Weighted Jaccard similarity J weighted Block edit distance similarity S block LCS similarity S LCS Concatenate them into a single feature vector f; The model is fused using a multilayer perceptron model or a weighted linear model to determine whether they are the same address. Weighted linear model: S fusion =w T f+ b ; Where w is the weight vector and b is the bias; Multilayer perceptron model: ； Where W1 and W2 are weight matrices, b1 and b2 are biases, and σ is the sigmoid function; Determined based on the dynamic threshold θ: 。

Images