Social network data mining method based on concept lattice convolutional neural network

Abstract

The invention provides a social network data mining method based on a concept lattice convolutional neural network. The social network data mining method comprises the following steps: S1, collecting and preprocessing social network data; s2, constructing a concept lattice; s3, concept lattice feature extraction and curved surface fitting optimization; s4, constructing a fusion model of the concept lattice convolutional neural network and the mixed density network; and S5, analyzing and feeding back a social network data mining result. Through collaborative fusion of multiple technical modules, the defects of a traditional method are effectively overcome, and a remarkable technical effect is achieved.

Description

Technical Field

[0001] This invention relates to the field of social network data mining technology, and in particular to a social network data mining method based on a concept lattice convolutional neural network. Background Technology

[0002] With the rapid development of social networks, the number of users and the scale of data are growing exponentially. Social network data encompasses multi-dimensional information such as basic user information, published content, and interactive behaviors, containing valuable knowledge such as user interests and behavioral patterns. However, this data is characterized by high dimensionality, non-linearity, and multimodality, posing numerous challenges to traditional data mining methods when processing it.

[0003] In traditional social network data mining methods, while concept lattice-based methods can express the hierarchical relationship between objects and attributes, they are difficult to capture the non-linear correlation between features, and the utilization of concept lattice features is mostly limited to a shallow level. Convolutional neural networks perform well in feature extraction, but when directly applied to structured data such as concept lattices, they are difficult to adapt to their hierarchical structure. Single probability models cannot accurately characterize the diversity and complexity of user interests, which may have multi-peak distributions, when dealing with such data, resulting in insufficient accuracy and robustness of the mining results.

[0004] Furthermore, in existing methods, different technical modules often work independently, lacking an effective collaborative integration mechanism. Data such as concept lattice features and neural network outputs are difficult to interact effectively, which prevents each module from fully leveraging its advantages and restricts the depth and accuracy of social network data mining. Summary of the Invention

[0005] This invention provides a social network data mining method based on a concept lattice convolutional neural network. Through the synergistic integration of multiple technical modules, it effectively overcomes the shortcomings of traditional methods and achieves significant technical results.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A social network data mining method based on concept lattice convolutional neural networks includes: S1: Social Network Data Acquisition and Preprocessing; Raw social network data is acquired through the application programming interface of the social network platform. The raw social network data includes basic user information, content information posted by users, and interaction information between users. The raw social network data is cleaned, deduplicated, and standardized to obtain preprocessed social network data. The preprocessed social network data includes a preprocessed set of basic user information, a preprocessed set of content information posted by users, and a preprocessed set of interaction information between users. S2: Concept lattice construction; Based on the preprocessed user basic information set, preprocessed user-published content information set, and preprocessed user interaction information set obtained in S1, a concept lattice is constructed. Each concept in the concept lattice consists of extension and intension. The extension is the set of users included in the concept, and the intension is the set of keywords and interaction features common to these users. The construction process includes extracting keywords from the user-published content text, calculating interaction features by combining user interaction information, constructing formal background, and then using the NextClosure algorithm to generate the concept lattice. S3: Concept lattice feature extraction and surface fitting optimization; Features are extracted from the concept lattice constructed in S2 to form a concept lattice feature vector. The concept lattice feature vector includes the extension size, connotation size, number of parent concepts, number of child concepts, and interaction feature statistics of the concept. The concept lattice feature vector is optimized by using binary polynomial surface fitting to obtain the fitted feature vector. S4: Construct a fusion model of concept lattice convolutional neural network and hybrid density network; the fusion model takes the fitted feature vector obtained in S3 as input and includes a concept lattice convolutional neural network sub-model and a hybrid density network sub-model. The concept lattice convolutional neural network sub-model is used to extract deep local features, and the hybrid density network sub-model is used to model the multi-peak distribution of user interests. The optimized fusion model is obtained through training, and the optimized fusion model is used to predict user interests. S5: Analysis and feedback of social network data mining results; compare the user interest prediction results obtained in S4 with the actual results, calculate the evaluation index, optimize and adjust the model or preprocessing process according to the evaluation index results until the model performance meets the requirements, and output the final social network data mining results.

[0007] In this specification, in S1, the basic user information includes a unique user identifier, user age, and user gender; the content information published by the user includes a unique content identifier, content text, and publication time; the interaction information between users includes a unique interaction identifier, interaction type, unique identifier of the user initiating the interaction, unique identifier of the user receiving the interaction, and interaction time; the cleaning process is to delete records containing missing values; the deduplication process is to delete duplicate user-published content information; the standardization process is to standardize the user age to the [0,1] interval, and the standardization formula is that the standardized user age is equal to (user age minus the minimum value among all user ages) divided by (the maximum value among all user ages minus the minimum value among all user ages).

[0008] In this specification, in S2, the keyword extraction of user-published content text uses the TF-IDF algorithm, including calculating word frequency, calculating inverse document frequency, calculating TF-IDF value, and selecting the top K keywords with the highest TF-IDF value in each content text; the formal background is a triple (U, A, R), where U is the user set, A is the union of the attribute sets of all users, and R is the set of relationships between users and attributes; the process of generating concept lattices using the NextClosure algorithm is initialization, generating new connotations and judging closure, and repeating the generation process until no new concepts can be generated.

[0009] In this specification, S3, extracting features from the concept lattice includes calculating the extension and intension of each concept, defining the parent-child relationship between concepts, and recording the number of parent concepts and child concepts for each concept; the feature vector of the concept lattice is (extension, intension, number of parent concepts, number of child concepts); the polynomial order of the bivariate polynomial surface fitting is 2, and the fitting formula is that the fitted feature value is equal to the sum of the products of multiple surface fitting coefficients and the powers of the corresponding variables. The surface fitting coefficients are solved by the least squares method to generate the fitted feature vector.

[0010] In this specification, S4, the conceptual lattice convolutional neural network sub-model includes an input layer, a convolutional layer, an activation function, a pooling layer, and a fully connected layer; the input layer receives the fitted feature vector; the convolutional layer uses multiple 1×2 convolutional kernels to perform convolution operations on the input features; the activation function is the ReLU function; the pooling layer uses max pooling; the fully connected layer maps the pooling results to a low-dimensional space and outputs feature values.

[0011] In this specification, in S4, the hybrid density network sub-model takes the feature values ​​output by the concept lattice convolutional neural network sub-model as input, including three types of parameters: mixing coefficient, mean, and standard deviation. The mixing coefficient is calculated using the softmax function, the mean is calculated using a linear transformation, and the standard deviation is calculated using an exponential function. The probability density function of the hybrid density network sub-model is the sum of the products of multiple Gaussian probability density functions and their corresponding mixing coefficients.

[0012] In this specification, S4, the training process of the fusion model includes preparing training data, with the fitted feature vector as input and the user's actual interest category as the label; initializing model parameters; calculating the loss function, which adopts the negative log-likelihood loss function; performing backpropagation through the Adam optimizer to update the model parameters; and iteratively training until the loss function converges or the preset maximum number of iterations is reached.

[0013] In this specification, in S4, the process of predicting user interests using the optimized fusion model is as follows: the fitted feature vector of the new concept is input into the concept lattice convolutional neural network sub-model to obtain feature values, the feature values ​​are input into the hybrid density network sub-model to calculate the probability density distribution of user interest categories, and the interest category with the highest probability is selected as the prediction result.

[0014] In this specification, S5 includes evaluation metrics such as accuracy, precision, recall, and F1 score. Accuracy is the ratio of correctly predicted samples to the total number of samples. Precision is the proportion of samples predicted as positive that are actually positive. Recall is the proportion of samples that are actually positive that are correctly predicted as positive. The F1 score is the harmonic mean of precision and recall.

[0015] In this specification, S5 includes optimizing and adjusting the model or preprocessing process, including adjusting the polynomial order of the surface fitting, adjusting the number of Gaussian components in the mixed density network, increasing the number of training samples, adjusting the weights of the training samples, and redefining the dimension of the concept lattice feature vector.

[0016] In summary, the present invention has at least the following beneficial effects: First, by constructing a concept lattice, the relationship between users and content is clearly expressed in the form of hierarchical concepts, laying a structured foundation for subsequent feature extraction and enabling the semantic relationships in the data to be preserved and utilized.

[0017] Secondly, the introduction of surface fitting technology optimizes the feature vector of the concept lattice, successfully capturing the nonlinear relationship between features, mapping high-dimensional discrete features to a continuous surface space, improving the expressive power of features, and providing better input for neural networks.

[0018] Furthermore, the fusion of Conceptual Lattice Convolutional Neural Network (CL-CNN) and Hybrid Density Network (MDN) fully leverages CL-CNN's ability to extract local features and MDN's advantage in modeling multimodal distributions. CL-CNN extracts deep local features from the optimized features, while MDN, based on its output, accurately characterizes the multimodal distribution of user interests, significantly improving the accuracy of tasks such as interest prediction.

[0019] Meanwhile, the various technical modules achieve close collaboration through a clear data transmission and feedback mechanism. Conceptual lattice features provide input for surface fitting, surface fitting results provide optimization features for CL-CNN, and CL-CNN output drives the parameter calculation of MDN, forming a complete technical closed loop. This allows the advantages of each module to be superimposed, effectively improving the overall performance of social network data mining. It can more accurately mine valuable information such as user interests and behavioral patterns, providing strong technical support for applications such as personalized recommendation and precision marketing. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are 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.

[0021] Figure 1 This is a schematic diagram of the social network data mining method based on concept lattice convolutional neural networks involved in this invention.

[0022] Figure 2 This is a schematic diagram of the social network data collection and preprocessing process involved in this invention.

[0023] Figure 3 This is a schematic diagram of the concept lattice construction process involved in this invention.

[0024] Figure 4 This is a schematic diagram of the feature extraction and surface fitting process involved in this invention.

[0025] Figure 5 This is a schematic diagram illustrating the construction and application process of the CL-CNN-MDN model involved in this invention. Detailed Implementation

[0026] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0027] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0028] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] like Figure 1 As shown, this embodiment provides a social network data mining method based on a concept lattice convolutional neural network, including: S1: Social Network Data Acquisition and Preprocessing; Raw social network data is acquired through the application programming interface of the social network platform. The raw social network data includes basic user information, content information posted by users, and interaction information between users. The raw social network data is cleaned, deduplicated, and standardized to obtain preprocessed social network data. The preprocessed social network data includes a preprocessed set of basic user information, a preprocessed set of content information posted by users, and a preprocessed set of interaction information between users. S2: Concept lattice construction; Based on the preprocessed user basic information set, preprocessed user-published content information set, and preprocessed user interaction information set obtained in S1, a concept lattice is constructed. Each concept in the concept lattice consists of extension and intension. The extension is the set of users included in the concept, and the intension is the set of keywords and interaction features common to these users. The construction process includes extracting keywords from the user-published content text, calculating interaction features by combining user interaction information, constructing formal background, and then using the NextClosure algorithm to generate the concept lattice. S3: Concept lattice feature extraction and surface fitting optimization; Features are extracted from the concept lattice constructed in S2 to form a concept lattice feature vector. The concept lattice feature vector includes the extension size, connotation size, number of parent concepts, number of child concepts, and interaction feature statistics of the concept. The concept lattice feature vector is optimized by using binary polynomial surface fitting to obtain the fitted feature vector. S4: Construct a fusion model of concept lattice convolutional neural network and hybrid density network; the fusion model takes the fitted feature vector obtained in S3 as input and includes a concept lattice convolutional neural network sub-model and a hybrid density network sub-model. The concept lattice convolutional neural network sub-model is used to extract deep local features, and the hybrid density network sub-model is used to model the multi-peak distribution of user interests. The optimized fusion model is obtained through training, and the optimized fusion model is used to predict user interests. S5: Analysis and feedback of social network data mining results; compare the user interest prediction results obtained in S4 with the actual results, calculate the evaluation index, optimize and adjust the model or preprocessing process according to the evaluation index results until the model performance meets the requirements, and output the final social network data mining results.

[0030] In some embodiments, in S1, the basic user information includes a unique user identifier, user age, and user gender; the content information published by the user includes a unique content identifier, content text, and publication time; the interaction information between users includes a unique interaction identifier, interaction type, unique identifier of the user initiating the interaction, unique identifier of the user receiving the interaction, and interaction time; the cleaning process involves deleting records containing missing values; the deduplication process involves deleting duplicate user-published content information; and the standardization process involves standardizing the user age to the [0,1] interval, with the standardization formula being that the standardized user age equals (user age minus the minimum of all user ages) divided by (maximum of all user ages minus the minimum of all user ages).

[0031] In some embodiments, in S2, keyword extraction of user-published content text uses the TF-IDF algorithm, including calculating word frequency, calculating inverse document frequency, calculating TF-IDF value, and selecting the top K keywords with the highest TF-IDF value in each content text; the formal background is a triple (U, A, R), where U is the user set, A is the union of the attribute sets of all users, and R is the set of relationships between users and attributes; the process of generating concept lattices using the NextClosure algorithm is initialization, generating new connotations and judging closure, and repeating the generation process until no new concepts can be generated.

[0032] In some embodiments, S3, extracting features from the concept lattice includes calculating the extension and intension of each concept, defining the parent-child relationship between concepts and recording the number of parent concepts and child concepts for each concept; the feature vector of the concept lattice is (extension, intension, number of parent concepts, number of child concepts); the polynomial order of the bivariate polynomial surface fitting is 2, and the fitting formula is that the fitted feature value is equal to the sum of the products of multiple surface fitting coefficients and the powers of the corresponding variables. The surface fitting coefficients are solved by the least squares method to generate the fitted feature vector.

[0033] In some embodiments, S4, the conceptual lattice convolutional neural network sub-model includes an input layer, a convolutional layer, an activation function, a pooling layer, and a fully connected layer; the input layer receives the fitted feature vector; the convolutional layer uses multiple 1×2 convolutional kernels to perform convolution operations on the input features; the activation function adopts the ReLU function; the pooling layer adopts max pooling operation; the fully connected layer maps the pooling result to a low-dimensional space and outputs feature values.

[0034] In some embodiments, in S4, the hybrid density network sub-model takes the feature values ​​output by the concept lattice convolutional neural network sub-model as input, including three types of parameters: mixing coefficients, mean, and standard deviation. The mixing coefficients are calculated using the softmax function, the mean is calculated using a linear transformation, and the standard deviation is calculated using an exponential function. The probability density function of the hybrid density network sub-model is the sum of the products of multiple Gaussian probability density functions and their corresponding mixing coefficients.

[0035] In some embodiments, in S4, the training process of the fusion model includes preparing training data, with the fitted feature vector as input and the user's actual interest category as label; initializing model parameters; calculating the loss function, which adopts the negative log-likelihood loss function; performing backpropagation through the Adam optimizer to update the model parameters; and iteratively training until the loss function converges or reaches the preset maximum number of iterations.

[0036] In some embodiments, in S4, the process of predicting user interests using the optimized fusion model involves inputting the fitted feature vector of the new concept into the concept lattice convolutional neural network sub-model to obtain feature values, inputting the feature values ​​into the hybrid density network sub-model to calculate the probability density distribution of user interest categories, and selecting the interest category with the highest probability as the prediction result.

[0037] In some embodiments, S5, the evaluation metrics include accuracy, precision, recall, and F1 score; accuracy is the ratio of correctly predicted samples to the total number of samples; precision is the proportion of samples predicted as positive that are actually positive; recall is the proportion of samples that are actually positive that are correctly predicted as positive; and F1 score is the harmonic mean of precision and recall.

[0038] In some embodiments, S5 includes optimizing the model or preprocessing process by adjusting the polynomial order of the surface fitting, adjusting the number of Gaussian components in the mixed density network, increasing the number of training samples, adjusting the weights of the training samples, and redefining the dimension of the concept lattice feature vector.

[0039] The technical concept of this invention is as follows: S1: Social Network Data Acquisition and Preprocessing In the initial stage of social network data mining, data collection and preprocessing are fundamental to all subsequent analyses, and their quality directly affects the reliability of the final results. This step systematically acquires and processes raw data, laying a solid foundation for subsequent operations. The social network data collection and preprocessing workflow is as follows: Figure 2 As shown.

[0040] First, the data collection process is implemented through the Application Programming Interface (API) provided by the social network platform. This process must strictly adhere to the platform's data access specifications to ensure the legality and integrity of the acquired data. The collected raw social network data covers three core dimensions: firstly, basic user information, including the user's unique identifier (using...). This indicates that i is the user serial number. N represents the total number of users), user age ( User gender () ,in =0 indicates female. =1 indicates male); secondly, the content information posted by the user, including a unique content identifier ( j is the content sequence number. M represents the total number of contents, and the content text ( ), release time ( Third, user interaction information involves unique interaction identifiers (...). k is the interaction number. P represents the total number of interactions, and the interaction type ( Likes, comments, and shares correspond to... =1、 =2、 =3), the unique identifier of the user who initiated the interaction ( ), interactive receiving user unique identifier ( ) and interactive time ( ).

[0041] After obtaining the raw data, multiple preprocessing steps are required to eliminate noise, standardize the format, and improve data usability: 1. Data Cleaning: Examine each record one by one and remove entries with missing key information. Specifically, records with missing age in user basic information, empty text in user-posted content, and missing initiator or recipient identifiers in user interaction information are all subject to cleaning, as these omissions will render subsequent analysis meaningless or introduce bias.

[0042] 2. Data Deduplication: For user-posted content, if two records contain duplicate text (…), then the deduplication process is performed. ) are completely identical and the release time is ( If the records are identical, they are considered duplicates, and only one of them is retained. This step avoids duplicate information interfering with feature extraction and model training, reducing data redundancy.

[0043] 3. Data standardization: Due to user age ( The numerical range of ) may vary considerably. To facilitate subsequent model processing, it needs to be standardized to the [0,1] interval. The standardization formula is: ; in, This represents the standardized user age. It is the minimum age among all users. It is the maximum age among all users. For example, if the users' ages are 18, 22, and 25, then... =18, =25, the standardized value of a user aged 22 is .

[0044] After the above preprocessing, a structured social network dataset is obtained, including a preprocessed set of basic user information. (each) Include and ), collection of user-published content information (each) Include and ), and user interaction information collection (each) Include , , and ).

[0045] S2: Concept Lattice Construction Concept lattices, as a mathematical model that clearly expresses the hierarchical relationships between objects and attributes, are a key tool in this method for integrating user and content features. Their construction process is based on the preprocessed user and content data in S1, systematically transforming scattered information into a structured conceptual hierarchy. The concept lattice construction process is as follows: Figure 3 As shown.

[0046] The core idea behind constructing a concept grid is to treat users as "objects" and keywords extracted from user-posted content as "attributes," thereby establishing a relationship between objects and attributes. The specific steps are as follows: First, keyword extraction is performed. The TF-IDF (Term Frequency-Inverse Document Frequency) algorithm is used to extract keywords from the user-published content text (…). This algorithm extracts representative keywords from a single text, effectively balancing the importance of words in a single text with their prevalence across the entire text collection. 1. Calculate word frequency ( ): refers to the keyword w in the content text The ratio of the number of times a word appears in a text to the total number of words in the text is given by the formula: ; in, The keyword is w. The number of times it appears in yes The total number of occurrences of all words in the text. For example, in the text "Social network data mining is important" which contains 6 words, if "social" appears once, then... .

[0047] 2. Calculate the inverse document frequency (IVF) The formula reflects the prevalence of keyword w across all texts: ; Where M is the total number of content texts. This is the number of texts containing "w". Adding 1 is to avoid a denominator of 0. For example, if M=100 and there are 20 texts containing "social", then... (Natural logarithm).

[0048] 3. Calculate the TF-IDF value: Multiply the term frequency by the inverse document frequency to obtain the result. The top K (e.g., K=5) TF-IDF values ​​of each text are selected as the keywords of that text.

[0049] Next, for each user Build property collection This is the union of the keyword sets of all content text posted by the user. Based on this, a formal context is constructed, which is a triple (U, A, R), where U is the set of users, A is the union of all user attribute sets, and R is the set of relationships between users and attributes: if user Has attributes ,but .

[0050] The concept lattice is constructed using the NextClosure algorithm, which iteratively generates the intensions of all concepts to gradually build a complete concept lattice. 1. Initialization: The empty set is taken as the first connotation, and its corresponding extension is the set of all users (that is, all users share the characteristic of "no attributes").

[0051] 2. Generating New Connotations and Determining Closures: For each generated connotation, a new connotation is generated by adding previously unincluded attributes. If the new connotation is a closure (i.e., all objects of the connotation share attributes that are also the connotation itself), its corresponding extension (i.e., all users who possess all attributes of the connotation) is calculated, and this concept (extension + connotation) is added to the concept lattice.

[0052] 3. Iteration Termination: Repeat step 2 until no new concepts can be generated, at which point the concept lattice construction is complete.

[0053] The resulting concept lattice L contains multiple concepts l = (Ext(l), Int(l)), where Ext(l) is the set of users covered by concept l, and Int(l) is the set of keywords shared by these users.

[0054] S3: Concept Lattice Feature Extraction and Surface Fitting Optimization Feature extraction of concept lattices is the process of transforming the structural information of concept lattices into computable numerical features, while surface fitting further explores the nonlinear relationships between features, enhancing their expressive power. The feature extraction and surface fitting processes are as follows: Figure 4 As shown.

[0055] First, extract the basic features from the concept lattice L: 1. Extension size: The extension Ext(l) of each concept l is a group of users, and its size |Ext(l)| is the number of users in the group, reflecting the scale of users covered by the concept.

[0056] 2. Connotation size: Connotation Int(l) is a set of keywords, and its size |Int(l)| is the number of keywords, which reflects the richness of the concept's attributes.

[0057] 3. Parent-child relationship: If the concept Extensional inclusion extension ( ), and its connotation was The connotation includes ( ), and there is no intermediate concept. If the same inclusion relationship is satisfied, then yes The parent concept, yes Sub-concepts. Record the number of parent concepts for each concept. Number of sub-concepts .

[0058] The above features are integrated into a concept lattice feature vector. This vector describes the characteristics of a concept from three dimensions: scale, attributes, and hierarchical relationships.

[0059] To capture the nonlinear relationships between features, a bivariate polynomial surface fitting pair is introduced. Optimize to obtain Let the order of the fitted polynomial be 2, and the formula be: ; Where x and y are The two dimensions in (e.g., x=|Ext(l)|, y=|Int(l)|). The fitting coefficients are p and q, where p and q are the order indices. These are the fitted feature values.

[0060] The fitting process is as follows: 1. Construct training samples: As input, with The output label is the weighted sum of the number of parent and child concepts (weights are set according to feature importance).

[0061] 2. Solve for the fitting coefficients Minimize the error using the least squares method. (L represents the total number of concepts). By calculating E... Taking the partial derivatives and setting them to zero, we obtain a system of linear equations, which can be solved to obtain... .

[0062] 3. Generate optimized features: Substitute the (x,y) values ​​of all concepts into the fitting formula to obtain... The optimized feature vector is (Preserve core features and integrate non-linear relationships). For example, =5, =3, ,but ,Right now .

[0063] S4: Constructing a fusion model of conceptual lattice convolutional neural networks and hybrid density networks To achieve accurate prediction of user interests, a fusion model of Concept Lattice Convolutional Neural Network (CL-CNN) and Hybrid Density Network (MDN) (CL-CNN-MDN) is constructed. This model uses the feature vectors optimized in S3.1. Using this as input, it fully combines the feature extraction capabilities of convolutional neural networks with the modeling capabilities of hybrid density networks for multi-peak distributions. The CL-CNN-MDN model construction and application process is as follows: Figure 5 As shown.

[0064] S4.1 Conceptual Lattice Convolutional Neural Network (CL-CNN) Sub-model The core function of CL-CNN is to extract deep local features from the optimized feature vectors. Its structure includes: 1. Input layer: receiving The number of neurons is 3 (and (The dimensions are consistent), responsible for feeding the feature vectors into the network.

[0065] 2. Convolutional layers: using =Eight 1×2 convolutional kernels are used to perform convolution operations on the input features through a sliding window to extract locally related features. The convolution formula is: ; in, yes The (i+m)th element It is the m-th weight of the k-th convolutional kernel. It is a bias term. For example, =(5,3,4.65), convolution kernel =0.2, =0.3, =0.1, then =5×0.2+3×0.3+0.1=2.0, =3×0.2+4.65×0.3+0.1=2.095.

[0066] 3. Activation function: The ReLU function ReLU(x)=max(0,x) is adopted to introduce nonlinear transformation and enhance the network's ability to express complex features.

[0067] 4. Pooling Layer: A 1×2 max-pooling window is used to downsample the convolution results, preserving key features and reducing the amount of data. The pooling formula is as follows: For example, the above convolution result after pooling =max(2.0,2.095)=2.095.

[0068] 5. Fully Connected Layer: Linearly combines all pooling results and maps them to a low-dimensional space, as shown in the formula: ; in, These are the weights of the fully connected layer. is the bias term, and h is the output of CL-CNN (dimension 1), which is used as the input to MDN.

[0069] S4.2 Hybrid Density Network (MDN) Sub-model MDN is used to model the multimodal distribution of user interests (i.e., user interests may belong to multiple categories simultaneously, and the probabilities of different categories are different). It takes the output h of CL-CNN as input and outputs three types of parameters to describe the mixture of K Gaussian distributions: 1. Mixing coefficient : represents the weight of the k-th Gaussian distribution in the mixture model, satisfying Calculated by the softmax function: ; in, It is a linear transformation ( (Parameters to be learned).

[0070] 2. Mean The center position of the k-th Gaussian distribution is calculated by a linear transformation. ( (For parameters).

[0071] 3. Standard deviation The degree of dispersion of the k-th Gaussian distribution, calculated using an exponential function to ensure non-negativity: ( (For parameters).

[0072] The probability density function of MDN is: ; in, It is a Gaussian probability density function, and y is the user interest category label (discretely represented by a value from 1 to C, such as C=5 representing 5 interest categories).

[0073] S4.3 Model Training Process The goal of model training is to optimize parameters to make the prediction results as close as possible to the true labels. The specific steps are as follows: 1. Training data preparation: Input is The tags represent the user's actual interest categories (obtained through the interest tags filled in by the user; if the user does not fill in any tags, they are determined by the topic category of the content they publish, such as using a text classification model to classify the content into categories such as "data mining" and "artificial intelligence" to determine the user's interests).

[0074] 2. Parameter initialization: For CL-CNN and MDN We use random values ​​that follow a normal distribution for initialization to provide a starting point for model training.

[0075] 3. Loss function calculation: The negative log-likelihood loss function is used, and the formula is as follows: ; in, , , Let be the mixing coefficients, mean, and standard deviation of the k-th Gaussian distribution output by the i-th sample in the MDN. Let be the true interest label for the i-th sample. This loss function guides the model parameters towards optimal adjustment by measuring how well the predicted distribution matches the true label.

[0076] 2. Backpropagation and Parameter Update: The Adam optimizer is used for efficient parameter updates. The Adam optimizer combines the advantages of momentum and adaptive learning rates, dynamically adjusting the update step size by calculating the gradient of the loss function with respect to each parameter. Specifically, for each parameter in the model... (Including the convolutional kernel weights of CL-CNN) Bias Fully connected layer weights Bias (and all parameters of MDN), its update formula is: ; Where t is the number of iterations. For learning rate (e.g.) =0.001), It is a first-order momentum estimate that has been corrected for bias (reflecting the average trend of the gradient). It is a second-order momentum estimate that has been corrected for bias (reflecting the degree of dispersion of the gradient). It is a very small value (e.g., 10) -8 This is used to avoid a denominator of 0. Through backpropagation, the gradient is passed from the loss function layer by layer to the parameters of each layer, ensuring that each parameter can be adjusted according to its contribution to the loss.

[0077] 3. Iterative Training and Convergence Assessment: Divide the training data into batches (e.g., 32 samples per batch) and input them into the model for training. After each batch is completed, calculate and record the loss value. After each round (traversing all training samples), calculate the loss value on the validation set to determine if the model is overfitting. If the validation set loss increases for several consecutive rounds (e.g., 5 rounds), it indicates overfitting, and training should be terminated early. If both the training and validation set losses stabilize (e.g., the difference between losses in adjacent rounds is less than 10%), the model is considered overfitting. -4 If the number of iterations reaches a preset maximum value (such as 100 rounds), training will terminate and the current model parameters will be saved as the final training result.

[0078] S4.4 Model Application Process The trained CL-CNN-MDN model can be directly used for interest prediction of new samples. The specific process is as follows: 1. Input feature preparation: For the concepts in the new concept lattice Extract its feature vector The optimized feature vector is generated using the surface fitting formula in S3.1. .

[0079] 2. Feature transfer and prediction: Inputting a CL-CNN sub-model, which is then processed through convolutional, pooling, and fully connected layers, yields... ;Will Input the MDN sub-model and calculate the mixing coefficients of K Gaussian distributions. mean and standard deviation This leads to the probability density distribution of interest category y. .

[0080] 3. Result Generation: In a discrete set of interest categories (e.g., 1-5 corresponding to "Technology", "Entertainment", "Sports", "Education", and "Finance" respectively), calculate the probability value for each category and select the category with the highest probability as the final prediction result. For example, if K=3, the calculated probability of category 3 is 0.6, the probability of category 2 is 0.3, and the sum of the probabilities of the remaining categories is 0.1. Therefore, the predicted main interest of the user group corresponding to this concept is category 3 ("Sports").

[0081] S5: Analysis and Feedback of Social Network Data Mining Results To verify the effectiveness of the model and continuously optimize it, the prediction results need to be systematically analyzed and adjusted based on feedback, forming a closed loop of "training-evaluation-optimization": S5.1 Results Evaluation Indicators Model performance is quantified using multi-dimensional metrics: 1. Accuracy: The proportion of correctly predicted samples out of the total sample size. The formula is: ; Wherein, TP (True Positives) is the number of samples that actually belong to a certain category and are correctly predicted, TN (True Negatives) is the number of samples that actually do not belong to a certain category and are correctly predicted, FP (False Positives) is the number of samples that actually do not belong to a certain category but are incorrectly predicted, and FN (False Negatives) is the number of samples that actually belong to a certain category but are incorrectly predicted.

[0082] 2. Precision: The proportion of samples predicted to belong to a certain class that actually belong to that class. The formula is: ; 3. Recall: The proportion of samples that actually belong to a certain category and are correctly predicted. The formula is: ; 4. F1 score: The harmonic mean of precision and recall, used to comprehensively evaluate model performance. The formula is: .

[0083] S5.2 Feedback Optimization Mechanism Based on the evaluation results, adjust the model or preprocessing procedure accordingly: 1. Model Structure Optimization: If the overall accuracy is lower than a preset threshold (e.g., 0.8), it may be due to insufficient feature representation or inadequate model complexity. In this case, the order of the polynomial in the surface fitting can be increased (e.g., from order 2 to order 3), and the number of convolutional kernels in CL-CNN can be increased (e.g., ...). Increase the number of Gaussian components K in the MDN (e.g., from 8 to 16), or adjust the number of Gaussian components K in the MDN (e.g., from 3 to 5) to enhance the model's ability to capture complex patterns.

[0084] 2. Sample and Weight Adjustment: If the recall rate of a certain interest category is low (e.g., the recall rate of "education" is <0.6), it indicates that the model has not learned enough of this type of feature. This can be optimized in two ways: first, increase the number of training samples corresponding to this interest category (e.g., crawl more user content containing the keyword "education" from social networks); second, add weights to this type of sample in the loss function (e.g., multiply the loss value by 1.5 when y_i = "education") to increase the model's attention to this category.

[0085] 3. Feature Engineering Optimization: If the features after surface fitting do not significantly improve model performance, the feature selection may be inappropriate. The feature vectors can be redefined. Dimensions (such as adding user interaction frequency features), or adjusting the input dimensions of the surface fitting (such as using...) and For x and y, fit a weighted sum of |Ext(l)| and |Int(l)|.

[0086] S5.3 Optimization Effect Verification Taking user interest prediction on a social network platform as an example, the accuracy of the initial model (without surface fitting and MDN) was 0.72. After adding surface fitting, the accuracy improved to 0.79, and after incorporating MDN, it further improved to 0.88. Specifically, the F1 score for the "data mining" interest category improved from 0.68 (initial) → 0.75 (with surface fitting) → 0.85 (with MDN), indicating that surface fitting effectively captured the nonlinear relationships of features, while MDN improved the modeling ability for multimodal interest distributions. The final optimized model can be used for personalized content recommendation, such as pushing relevant technical articles to users predicted to have a "data mining" interest, resulting in a 35% increase in click-through rate compared to random recommendations.

[0087] The following examples illustrate the application of social network data mining methods based on concept lattice convolutional neural networks (including surface fitting and hybrid density network optimization) in specific scenarios, fully demonstrating the actual operation process and effects of the methods.

[0088] Application scenario: User interest clustering and academic activity recommendation on campus social platforms A campus social platform has 5,000 users (N=5,000). Users mainly post content such as academic discussions, campus activities, and life sharing. The platform hopes to identify user interest clusters through data mining and recommend matching academic lectures, seminars, and other activities to users.

[0089] Step 1: Social Network Data Collection and Preprocessing Collect raw data from the past 3 months via the platform API: User basic information: includes (Student ID) (Ages 18-25) (Gender). After cleaning and deduplication, 4800 valid records were retained, with standardized age. (e.g., 20-year-old users) =(20-18) / (25-18)=0.29).

[0090] User-submitted content: A total of 12,000 entries (M=12,000), after deduplication, 10,500 entries were retained. (Content text) Examples include "Applications of Machine Learning in Image Recognition" and "Announcement of Next Week's Artificial Intelligence Lecture".

[0091] User interaction information: including likes ( =1), Comments ( =2), forwarding ( =3), a total of 35,000 interaction records (P=35,000), of which 32,000 were retained after cleaning.

[0092] Step 2: Concept Lattice Construction Keyword extraction: For 10,500 text entries, extract the top 5 keywords using TF-IDF (K=5). For example, the keywords for the text "Application of machine learning in image recognition" are "machine learning", "image recognition", "application", "algorithm", and "model" (the 5 with the highest TF-IDF values).

[0093] User attribute set: for each user Build , such as users Three pieces of content related to machine learning were published. ={Machine Learning, Image Recognition, Algorithms, Models, Deep Learning}.

[0094] Formal Context and Concept Lattice: Construct (U, A, R), where A is the union of all keywords (800 keywords in total). Use the NextClosure algorithm to generate the concept lattice L, resulting in 120 concepts, such as the concept... ,in (30 users) ={machine learning, algorithm, model} (these are common keywords among users).

[0095] Step 3: Feature Extraction and Surface Fitting Basic feature vectors: for concepts , =30, =3, =2 (2 parent concepts) =4 (4 sub-concepts), therefore =(30,3,2,4).

[0096] Surface fitting: Assume a second-order polynomial fitting, with x=|Ext(l)| and y=|Int(l)| as inputs. For the labels. Solve for the coefficients. .right Given x=30, y=3, the fitted expression is: =0.05+0.3+0.27+6+7.2+9=22.82, therefore .

[0097] Step 4: Training and Application of CL-CNN-MDN Model Training data: 120 concepts The input is the user's actual interest tags (obtained from the "Research Direction" field filled in by the user, such as "Artificial Intelligence", "Computer Vision", etc., a total of 5 categories) and the output is the user's actual interest tags (obtained from the "Research Direction" field filled in by the user, such as "Artificial Intelligence", "Computer Vision", etc., a total of 5 categories).

[0098] CL-CNN Training: =8 1×2 convolutional kernels, after convolution, ReLU activation, and pooling, the fully connected layer outputs h. For example =(30,3,22.82) is obtained after processing. =5.2.

[0099] MDN training: Assume K = 3 Gaussian components, and obtain the parameters after training. For example... =5.2 corresponding =0.6, =1 (corresponding to "artificial intelligence"), =0.3; =0.3, =2 (corresponding to "computer vision"); =0.1, =3 (corresponding to "Data Science").

[0100] Prediction results: ; Therefore, it is predicted The corresponding user group's interest is "artificial intelligence".

[0101] Step 5: Results Analysis and Feedback Optimization Evaluation metrics: Model accuracy was 0.89, precision was 0.87, recall was 0.86, and F1 score was 0.86 for the "Artificial Intelligence" interest category.

[0102] Feedback optimization: The recall rate of the "Data Science" category was found to be low (0.72). After increasing the weight of this category of samples (multiplying the loss function by 1.5), the recall rate was improved to 0.81.

[0103] Application results: Recommending "Frontiers in Deep Learning" lectures to users interested in "artificial intelligence" resulted in a participation rate of 42%, an increase of 28% compared to before optimization.

[0104] Through the above application examples, this method fully realizes the entire process from data collection to interest prediction. The fusion of surface fitting and MDN significantly improves the mining accuracy, verifying the effectiveness of the method.

[0105] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0106] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0107] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0108] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0109] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0110] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

[0111] Furthermore, those skilled in the art will understand that aspects of this application can be described and illustrated through several patentable types or situations, including any new and useful combination of processes, machines, products, or substances, or any new and useful improvements thereof. Therefore, aspects of this application can be implemented entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. All of the above hardware or software can be referred to as a “unit,” “module,” or “system.” Furthermore, aspects of this application can take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0112] The computer program code required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, and Python; general programming languages ​​such as C; Visual Basic, Fortran2103, Perl, COBOL2102, PHP, and ABAP; dynamic programming languages ​​such as Python, Ruby, and Groovy; or other programming languages. This program code can run entirely on the user's computer, or as a standalone software package on the user's computer, or partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any network, such as a local area network (LAN) or wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as Software as a Service (SaaS).

[0113] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this application are not intended to limit the order of the processes and methods of this application. Although some currently considered useful embodiments of the invention have been discussed in the foregoing disclosure by way of various examples, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the substance and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a purely software solution, such as an installation on an existing server or mobile device.

[0114] Similarly, it should be noted that, in order to simplify the description of the present application and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments of the present application sometimes combines multiple features into a single embodiment, drawing, or description thereof. However, this approach of the present application should not be construed as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject of the invention should possess fewer features than in any single embodiment described above.

Claims

1. A method for social network data mining based on concept lattice convolutional neural networks, characterized in that, include: S1: Social network data collection and preprocessing; Raw social network data is obtained through the application programming interface of the social network platform. The raw social network data includes basic user information, content information posted by users, and interaction information between users. The raw social network data is cleaned, deduplicated, and standardized to obtain preprocessed social network data. The preprocessed social network data includes a preprocessed set of basic user information, a preprocessed set of content information posted by users, and a preprocessed set of interaction information between users. S2: Concept lattice construction; Based on the preprocessed user basic information set, preprocessed user-published content information set, and preprocessed user interaction information set obtained in S1, a concept lattice is constructed. Each concept in the concept lattice consists of extension and intension. The extension is the set of users included in the concept, and the intension is the set of keywords and interaction features common to these users. The construction process includes extracting keywords from the user-published content text, calculating interaction features by combining user interaction information, constructing formal background, and then using the NextClosure algorithm to generate the concept lattice. S3: Concept lattice feature extraction and surface fitting optimization; Features are extracted from the concept lattice constructed in S2 to form a concept lattice feature vector. The concept lattice feature vector includes the extension size, connotation size, number of parent concepts, number of child concepts, and interaction feature statistics of the concept. The concept lattice feature vector is optimized by using binary polynomial surface fitting to obtain the fitted feature vector. S4: Construct a fusion model of concept lattice convolutional neural network and hybrid density network; the fusion model takes the fitted feature vector obtained in S3 as input and includes a concept lattice convolutional neural network sub-model and a hybrid density network sub-model. The concept lattice convolutional neural network sub-model is used to extract deep local features, and the hybrid density network sub-model is used to model the multi-peak distribution of user interests. The optimized fusion model is obtained through training, and the optimized fusion model is used to predict user interests. S5: Analysis and feedback of social network data mining results; compare the user interest prediction results obtained in S4 with the actual results, calculate the evaluation index, optimize and adjust the model or preprocessing process according to the evaluation index results until the model performance meets the requirements, and output the final social network data mining results.

2. The social network data mining method based on concept lattice convolutional neural networks according to claim 1, characterized in that, In S1, the basic user information includes a unique user identifier, user age, and user gender; the content information published by the user includes a unique content identifier, content text, and publication time; and the interaction information between users includes a unique interaction identifier, interaction type, unique identifier of the user initiating the interaction, unique identifier of the user receiving the interaction, and interaction time. The cleaning process involves deleting records containing missing values; the deduplication process involves deleting duplicate user-posted content information; and the standardization process involves standardizing the user's age to the [0,1] interval. The standardization formula is that the standardized user age equals (user age minus the minimum of all user ages) divided by (the maximum of all user ages minus the minimum of all user ages).

3. The social network data mining method based on concept lattice convolutional neural networks according to claim 1, characterized in that, In S2, keyword extraction for user-published content text uses the TF-IDF algorithm, which includes calculating word frequency, calculating inverse document frequency, calculating TF-IDF value, and selecting the top K keywords with the highest TF-IDF value in each content text. The formal background is a triple (U, A, R), where U is the user set, A is the union of the attribute sets of all users, and R is the set of relationships between users and attributes. The process of generating concept lattices using the NextClosure algorithm is as follows: initialization, generating new connotations and judging closure, and repeating the generation process until no new concepts can be generated.

4. The social network data mining method based on concept lattice convolutional neural networks according to claim 1, characterized in that, In S3, feature extraction from the concept lattice includes calculating the extension and intension of each concept, defining the parent-child relationship between concepts, and recording the number of parent and child concepts for each concept; the feature vector of the concept lattice is (extension, intension, number of parent concepts, number of child concepts); the polynomial order of the bivariate polynomial surface fitting is 2, and the fitting formula is that the fitted feature value is equal to the sum of the products of multiple surface fitting coefficients and the powers of the corresponding variables. The surface fitting coefficients are solved by the least squares method to generate the fitted feature vector.

5. The social network data mining method based on concept lattice convolutional neural networks according to claim 1, characterized in that, In S4, the concept lattice convolutional neural network sub-model includes an input layer, convolutional layers, activation functions, pooling layers, and fully connected layers. The input layer receives the fitted feature vectors. The convolutional layers use multiple 1×2 convolutional kernels to perform convolution operations on the input features. The activation function is the ReLU function. The pooling layers use max pooling. The fully connected layers map the pooling results to a low-dimensional space and output feature values.

6. The social network data mining method based on concept lattice convolutional neural networks according to claim 1, characterized in that, In S4, the hybrid density network sub-model takes the feature values ​​output by the concept lattice convolutional neural network sub-model as input, including three types of parameters: mixing coefficients, mean, and standard deviation. The mixing coefficients are calculated using the softmax function, the mean is calculated using a linear transformation, and the standard deviation is calculated using an exponential function. The probability density function of the hybrid density network sub-model is the sum of the products of multiple Gaussian probability density functions and their corresponding mixing coefficients.

7. The social network data mining method based on concept lattice convolutional neural networks according to claim 1, characterized in that, In S4, the training process of the fusion model includes preparing training data, with the fitted feature vector as input and the user's actual interest category as label; initializing model parameters; calculating the loss function, which adopts the negative log-likelihood loss function; performing backpropagation through the Adam optimizer to update the model parameters; and iteratively training until the loss function converges or the preset maximum number of iterations is reached.

8. The social network data mining method based on concept lattice convolutional neural networks according to claim 1, characterized in that, In S4, the process of predicting user interests using the optimized fusion model involves inputting the fitted feature vector of the new concept into the concept lattice convolutional neural network sub-model to obtain feature values, inputting the feature values ​​into the hybrid density network sub-model to calculate the probability density distribution of user interest categories, and selecting the interest category with the highest probability as the prediction result.

9. The social network data mining method based on concept lattice convolutional neural networks according to claim 1, characterized in that, In S5, the evaluation metrics include accuracy, precision, recall, and F1 score. Accuracy is the ratio of correctly predicted samples to the total number of samples. Precision is the proportion of samples predicted as positive that are actually positive. Recall is the proportion of samples that are actually positive that are correctly predicted as positive. The F1 score is the harmonic mean of precision and recall.

10. The social network data mining method based on concept lattice convolutional neural networks according to claim 1, characterized in that, In S5, optimizations and adjustments to the model or preprocessing process include adjusting the polynomial order of the surface fitting, adjusting the number of Gaussian components in the mixed density network, increasing the number of training samples, adjusting the weights of the training samples, and redefining the dimension of the concept lattice feature vector.

Images