Intelligent matching system based on big data resources

Abstract

The invention discloses an intelligent matching system based on big data resources, and the system comprises a data acquisition module which collects multi-source matching data, carries out the fusion processing, and obtains user behavior characteristics and resource behavior characteristics from the processed multi-source matching data; the feature adaptation module is used for acquiring a user behavior feature matrix and a resource behavior feature matrix, and inputting the user behavior feature matrix and the resource behavior feature matrix by constructing an improved space-time convolutional network model to obtain an optimal preliminary matching factor; and the intelligent matching module is used for labeling high-quality matching tags, taking the high-quality matching tags as a labeling data set of a parameter calculation model based on a neural network architecture, obtaining optimal adjustment parameters through the parameter calculation model, making an optimal decision according to the optimal adjustment parameters, and providing a fine decision basis for intelligent matching of big data.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing, and particularly to an intelligent matching system based on big data resources. Background Art

[0002] At present, the development of big data technology has brought new benefits to many fields. For example, in many scenarios involving user-resource interaction, such as product recommendation on e-commerce platforms, course matching in online education, and talent-position docking in the job recruitment market, achieving efficient and accurate user-resource intelligent matching has become a key requirement. However, existing matching technologies have many limitations in practical applications;

[0003] Currently, in existing resource matching systems, the behaviors of users and the supply situations of resources change dynamically over time and space. For example, in the online education scenario, the learning time and location of users may change, and the opening time and location of courses will also be adjusted. However, existing technologies often adopt static matching strategies and cannot be adjusted in a timely manner according to these dynamic changes, greatly reducing the timeliness and accuracy of matching;

[0004] Moreover, the associations between users and resources may vary in different spatio-temporal local regions. For example, in a specific time period or specific geographical location, users' preferences for certain resources may be more obvious. However, existing matching models generally lack in-depth exploration of such local associations and cannot provide users with resource recommendations that are more in line with their current needs. At the same time, the behaviors of users and the supply of resources often have periodic patterns, but traditional methods have not fully utilized these periodic characteristics, resulting in the inability to push the most relevant resources to users at the right time, reducing the quality and efficiency of matching;

[0005] Therefore, an intelligent matching system based on big data resources is proposed herein. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art and achieve the above object, the present invention proposes the following technical solutions:

[0007] An intelligent matching system based on big data resources, comprising:

[0008] A data acquisition module: collecting multi-source matching data and performing fusion processing, and obtaining user behavior characteristics and resource behavior characteristics from the processed multi-source matching data;

[0009] A feature adaptation module: obtaining a user behavior feature matrix and a resource behavior feature matrix, and obtaining an optimal preliminary matching factor by constructing an improved spatio-temporal convolutional network model and inputting the user behavior feature matrix and the resource behavior feature matrix;

[0010] Intelligent matching module: Mark high-quality matching tags, use the high-quality matching tags as the labeled data set of a parameter calculation model based on the neural network architecture, obtain the optimal adjustment parameters through the parameter calculation model, dynamically adjust the optimal preliminary matching factor through the optimal adjustment parameters, obtain the final associated matching factor, and make the optimal user-resource matching decision according to the associated matching factor;

[0011] The improved spatio-temporal convolutional network model is obtained by adding a custom layer to the basic spatio-temporal convolutional network model;

[0012] The custom layer includes a spatio-temporal local association enhancement layer and a temporal periodic feature capture layer.

[0013] The multi-source matching data includes structured matching data, unstructured matching data, and spatio-temporal matching data;

[0014] The user behavior feature matrix is obtained by directly representing the user behavior features in matrix form;

[0015] The resource behavior feature matrix is obtained by directly representing the resource behavior features in matrix form.

[0016] The improved spatio-temporal convolutional network model includes an input layer, a basic spatio-temporal convolutional layer, a spatio-temporal local association enhancement layer, a temporal periodic feature capture layer, and a fully connected layer;

[0017] The input layer receives the user behavior feature matrix and the resource behavior feature matrix;

[0018] The basic spatio-temporal convolutional layer outputs a basic feature representation according to the user behavior feature matrix and the resource behavior feature matrix;

[0019] The spatio-temporal local association enhancement layer outputs strong local association features based on the basic feature representation;

[0020] The temporal periodic feature capture layer outputs spatio-temporal periodic features based on the strong local association features;

[0021] The fully connected layer outputs the optimal preliminary matching factor based on the spatio-temporal periodic features.

[0022] The process of the spatio-temporal local association enhancement layer outputting strong local association features is as follows:

[0023] Divide the basic feature representation into local spatio-temporal blocks through a fixed window to obtain user behavior local feature blocks and resource behavior local feature blocks;

[0024] Obtain a set of user behavior local feature blocks and a set of resource behavior local feature blocks based on the user behavior local feature blocks and the resource behavior local feature blocks;

[0025] Combine the user behavior local feature blocks and the resource behavior local feature blocks to form a local spatio-temporal block;

[0026] Extract local user behavior features and local resource behavior features from the local spatio-temporal block, and calculate the correlation weight scores;

[0027] Normalize all the correlation weight scores to obtain a weight matrix;

[0028] Based on the weight matrix, perform weighted fusion on the local features to obtain an enhanced local feature representation for each local spatio-temporal block;

[0029] Concatenate the enhanced features of all local spatio-temporal blocks, and the spatio-temporal local correlation enhancement layer finally outputs strong local correlation features.

[0030] The process by which the temporal cycle feature capture layer outputs spatio-temporal periodic features is as follows:

[0031] Extract user time features and resource time features from the strong local correlation features;

[0032] Based on the user time features and resource time features, obtain user time feature segments and resource time feature segments through a sliding window;

[0033] For the user time feature segments and resource time feature segments within each sliding window, perform feature weight allocation through a self-attention mechanism to obtain attention weights;

[0034] Based on the attention weights, obtain weighted user time feature segments and weighted resource time feature segments;

[0035] Input the weighted user time feature segments and weighted resource time feature segments into a recurrent neural network LSTM, and select the output of the last time step as the periodic features of user behavior and the periodic features of resource behavior;

[0036] Concatenate the periodic features of user behavior and the periodic features of resource behavior to obtain spatio-temporal periodic features.

[0037] The specific acquisition process by which the fully connected layer outputs the optimal preliminary matching factor is as follows:

[0038] Use the Flatten function to flatten the input spatio-temporal periodic features into a one-dimensional vector to obtain a flattened feature vector;

[0039] Through bilinear transformation and activation operation on the flattened feature vector, obtain the optimal preliminary matching factor.

[0040] The acquisition process of the correlation matching factor is as follows:

[0041] Obtain the mean and standard deviation of the preliminary matching factors of the current batch, and dynamically adjust the preliminary matching factors based on the optimal adjustment parameters;

[0042] Normalize the dynamic adjustment result through the Softmax function and output the associated matching factors.

[0043] The process of obtaining the optimal adjustment parameters is as follows:

[0044] Construct a parameter calculation model;

[0045] Label high-quality matching labels for historical matching cases, and construct an annotation dataset based on the high-quality matching labels;

[0046] The parameter calculation model includes an input layer, a hidden layer, and an output layer;

[0047] The input layer receives the annotation dataset and inputs the features in the annotation dataset into the network;

[0048] The hidden layer performs weighted calculation and activation processing on the input features by neurons;

[0049] The output layer integrates the information processed by the hidden layer and outputs the optimal adjustment parameters through linear transformation.

[0050] The hidden layer of the parameter calculation model includes l hidden layers, each containing n l neurons, and the number of neurons in the output layer of the parameter calculation model is one.

[0051] The annotation rules for the high-quality labels are as follows:

[0052] Count the number of times a user actually selects a certain resource in historical matching cases. When the number of times exceeds the preset threshold, it is labeled as a high-quality match; when it is less than or equal to the preset threshold, it is labeled as a non-high-quality match. Based on this rule, construct an annotation dataset to convert user behavior into labels recognizable by the model.

[0053] The present invention has the following beneficial effects:

[0054] In the present invention, first, through the collection and fusion of multi-source matching data, user behavior characteristics and resource behavior characteristics are comprehensively obtained, and by improving the spatio-temporal local correlation enhancement layer and the temporal periodic feature capture layer in the spatio-temporal convolutional network model, the dynamic correlation and periodic features between users and resources in the spatio-temporal dimension are effectively mined, making the matching result more timely and adaptable;

[0055] Secondly, a labeled dataset is constructed by labeling high-quality matching tags to train the parameter calculation model, enabling the parameter calculation model to learn the characteristic patterns of historical high-quality matches, ensuring that the obtained optimal adjustment parameters are more in line with the matching logic. At the same time, the preliminary matching factors are dynamically processed through the optimal adjustment parameters, which can flexibly handle the data distribution differences in different scenarios, making the adjustment process adapt to the real-time data characteristics and enhancing the dynamic response ability of the matching factors to complex scenarios;

[0056] Finally, a decision is made based on the final associated matching factors to accurately quantify the matching degree between the user and the resource, making the matching decision more scientific and reasonable, effectively improving the accuracy and efficiency of user-resource matching, and achieving the efficient allocation of high-quality resources. Brief Description of the Drawings

[0057] Figure 1 It is a system block diagram of an intelligent matching system based on big data resources proposed by the present invention. Detailed Embodiments

[0058] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0059] Embodiment 1

[0060] As Figure 1 shown, an intelligent matching system based on big data resources proposed by the present invention includes:

[0061] Data acquisition module: Collect multi-source matching data and perform fusion processing, and obtain user behavior characteristics and resource behavior characteristics from the processed multi-source matching data;

[0062] The multi-source matching data includes structured matching data, unstructured matching data, and spatio-temporal matching data; let the structured matching data be S, the vectorized representation of the unstructured matching data be N, and the spatio-temporal matching data be M;

[0063] Collection of structured matching data S: For structured matching data, including the basic information of users (age, gender, occupation, etc.) and the attribute information of resources (product category, specification parameters, etc.), it is directly read and collected from the database;

[0064] Collection of unstructured matching data N: The unstructured matching data includes user comment data and resource description text data, and relevant text content is directly crawled from the web based on web scraping technology;

[0065] Spatiotemporal Matching Data M Collection: The spatiotemporal matching data includes the user's GPS trajectory data and the geographical location data of resources. The user's GPS trajectory is obtained through the positioning function on the mobile device, and the geographical location data of resources is obtained from the API of the map service provider;

[0066] Fuse the collected structured, unstructured, and spatiotemporal matching data. Specifically:

[0067] For structured matching data, merge the data in different tables through the primary keys (user ID, resource ID). For unstructured matching data, use natural language processing technology for vector representation. For example, use word embedding technology to convert the text into vector form. For spatiotemporal data, encode information such as longitude, latitude, and timestamp to form a unified feature vector;

[0068] Fuse the three types of data into the processed multi-source matching data F through vector concatenation operation. The fusion formula is:

[0069]

[0070] where represents the vector concatenation operation;

[0071] Extract the features related to user behavior from the fused multi-source matching data F as user behavior features U, and extract the features related to resource behavior as resource behavior features R;

[0072] Specifically, the features related to user behavior (user behavior features U) are reflected in the multi-source matching data F as follows:

[0073] In the structured matching data, it is reflected in the user's basic information, including age, gender, occupation, etc. These information can be used to analyze the user's behavior preferences, demand tendencies and other behavior characteristics. In the unstructured matching data, it is reflected in the user's comment data. By analyzing the comment content, we can understand the user's evaluation of the resource, usage experience, demand feedback and other behavior information. In the spatiotemporal matching data, it is reflected in the user's GPS trajectory data, which is used to analyze the user's activity range, action path, frequently visited locations and other behavior trajectory characteristics;

[0074] The features related to resource behavior (resource behavior features R) are reflected in the multi-source matching data F as follows:

[0075] The attribute information reflected as resources in structured matching data, such as product categories, specification parameters, etc., which reflect the basic characteristics and functions of resources, belongs to the category of behavioral characteristics of resources. The descriptive text data reflected as resources in unstructured matching data can reflect the characteristics, advantages, applicable scenarios and other behavioral attributes of resources through the text content. The geographical location data reflected as resources in spatio-temporal matching data is used to analyze the distribution characteristics, coverage range and other behavioral attributes of resources (such as the layout strategy of resources in different regions, the service radiation range).

[0076] Feature adaptation module: Obtain the user behavior feature matrix and the resource behavior feature matrix, and obtain the optimal preliminary matching factor by constructing an improved spatio-temporal convolutional network model and inputting the user behavior feature matrix and the resource behavior feature matrix;

[0077] Directly represent the user behavior features in a matrix form to obtain the user behavior feature matrix

[0078] Directly represent the resource behavior features in a matrix form to obtain the resource behavior feature matrix

[0079] Construct an improved spatio-temporal convolutional network model. The improved spatio-temporal convolutional network model consists of a basic spatio-temporal convolutional network model and a custom layer. The custom layer includes a spatio-temporal local association enhancement layer and a temporal periodic feature capture layer;

[0080] The basic spatio-temporal convolutional network model is specifically an input layer, a basic spatio-temporal convolutional layer and a fully connected layer;

[0081] The improved spatio-temporal convolutional network model is specifically an input layer, a basic spatio-temporal convolutional layer, a spatio-temporal local association enhancement layer, a temporal periodic feature capture layer and a fully connected layer;

[0082] Specifically, receive the user behavior feature matrix and the resource behavior feature matrix through the input layer, output the basic feature representation through the basic spatio-temporal convolutional layer according to the user behavior feature matrix and the resource behavior feature matrix, output the strong local association feature based on the basic feature representation through the spatio-temporal local association enhancement layer, output the spatio-temporal periodic feature based on the strong local association feature through the temporal periodic feature capture layer, and output the optimal preliminary matching factor based on the spatio-temporal periodic feature through the fully connected layer;

[0083] The input layer receives the input data, the user behavior feature matrix and the resource behavior feature matrix

[0084] The process of the basic spatio-temporal convolutional layer outputting the basic feature representation is as follows:

[0085] The basic spatio-temporal convolutional layer performs basic feature extraction on the input user behavior feature matrix and resource behavior feature matrix, capturing the basic feature representation of the multi-source matching data F in the spatio-temporal dimension. The specific process is as follows:

[0086] Adopt spatio-temporal convolution operation and use a two-dimensional convolution kernel Slide in the time and space dimensions and perform temporal convolution on the input user behavior feature matrix and resource behavior feature matrix Perform temporal convolution in the time dimension and spatial feature aggregation in the space dimension, initially generating a basic feature representation containing spatio-temporal information

[0087] The k in t represents the size of the convolution kernel in the time dimension, and k s represents the size of the convolution kernel in the space dimension;

[0088] The implementation process of the spatio-temporal local correlation enhancement layer is as follows:

[0089] For the basic feature representation Divide the local spatio-temporal blocks through a fixed window. Divide the fixed window into 3 consecutive time units. Among them, the fixed window includes a time window and a spatio-temporal window, and the formula is expressed as:

[0090]

[0091] Among them, represents the local user behavior feature block, represents the local resource behavior feature block. For the divided local user behavior feature block and the local resource behavior feature block Within the time range t and space range s, generate the set of local user behavior feature blocks and the set of local resource behavior feature blocks The set of local user behavior feature blocks and the set of local resource behavior feature blocks form a local spatio-temporal block H, and its dimension is d, where d = t × s;

[0092] Specifically, the time range t refers to an interval set in the time dimension when dividing the basic feature representation, used to intercept the feature data within a specific time period, and the space range s is a region set in the space dimension, used to limit the feature data within a specific space region;

[0093] After obtaining the local spatio-temporal block H, extract the local user behavior feature U local and the local resource behavior feature Rlocal ;

[0094] Specifically, it is known that the local spatio-temporal block H is composed of a set of local user behavior feature blocks and a set of local resource behavior feature blocks , and its dimension is d = t × s. Since and respectively contain the feature data of the user and the resource within a specific spatio-temporal range, they can be extracted according to their dimensional positions and indices in H. For example:

[0095] Suppose is in the first half part of H, is in the second half part of H. H is represented in the form of a one-dimensional vector (flattening the original spatio-temporal block data). By calculating the dimensional ratio to determine the splitting point, the first elements belong to and the subsequent elements belong to Then:

[0096] The local user behavior feature U local is expressed as the set of elements from the starting position of H to the position, that is,

[0097]

[0098] The local resource behavior feature R local is expressed as the set of elements from the position of H to the end position, that is,

[0099]

[0100] To measure the correlation degree between the local user behavior feature and the local resource behavior feature within the local spatio-temporal block, calculate the correlation weight score:

[0101]

[0102] where, (R local ) * represents the transpose of the local resource behavior feature R local , represents the correlation weight score;

[0103] Specifically, the correlation weight score is used to measure the correlation degree between the local user behavior feature and the local resource behavior feature within the local spatio-temporal block;

[0104] Normalize all the correlation weight scores to obtain a weight matrix

[0105] Based on the weight matrix The enhanced local feature representation X of each local spatio-temporal block H is obtained by weighted fusion of local features, and the formula is expressed as:

[0106]

[0107] After fusion, the enhanced features X of all local spatio-temporal blocks H are concatenated, and the spatio-temporal local correlation enhancement layer finally outputs strong local correlation features

[0108] Specifically, this way of feature fusion and concatenation can integrate the correlation information within each local spatio-temporal block to form a more comprehensive and representative spatio-temporal feature representation. Through the spatio-temporal local correlation enhancement layer, the correlation degree between user behavior features and resource behavior features within the same spatio-temporal local region is enhanced, enabling the model to capture the dynamic evolution details within the local spatio-temporal range more precisely;

[0109] The implementation process of the time series period feature capture layer is as follows:

[0110] Taking the output of the spatio-temporal local correlation enhancement layer (strong local correlation features containing strong local correlation information ) as the input of the time series period feature capture layer, extracting the user time feature U related to user behavior t and the resource time feature R related to resource behavior t ;

[0111] Using a sliding window to process the user time feature U t and the resource time feature R t , and regarding the user time feature U t and the resource time feature R t within each window as a time series segment, denoted as the user time feature segment U window and the resource time feature segment R window ;

[0112] For the user time feature segment and the resource time feature segment within each sliding window, performing feature weight allocation through a self-attention mechanism;

[0113] Mapping them to three different spaces respectively to obtain the query vector Q U , key vector K U , value vector V U (for user behavior features), and query vector Q R , key vector K R , value vector V R (for resource behavior features), and then normalizing through the softmax function to obtain the attention weights α U and αR ;

[0114] Calculate the weighted user time feature segment U based on the attention weights window and the resource time feature segment R window It is represented that the weighted representation of the user time feature segment is:[[]]

[0115]

[0116] where U attended represents the weighted user time feature segment;

[0117] The resource time feature segment R attended The weighted representation of the feature segment is:[[]]

[0118]

[0119] where R attended represents the weighted resource time feature segment;

[0120] Input the weighted user time feature segment U attended and the weighted resource time feature segment R attended processed by the self-attention mechanism into a recurrent neural network LSTM respectively. Let the LSTM cell be LSTM U and LSTM R , Extract the periodic features from the output of the recurrent neural network, which is represented as:[[]]

[0121] U cycle = LSTM U (U attended )

[0122] R cycle = LSTM R (R attended )

[0123] Specifically, the recurrent neural network can effectively process time series data and capture long-term dependencies and periodic patterns in the time series through the update of hidden states;

[0124] Select the output of the last time step as the periodic feature U of the user behavior cycle and the periodic feature R of the resource behavior cycle , Concatenate the periodic feature U of the user behavior cycle and the periodic feature R of the resource behavior cycle to finally output the spatio-temporal periodic feature containing the time series periodic feature

[0125] Specifically, when dealing with information in the time dimension, traditional spatio-temporal convolutional networks focus on capturing continuous time series changes and are insufficient in extracting periodic features. This time series periodic feature capture layer extracts periodic features in the time dimension through specially designed steps to assist the model in more accurately modeling dynamic evolution relationships;

[0126] The process of the fully connected layer outputting the optimal preliminary matching factor is as follows:

[0127] For spatio-temporal periodic features perform a bilinear transformation and activation to obtain the optimal preliminary matching factor. The specific process is as follows:

[0128] Through (Flatten is a flattening function) flatten the input spatio-temporal periodic features into a one-dimensional vector to obtain a flattened feature vector;

[0129] The first linear transformation: Multiply the flattened feature vector by an auxiliary weight matrix W2 (with dimension p, where p is the dimension of the spatio-temporal periodic features after flattening), and add a bias b2, which is expressed as: of the dimension, ) and the flattened feature vector, and add a bias b2, which is expressed as:

[0130]

[0131] The second linear transformation: After activation through the tanh function, multiply it by a main weight matrix W1 (with dimension m, where m is the output dimension of the matching score) and add a bias b1 to obtain the optimal preliminary matching factor Z. The formula is expressed as:

[0132] Specifically, the optimal preliminary matching factor Z is the value reflecting the preliminary matching degree between the user and the resource output after the fully connected layer performs a bilinear transformation and activation operation on the input spatio-temporal periodic features It is the optimal intermediate result of the model processing the correlation relationship between user behavior features and resource behavior features, providing the best data basis for generating the final correlation matching factor.

[0133] Intelligent matching module: Mark high-quality matching labels, use the high-quality matching labels as the labeled data set of a parameter calculation model based on a neural network architecture, obtain the optimal adjustment parameters through the parameter calculation model, dynamically adjust the optimal preliminary matching factor through the optimal adjustment parameters, obtain the final correlation matching factor, and make the optimal user-resource matching decision according to the correlation matching factor;

[0134] The implementation process of dynamic adjustment is as follows:

[0135] Introduce the optimal adjustment parameter γ to adjust the preliminary matching factor Z, and calculate the mean value Z μ and standard deviation Zσ and dynamically adjust the preliminary matching factor with the optimal adjustment parameter γ, which is expressed by the formula:

[0136]

[0137] where ∈ represents a minimum value to prevent the denominator from being zero;

[0138] Normalize the dynamic adjustment result through the Softmax function and output the associated matching factor F match ;

[0139] The process of obtaining the optimal adjustment parameter γ is as follows:

[0140] Obtain the optimal adjustment parameter γ through a parameter calculation model;

[0141] The process of constructing the parameter calculation model is as follows:

[0142] Annotate high-quality matching labels for historical matching cases. The high-quality matching labels are based on the statistical times of the user's actual selection of a certain resource. If it exceeds the preset statistical times, it is marked as a high-quality match;

[0143] Specifically, the high-quality matching label is used as the supervision signal for model training, and the annotation rule is:

[0144] Statistically count the number of times the user actually matches and selects a certain resource. If the number of times exceeds the preset threshold U (for example, the preset statistical times is U, when the number of times the user selects a certain resource ≥ U), it is marked as a high-quality match (label value is 1), otherwise (less than or equal to) it is marked as a non-high-quality match (label value is 0). Based on this rule, construct the annotation dataset D to convert the user's behavior into labels recognizable by the model;

[0145] Construct the annotation dataset D based on the high-quality matching labels,

[0146] The parameter calculation model is based on a neural network architecture, including an input layer, a hidden layer, and an output layer;

[0147] Input layer: Receive the annotation dataset D and input the features (historical matching statistics, current matching features) in the annotation dataset into the network as the original information for model learning;

[0148] Hidden layer: Design l hidden layers, each layer contains n l neurons. The hidden layer performs weighted calculation and activation processing (ReLU activation function) on the input features through neurons, extracts the deep correlation information in the features layer by layer, and gradually abstracts the complex patterns related to the optimal adjustment parameter γ. The output process of the l-th hidden layer is expressed as:

[0149] H l = W l Hl-1 +b l

[0150] wherein, H l is the output of the l-th hidden layer, H l-1 is the output of the layer before the l-th layer, W l is the weight of the l-th hidden layer, and b l is the bias vector of the l-th hidden layer;

[0151] Output layer: It only contains one neuron. By integrating the information processed by the hidden layer, it outputs the optimal adjustment parameter γ through linear transformation, completing the mapping from input features to target parameters;

[0152] Example: In a recruitment resource matching scenario, during the peak recruitment season, the optimal adjustment parameter dynamically adjusts the correlation matching factor to expand the screening range of the matching criteria. During the off-peak recruitment season, the optimal adjustment parameter γ increases the correlation matching factor to ensure the matching quality, thereby making the optimal user-resource matching decision and enhancing the system's matching ability in different scenarios;

[0153] Specifically, the larger the correlation matching factor obtained by adjusting through the optimal adjustment parameter γ, the higher the matching degree between the user and resource behavior characteristics. By adding a spatio-temporal local correlation enhancement layer and a temporal periodic feature capture layer to the basic spatio-temporal convolutional network model, although the basic spatio-temporal convolutional network model can extract some overall spatio-temporal features, its ability to mine the correlation between user behavior characteristics and resource behavior characteristics within local spatio-temporal regions is limited. In practical applications, the basic model is difficult to accurately capture these local correlation information, resulting in the inability to provide a sufficiently fine basis for intelligent matching. By improving the spatio-temporal convolutional network combined with a dynamic adjustment strategy, the ability to mine local spatio-temporal correlations is effectively strengthened, and the deep correlation information between user and resource behavior characteristics is accurately captured, providing a fine and reliable decision basis for the intelligent matching of big data.

[0154] In the application, several formulas involved are all calculated by taking the numerical value after dimensionless, and the establishment of the formula is obtained by collecting a large amount of data for software simulation to get a formula closest to the real situation. Some coefficients or weights in the formula are set by those skilled in the art according to the actual situation, so no more details will be elaborated here.

[0155] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution.

[0156] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent matching system based on big data resources, characterized in that, Including: Data acquisition module: Collect multi-source matching data and perform fusion processing, and obtain user behavior characteristics and resource behavior characteristics from the processed multi-source matching data; Feature adaptation module: Obtain the user behavior feature matrix and the resource behavior feature matrix, and obtain the optimal preliminary matching factor by constructing an improved spatio-temporal convolutional network model and inputting the user behavior feature matrix and the resource behavior feature matrix; Intelligent matching module: Label high-quality matching labels, use the high-quality matching labels as the labeled data set of a parameter calculation model based on the neural network architecture, obtain the optimal adjustment parameters through the parameter calculation model, dynamically adjust the optimal preliminary matching factor through the optimal adjustment parameters, obtain the final associated matching factor, and make the optimal user-resource matching decision according to the associated matching factor; The improved spatio-temporal convolutional network model is obtained by adding a custom layer to the basic spatio-temporal convolutional network model; The custom layer includes a spatio-temporal local association enhancement layer and a temporal periodic feature capture layer.

2. The intelligent matching system based on big data resources according to claim 1, wherein The multi-source matching data includes structured matching data, unstructured matching data, and spatio-temporal matching data; The user behavior feature matrix is obtained by directly representing the user behavior characteristics in matrix form; The resource behavior feature matrix is obtained by directly representing the resource behavior characteristics in matrix form.

3. An intelligent matching system based on big data resources according to claim 1, characterized in that, The improved spatio-temporal convolutional network model includes an input layer, a basic spatio-temporal convolutional layer, a spatio-temporal local association enhancement layer, a temporal periodic feature capture layer, and a fully connected layer; The input layer receives the user behavior feature matrix and the resource behavior feature matrix; The basic spatio-temporal convolutional layer outputs a basic feature representation according to the user behavior feature matrix and the resource behavior feature matrix; The spatio-temporal local association enhancement layer outputs strong local association features based on the basic feature representation; The temporal periodic feature capture layer outputs spatio-temporal periodic features based on the strong local association features; The fully connected layer outputs the optimal preliminary matching factor based on the spatio-temporal periodic features.

4. An intelligent matching system based on big data resources according to claim 3, characterized in that, The process of the spatio-temporal local association enhancement layer outputting strong local association features is as follows: Divide the basic feature representation into local spatio-temporal blocks through a fixed window to obtain user behavior local feature blocks and resource behavior local feature blocks; Obtain a set of user behavior local feature blocks and a set of resource behavior local feature blocks based on the user behavior local feature blocks and the resource behavior local feature blocks; Combine the set of user behavior local feature blocks and the set of resource behavior local feature blocks into a local spatio-temporal block; Extract local user behavior features and local resource behavior features from the local spatio-temporal block and calculate the association weight score; Normalize all the association weight scores to obtain a weight matrix; Perform weighted fusion on the local features based on the weight matrix to obtain the enhanced local feature representation of each local spatio-temporal block; Stitch the enhanced features of all local spatio-temporal blocks, and the spatio-temporal local association enhancement layer finally outputs strong local association features.

5. An intelligent matching system based on big data resources according to claim 3, characterized in that, The process of the temporal periodic feature capture layer outputting spatio-temporal periodic features is as follows: Extract user time features and resource time features from the strong local association features; Obtain user time feature segments and resource time feature segments based on the user time features and the resource time features through a sliding window; For the user time feature segment and the resource time feature segment within each sliding window, attention weights are obtained through a self-attention mechanism for feature weight assignment; Based on the attention weights, the weighted user time feature segment and the weighted resource time feature segment are obtained; The weighted user time feature segment and the weighted resource time feature segment are input into a recurrent neural network LSTM, and the output of the last time step is selected as the periodic feature of the user behavior and the periodic feature of the resource behavior; The periodic feature of the user behavior and the periodic feature of the resource behavior are concatenated to obtain the spatio-temporal periodic feature.

6. An intelligent matching system based on big data resources according to claim 3, characterized in that, The specific acquisition process of the optimal preliminary matching factor output by the fully connected layer is as follows: The input spatio-temporal periodic feature is flattened into a one-dimensional vector through the Flatten function to obtain a flattened feature vector; The optimal preliminary matching factor is obtained through bilinear transformation and activation operation on the flattened feature vector.

7. An intelligent matching system based on big data resources according to claim 1, characterized in that, The acquisition process of the correlation matching factor is as follows: Obtain the mean and standard deviation of the preliminary matching factors in the current batch, and dynamically adjust the preliminary matching factors based on the optimal adjustment parameter; The dynamic adjustment result is normalized through the Softmax function, and the correlation matching factor is output.

8. An intelligent matching system based on big data resources according to claim 7, characterized in that The acquisition process of the optimal adjustment parameter is as follows: Construct a parameter calculation model; Label high-quality matching labels for historical matching cases, and construct a labeled data set based on the high-quality matching labels; The parameter calculation model includes an input layer, a hidden layer, and an output layer; The input layer receives the labeled data set and inputs the features in the labeled data set into the network; The hidden layer performs weighted calculation and activation processing on the neurons for the input features; The output layer integrates the information processed by the hidden layer and outputs the optimal adjustment parameter through linear transformation.

9. An intelligent matching system based on big data resources according to claim 8, characterized in that, The hidden layer of the parameter calculation model includes l hidden layers, each containing n l neurons, and the number of neurons in the output layer of the parameter calculation model is one.

10. An intelligent matching system based on big data resources according to claim 8, characterized in that, The labeling rule for the high-quality label is as follows: Count the number of times a user actually selects a certain resource in historical matching cases. When the number exceeds a preset threshold, it is labeled as a high-quality match, and when it is less than or equal to the preset threshold, it is labeled as a non-high-quality match. Based on this rule, a labeled data set is constructed to convert user behavior into labels recognizable by the model.

Images