A system and method for evaluating the propagation effect based on multi-platform data fusion

Abstract

The present application relates to the technical field of propagation effect evaluation, in particular to a propagation effect evaluation system and method based on multi-platform data fusion, in the present application, the behavior data is converted into a structured unit by normalizing the forwarding frequency and mapping the position with the user ranking index, the fusion ability of different platform data in the time dimension and the frequency dimension is enhanced, the sequence vector of the propagation path is established by processing the order difference and the frequency difference between nodes through the graph neural network, the direction judgment and the weight combination are combined to form the coding structure, which is used to reflect the sequence and the propagation strength in the node propagation path, the similarity of the trajectory sequence is judged through the continuity matching operation of the path sequence structure, the core path with actual diffusion continuity is effectively screened out, the weighted relationship among the node strength, the path position sequence and the repeated index is combined, the aggregation processing is carried out according to the platform, the regional difference value is extracted, and the propagation strength index of each platform is obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of propagation effect evaluation, in particular to a propagation effect evaluation system and method based on multi-platform data fusion. BACKGROUND

[0002] The technical field of propagation effect evaluation aims to quantitatively analyze and systematically evaluate the coverage, propagation speed, user response degree and influence of information in the process of propagation on various media platforms. By constructing a propagation model, setting an evaluation index system, and introducing data processing and analysis algorithms, an objective judgment of the effect of propagation activities is realized, thereby providing decision support for content optimization, media strategy adjustment and marketing effect improvement.

[0003] A propagation effect evaluation system based on multi-platform data fusion aims to solve the problems of fragmented data sources, non-uniform evaluation results and strong subjectivity of evaluation dimensions in current propagation effect evaluation. It can realize the fusion processing of data from multiple heterogeneous platforms, establish a unified propagation effect evaluation model, and present the results with multi-dimensional indicators, so as to achieve quantitative evaluation of the propagation breadth, depth, continuity and user response.

[0004] The prior art has the problem that the heterogeneous data structure makes it difficult to analyze and synchronize. In the process of path recognition and influence factor extraction, it relies on static statistical values such as node number, forwarding volume or coverage, lacks dynamic structure judgment based on behavior path, cannot reveal the real backbone path and strong node distribution relationship in the propagation structure, and it is difficult to consider the interactive influence of user propagation timing and behavior frequency at the same time, resulting in high-frequency but non-continuous behavior being misidentified as high-influence nodes. Simple classification methods are commonly used to divide path segments by platform or time, ignoring the reconstruction and matching of structural features in the propagation trajectory, resulting in the same propagation chain being fragmented and attributed, and lacking the ability to judge the continuity and propagation extensibility across platforms. SUMMARY

[0005] The purpose of the present application is to solve the shortcomings in the prior art and to propose a propagation effect evaluation system and method based on multi-platform data fusion.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions: a propagation effect evaluation system based on multi-platform data fusion comprises:

[0007] A node behavior decomposition module: based on multi-platform user behavior records, time point set extraction, target user sorting and forwarding frequency normalization are performed, the frequency value and the sequence index are positionally mapped, and a node propagation feature set is generated;

[0008] The propagation trajectory construction module extracts the forwarding node sequence and interval time based on the node propagation feature set. It uses a graph neural network to construct the propagation path sequence vector based on the node sequence difference, and combines the frequency difference to perform direction judgment and weight mapping to establish a path sequence relationship model.

[0009] Impact signal extraction module: Based on the path sequence relationship model, extract the number of upstream and downstream nodes and calculate the propagation difference. Select nodes with a propagation difference greater than zero, and perform a product judgment by combining the path length difference and frequency value to screen out nodes that meet the conditions and obtain the high impact signal node group.

[0010] The propagation core identification module: Based on the high-impact signal node group, it counts the frequency of the path segment and judges the density, matches the time coverage range and sets an upper limit, extracts the trajectory and performs path distribution aggregation and sequence matching to obtain the core propagation path set;

[0011] Intensity index generation module: Based on the core propagation path set, extract node strength and path order, perform weighted operation on duplicate index, aggregate by platform group, and filter platform difference values ​​by combining regional identifiers to obtain the propagation platform intensity index.

[0012] As a further aspect of the present invention, the node behavior decomposition module includes:

[0013] Time series extraction submodule: Based on user behavior records from multiple platforms, extract the set of user forwarding time points, perform time format validity verification and sort the time point set, build a time mapping index set in ascending order, bind the index results with event numbers and output them to generate a time series index set;

[0014] User sequence sorting submodule: Based on the time series index set, extract the target user list, match the target identifier value in the forwarding node with the index sequence one by one, rearrange the sequence according to the time order, and integrate the user mapping order with the event order to generate the target user sequence group;

[0015] Frequency sequence mapping submodule: Based on the target user sequence group, it counts the number of times users appear repeatedly in each node, performs frequency value normalization and establishes a frequency vector, and then matches the normalization results to the corresponding sorting index positions in sequence to construct an ordered frequency vector group and generate a node propagation feature set.

[0016] As a further aspect of the present invention, the propagation trajectory construction module includes:

[0017] The node sequence generation submodule: based on the node propagation feature set, the starting and target nodes are extracted, the node pair combination in the user sequence is matched and the time index sequence is read, the index difference is extracted and mapped to the node number position, the structure of the node time sequence arrangement is constructed, and the node position index vector is obtained;

[0018] The propagation path vector construction submodule: based on the node position index vector, the time difference and the corresponding frequency difference of each node pair are extracted, the two types of difference values are linearly superimposed and node bit sequence splicing operations are performed through the graph neural network to construct a composite structure, and the structure is grouped and combined according to the path order to output, and a path order weight structure is generated;

[0019] The direction weight judgment submodule: based on the path order weight structure, the node propagation direction comparison operation is performed, the path direction is judged by comparing the time difference symbol and the bit sequence change value, and then the node pairs are classified according to the weight ascending order, the path structure information is combined and output, and the path order relationship model is established.

[0020] As a further scheme of the application, the graph neural network is according to the formula:

[0021]

[0022] Wherein: Indicates in the propagation effect evaluation system, the node After the fusion representation vector of the first Layer graph neural network is updated, Indicates a nonlinear activation function, Indicates the weight matrix of the first Layer, the parameters are optimized by gradient descent in the training process, Indicates the node In the first Layer representation vector, Indicates a splicing operation for connecting the current node representation and neighbor aggregation information, Indicates the aggregation operation on the neighbor node set Of the node , Indicates the neighbor node In the first Layer representation vector, Indicates the propagation time difference value of the node pair , Indicates the propagation frequency difference value of the node pair , , Indicates the structure tightness coefficient of the node pair , , , Indicates the node path order encoding, 、 、 、 represents the weight coefficient of each factor.

[0023] As a further scheme of the present application, the graph neural network first constructs a directed graph structure between nodes, each node of the graph corresponds to a user entity, and each edge represents a propagation event, the features of the edge include the time difference value and the propagation frequency difference value between the node pair, and then the features are input into the graph neural network model, the graph convolution is used to aggregate the features of the nodes, the edge features of the adjacent nodes are calculated by weighting for each node, the adjacent node features are superimposed and then linearly transformed to obtain the node embedding vector, then the node embedding vector is spliced with the matching position index vector to form a composite representation structure, and the path order division operation is performed in the output layer of the graph neural network, the composite representation is combined according to the propagation path logic through the clustering method, the embedding representation of each path is output, and the path order weight structure with the propagation strength and the time sequence features is generated.

[0024] As a further scheme of the present application, the influence signal extraction module comprises:

[0025] An upstream and downstream quantity extraction submodule: based on the path order relationship model, a node pair set construction operation is performed, the roles of the source nodes and the target nodes in all node pairs are counted and the nodes are divided into upstream and downstream categories, the upstream and downstream quantities of each node are summarized and the quantity difference before and after is calculated to obtain an upstream and downstream difference set;

[0026] A propagation strength screening submodule: based on the upstream and downstream difference set, a node screening operation is performed, the nodes with a difference greater than zero are extracted to complete preliminary screening, the total length of the path corresponding to the node and the number of repetitions of the node in the path are extracted, the product is calculated, and a fixed threshold value is set for the product value to screen out low-value nodes to obtain a propagation strength node set;

[0027] An influence node identification submodule: based on the propagation strength node set, a local density statistical operation of the nodes in the network is performed, the intersection quantity of the nodes as each other's target is calculated, and a classification partition operation is performed, a node set with the maximum connection strength in each partition is extracted as the screening result to obtain a high-influence signal node group.

[0028] As a further scheme of the present application, the propagation core identification module comprises:

[0029] The path frequency statistics submodule: based on the high-impact signal node group, a path segment extraction operation is performed, the occurrence frequency value of each node belonging to the propagation path segment is obtained, the path density value is calculated combined with the number of nodes covered by each segment, and the frequency value and the density value are double-filtered to obtain the frequency-intensive path segment set;

[0030] The time coverage filtering submodule: based on the frequency-intensive path segment set, a time interval processing operation is performed, the time index of the starting node and the ending node of the path segment is extracted and the span value is calculated, the span value is compared with the set upper time limit and the out-of-limit elimination operation is performed, and the time-constrained path segment group is obtained;

[0031] The trajectory path extraction submodule: based on the time-constrained path segment group, a node sequence extraction operation is performed, a trajectory set is constructed in time sequence for the nodes in the path segment, and a structure similarity matching operation is performed between the trajectory sets, the node sequence with high trajectory structure continuity in the matching result is taken as the output set, and the core propagation path set is obtained.

[0032] As a further scheme of the application, the double filtering operation of the frequency value and the density value includes two stages, the first stage, the frequency value of each path segment appearing in the propagation graph is counted, and a frequency threshold is set, and all path segments with a frequency higher than the threshold are retained, and the second stage, the density value of the path segment is further judged, and a density lower limit standard is set, and the path segment with a density lower than the standard is removed.

[0033] As a further scheme of the application, the strength index generation module includes:

[0034] The node index calculation submodule: based on the core propagation path set, a path node extraction operation is performed, the product of the node position sequence number and the corresponding propagation strength value is calculated, and the node repetition frequency is extracted and weighted and added to the product result to generate the node propagation influence value set;

[0035] The platform grouping aggregation submodule: based on the node propagation influence value set, a node platform identification extraction operation is performed, the platform identification content attached to each node is read for classification operation and a platform index table is constructed, the sum of the node values in the platform is calculated and the corresponding total value of the platform is recorded, and the platform influence value group is generated;

[0036] The regional difference extraction submodule: based on the platform influence value group, a regional identification field extraction operation is performed, a regional classification mapping table is constructed by identifying the regional label value of the platform, the platform influence value is aggregated according to the region and the aggregation processing is performed, and the propagation platform strength index is obtained.

[0037] A propagation effect evaluation method based on multi-platform data fusion, the propagation effect evaluation method based on multi-platform data fusion is executed based on the above-mentioned propagation effect evaluation system based on multi-platform data fusion, comprising the following steps:

[0038] S1: based on the multi-platform user behavior record data set, the time index value involved in each record is extracted, the user identification set with record value is selected, the time index order in the record is sorted, whether there is a continuous relationship between the time adjacent and the identification different of each user identification is identified, the appearance frequency value of each user identification in the relationship is counted and the frequency value is normalized, then the normalized frequency value is matched with the position index in the user identification sorting sequence, and a node propagation feature set is generated;

[0039] S2: based on the node propagation feature set, all user identification pairs are extracted, the time index difference and the difference between the frequency normalized value of each identification pair are obtained, the two difference value domains are linearly combined using a graph neural network, and a three-tuple structure is formed by splicing the user identification sorting bit, all three-tuple structures are grouped in time sequence, each group structure is used as a path component, the processing result is output in sequence order, and a path sequence relationship model is obtained;

[0040] S3: based on the path sequence relationship model, the number of forward connections and the number of backward connections of each user identification in the path are identified, the propagation difference value between the connection numbers is calculated, the user identification with positive propagation difference value is selected, the bit difference value of the identification corresponding path segment and the appearance frequency value in the path are read, the product of the two is judged, the user identification with product value higher than the screening standard is screened out, and a high-impact signal node group is constructed;

[0041] S4: based on the high-impact signal node group, the unique number of each user identification corresponding to the propagation path segment is identified, the appearance frequency value in all propagation paths is counted, the number of user identifications after deduplication in each path segment is read, the frequency value and the number of user identifications are respectively judged with the preset screening standard, only the path segment with frequency value and identification number higher than the standard is retained, and a frequency-intensive path segment set is obtained;

[0042] S5: the time span is calculated by extracting the time index of the first and last user, the frequency-intensive path segment set is removed based on the time span and the trajectory sequence is constructed, the trajectory sequence with high structural continuity is screened out through structural similarity, a core propagation path set is obtained, and based on the core propagation path set, the node strength and path bit sequence, repeated index weighting are extracted, grouped and aggregated according to the platform, the platform difference value is screened out combined with the region identification, and a propagation platform strength index is generated.

[0043] Compared with the prior art, the advantages and positive effects of the present application are:

[0044] In the application, the behavior data is converted into a structured unit by normalizing the forwarding frequency and mapping the position with the user ranking index, thereby enhancing the fusion capability of different platform data in the time dimension and the frequency dimension.

[0045] In the application, the sequence difference and the frequency difference between nodes are processed by a graph neural network to establish a sequence vector of the propagation path, combined with direction judgment and weight combination, to form an encoding structure for reflecting the sequence and propagation intensity in the node propagation path.

[0046] In the application, the number of upstream and downstream connections of each node is counted and the propagation difference is calculated, supplemented by the product calculation of the path length difference and the frequency value, to realize fine screening of nodes with dominant propagation capability, thereby improving the identification accuracy of the key influence source of propagation, and combining the frequency statistics and time coverage constraints to build a double screening mechanism of the path segment in the propagation intensity and duration, to exclude non-main path segments with low density or abnormal time length.

[0047] In the application, the similarity of the trajectory sequence is judged by the continuity matching operation of the path sequence structure, effectively screening out the core path with actual diffusion continuity, combining the weighted relationship between the node intensity, path position sequence and repeated index, and performing aggregation processing according to the platform to extract the regional difference value, and obtaining the propagation intensity index of each platform. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The system flowchart of the application is shown in the figure.

[0049] Figure 2 The method step schematic diagram of the application is shown in the figure. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and advantages of the application clearer and more apparent, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0051] In the description of the application, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only used to facilitate the description of the application and simplify the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the application. In addition, in the description of the application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0052] Example one

[0053] Referring to Figure 1 The application provides a technical solution: a propagation effect evaluation system based on multi-platform data fusion includes:

[0054] Node behavior decomposition module: based on multi-platform user behavior records, time point set extraction, target user sorting and forwarding frequency normalization are performed, the frequency value and the sequence index are positionally mapped to generate a node propagation feature set;

[0055] Propagation trajectory construction module: based on the node propagation feature set, the forwarding node sequence and the interval time are extracted, the propagation path sequence vector is constructed according to the node sequence difference by using the graph neural network, the direction judgment and the weight mapping are performed in combination with the frequency difference, and the path sequence relationship model is established;

[0056] Influence signal extraction module: based on the path sequence relationship model, the number of upstream and downstream nodes is extracted and the propagation difference is calculated, the nodes with a propagation difference greater than zero are selected, the product judgment is performed in combination with the path length difference and the frequency value, the nodes meeting the conditions are screened out, and a high-influence signal node group is obtained;

[0057] Propagation core identification module: based on the high-influence signal node group, the path segment frequency is counted and the density is judged, the time coverage range is matched and the upper limit is set, the trajectory is extracted, the path distribution is aggregated and the sequence is matched, and a core propagation path set is obtained;

[0058] Strength index generation module: based on the core propagation path set, the node strength and the path bit sequence are extracted, the weighted operation is performed on the repetition index, the aggregation is performed according to the platform grouping, the platform difference value is selected in combination with the region identifier, and the propagation platform strength index is obtained.

[0059] The node behavior decomposition module includes:

[0060] Time series extraction submodule: based on multi-platform user behavior records, user forwarding time point set extraction is performed, the time point set is subjected to time format legality verification and sorting, the time mapping index set is established in ascending order, the index result is bound with the event number and output, and a time series index set is generated;

[0061] User sequence sorting submodule: based on the time series index set, target user list extraction is performed, the target identifier value in the forwarding node is positionally and one-to-one corresponded with the index sequence, the sequence is rearranged according to the time sequence, the user mapping bit sequence and the event sequence are uniformly integrated, and a target user sequence group is generated;

[0062] The frequency sequence mapping submodule: based on the target user sequence group, the number of repeated occurrences of each node user is counted, the frequency value is normalized, and the frequency vector is established. The normalized result is matched to the corresponding sorting index position in turn to construct the sequential frequency vector group, and the node propagation feature set is generated;

[0063] The time sequence extraction submodule: based on the multi-platform user behavior record, the time format standardization algorithm is used to process the user forwarding time point set. The time field value in the forwarding record is read one by one. The year, month, day, hour, minute, and second involved in each record are judged by format. The legality is filtered according to whether each component of the time is an integer value and within the time range. After filtering out the record items that do not meet the conditions, a legal time set is constructed. The year field is arranged from small to large. In the case of the same year value, the month, day, hour, minute, and second field values are compared in turn. The ascending rearrangement in the set is completed in order. The sorted time set is set with a position index. The index starts from zero and is numbered continuously. Each time item corresponds to a unique position index. The original event number corresponding to the time item and the index value are paired. The number and index binding structure is output. The time sequence index set is generated;

[0064] The user sequence sorting submodule: based on the time sequence index set, the user mapping sequence construction method is used to extract the target user list. The record items with the target identification value are filtered from the forwarding record. The time index value bound in the item is extracted. The user identification is reordered from small to large according to the index value. The mapping bit sequence is assigned according to the position in the sorting queue. The mapping bit sequence and the original event sequence number are processed in both directions. The position matching set between the user identification and the event record is formed. The sequence bit of each user identification in the set and the event number bit are integrated into a unified sorting structure. The integrated user sequence structure set is output. The target user sequence group is generated;

[0065] The frequency sequence mapping submodule: based on the target user sequence group, the user behavior frequency normalization algorithm is used to count and normalize the number of repetitions of each user identification in the sequence. Each identification value in the user sequence is read. The same identification value is accumulated and counted. After the counting is completed, the cumulative frequency corresponding to each identification value is converted by ratio with the length of the user sequence to obtain the normalized frequency value. The normalized result is matched according to the sorting index position of the user identification in the sequence. Each frequency value is written into the corresponding position index of the user identification. The frequency vector sequence is formed according to the sorting index structure. The vector sequence and the sorting index are output synchronously to construct the position mapping structure set. The sequential frequency vector group is constructed. The node propagation feature set is generated.

[0066] The propagation trajectory construction module includes:

[0067] The node sequence generation submodule: based on the node propagation feature set, the starting and target nodes are extracted, the node pair combination in the user sequence is matched and the time index sequence is read, the index difference is extracted and mapped to the node number position, the structure of the node time sequence arrangement is constructed, and the node position index vector is obtained;

[0068] The propagation path vector construction submodule: based on the node position index vector, the time difference and the corresponding frequency difference of each node pair are extracted, the two types of difference values are linearly superimposed and the node bit sequence is spliced through the graph neural network to construct a composite structure, and the structure is grouped and combined according to the path order to output, and the path order weight structure is generated;

[0069] The direction weight judgment submodule: based on the path order weight structure, the node propagation direction comparison operation is performed, the path direction is judged by comparing the time difference symbol and the bit sequence change value, and the node pairs are classified according to the weight ascending order, the path structure information is combined and output, and the path order relationship model is established;

[0070] The node sequence generation submodule: based on the node propagation feature set, the starting and target nodes are extracted, the node pair sequence construction method is used, the user identifier in each record is used as the node number, all nodes are combined in pairs according to the time index sequence to form a node pair set, the node pairs are screened according to the screening standard that the starting node time is earlier than the target node, the combination items with equal time index or the starting node later than the target node are removed, the starting node number, the target node number and the corresponding time index value difference of the screened node pair structure record are read, the time index difference value is read, and the difference value is mapped to the starting node number to form a time mapping structure. The node numbers in the mapping structure are rearranged according to the mapping value to generate a node sequence, the position of each node in the node sequence is numbered in sequence, and the continuous index vector is output, and the node position index vector is obtained;

[0071] The propagation path vector construction submodule: based on the node position index vector, the time difference and the frequency difference of each node pair are extracted, the time index difference and the frequency normalized value difference are read in sequence according to the node pair, the two difference value vectors are linearly integrated by using the information superposition structure in the graph neural network, the linear integration takes the starting time index between nodes as the input initial parameter, the superposition weight is set to be one-to-one for the time difference value and the frequency difference value, and a double-dimensional mixed difference structure is output. The structure and the node position index are spliced, the position index is taken as the last dimension of the structure during splicing, the path aggregation processing is performed on the spliced result according to the node order, the grouping is performed according to the judgment rule that each group of continuous nodes is a path, the all path groups are combined and output, and the path order weight structure is generated;

[0072] The direction weight judgment submodule, based on the path order weight structure, uses a propagation direction state judgment method to compare the propagation direction of nodes. For each node pair, it reads the time index difference and judges the positive or negative state. At the same time, it extracts the position difference of the node pair in the path structure. The position difference is the position number of the target node in the path minus the position number of the starting node. Both values ​​are non-negative integers. If the time difference and the position difference are in the same direction, it is forward propagation. If the directions are opposite, it is reverse propagation. All node pairs are classified into forward set and reverse set according to their direction state. Then, the path weight value corresponding to each node pair is extracted. All node pairs are sorted and classified according to the weight value from large to small. The classification results are integrated into path structure units according to the combination relationship of node pairs. After outputting the integrated structure, a path order relationship model is established.

[0073] Graph neural networks, according to the formula:

[0074]

[0075] in: In the dissemination effectiveness evaluation system, nodes After the first The fused representation vector updated by the layered graph neural network Represents a non-linear activation function. Indicates the first The layer's weight matrix and parameters are optimized during training using gradient descent. Represents a node In the Layer representation vector, This indicates a concatenation operation, used to connect the current node's representation with information aggregated from its neighbors. Indicates a node The set of neighboring nodes Perform aggregation operation. Representing neighboring nodes In the Layer representation vector, Represents node pairs , The difference in propagation time, Represents node pairs , The difference in propagation frequency, Represents node pairs The structural compactness coefficient, Represents a node Path bit order encoding, , , , This represents the weighting coefficient for each influencing factor;

[0076] Execution process: first, for any node to be updated , collect the set of neighbor nodes connected in the propagation path from multiple platforms , extract the current representation vector of each neighbor node , then calculate the propagation time difference of node pairs according to the timestamp record in the multi-platform data , and count the propagation frequency difference of node pairs in unit time through the event log , then calculate the structure tightness coefficient , which is determined by the ratio of the number of common neighbors to the number of the union set of adjacency sets, reflecting the coupling degree in the inter-platform propagation path structure, and determine the position sequence of the node in the propagation path , and input the position sequence code generated by the sine function , then multiply the above neighbor representation vector , time difference , frequency difference , structure tightness and position sequence coupling term · by weight coefficients , , , , all weights are initially set to a uniform distribution, and are dynamically adjusted based on loss function feedback in the model training stage to minimize evaluation error, sum the weighted results to form neighbor aggregation features, concatenate with the node's own representation , map through the weight matrix and input the nonlinear activation function , output the propagation representation vector of the node in the next layer , iterate in all propagation path nodes, and finally output the complete path order weight structure to evaluate the multi-platform propagation effect.

[0077] ​The graph neural network firstly constructs a directed graph structure between nodes, each node of the graph corresponds to a user entity, and each edge represents a propagation event. The edge features include the time difference and propagation frequency difference between node pairs. Then the features are input into the graph neural network model. The graph convolution is used to aggregate the features of the nodes. The edge features of adjacent nodes are weighted and calculated, and the adjacent node features are linearly transformed to obtain the node embedding vector. Then the node embedding vector is spliced with the matching position index vector to form a composite representation structure. The path order division operation is performed in the output layer of the graph neural network. The composite representation is logically combined according to the propagation path by the clustering method, and the embedding representation of each path is output. The path order weight structure with propagation strength and time sequence features is generated.

[0078] The influence signal extraction module includes:

[0079] The upstream and downstream quantity extraction submodule: based on the path order relationship model, the node pair set construction operation is performed. The roles of the source nodes and the target nodes in all node pairs are counted and classified, and the upstream and downstream categories of each node are summarized and calculated. The upstream and downstream difference set is obtained.

[0080] The propagation strength screening submodule: based on the upstream and downstream difference set, the node screening operation is performed. The nodes with a difference greater than zero are extracted to complete the preliminary screening. The total length of the corresponding node path and the number of repeated times of the node in the path are extracted. The product is calculated, and the product value is set to a fixed threshold to screen out low-value nodes. The propagation strength node set is obtained.

[0081] The influence node identification submodule: based on the propagation strength node set, the local density statistics operation of the node in the network is performed. The intersection number of nodes as targets is calculated, and the classification partition operation is performed. A group of nodes with the maximum connection strength in each partition is extracted as the screening result, and the high-influence signal node group is obtained.

[0082] Upstream and downstream quantity extraction submodule: based on the path order relationship model, the node role division statistical method is used to construct the node pair set, the source node number and the target node number two fields are extracted from each node pair in the path encoding result as a record item, the source node number is recorded as the output node, and the target node number is recorded as the input node, the source node number in all records is uniquely de-duplicated to establish the upstream node set, the target node number is uniquely de-duplicated to establish the downstream node set, the number of occurrences of each node as a source node is counted to obtain the upstream count value, and the number of occurrences of each node as a target node is counted to obtain the downstream count value, the difference between the upstream count value and the downstream count value of each node is calculated to construct the upstream and downstream count difference item of each node, and the node number and the corresponding upstream and downstream difference value pair are outputted to obtain the upstream and downstream difference value set;

[0083] Propagation intensity screening submodule: based on the upstream and downstream difference value set, the node screening processing is performed by using the propagation quantity intensity threshold calculation method, the nodes with a difference value greater than zero are preliminarily screened, the nodes with a difference value less than or equal to zero are directly removed, only the positive node numbers are reserved, the path number set of the path segment participated by the reserved node numbers is read, the number of nodes of each path in the set is counted to obtain the total length of the path, the number of occurrences of each node in the path is independently counted to obtain the repetition frequency value, and the product of the total length of the path and the node repetition frequency is calculated, the product value is recorded as the node propagation quantity value, a fixed screening threshold is set in the set of all node quantity values, the node numbers less than the threshold are all deleted, and the reserved node number list is outputted to obtain the propagation intensity node set;

[0084] Influence node identification submodule: based on the propagation intensity node set, the local network connection is identified by using the density statistical method based on the mutual target relationship, each node in the node number set is taken as a starting node, the target node set in the path structure is counted, whether there is a node number that points to the original starting node again in the target node set is found, the node numbers that meet the mutual target condition are extracted as a bidirectional connection unit, the number of occurrences of each node in the bidirectional connection unit is calculated as the connection density count value, the density values of all nodes are grouped and sorted, the nodes with the same density count value are classified into the same partition, the propagation weight in the path structure between the nodes in each partition is compared and the node number group corresponding to the maximum value is taken as the strongest connection set in the partition, all the strongest connection sets are merged and the identification result is outputted to obtain the high-influence signal node group.

[0085] The propagation core identification module comprises:

[0086] The path frequency statistics submodule: based on the high-impact signal node group, the path segment extraction operation is performed, the frequency value of each node belonging to the propagation path segment is obtained, the path density value is calculated combined with the number of nodes covered by each segment, and the frequency value and the density value are double-screened to obtain the frequency-dense path segment set;

[0087] The time coverage screening submodule: based on the frequency-dense path segment set, the time interval processing operation is performed, the time index of the starting node and the ending node of the path segment is extracted and the span between the indexes is calculated, the span value is compared with the set upper limit of time one by one, and the out-of-limit elimination operation is performed to obtain the time-constrained path segment group;

[0088] The trajectory path extraction submodule: based on the time-constrained path segment group, the node sequence extraction operation is performed, the trajectory set is constructed in time sequence for the nodes in the path segment, and the structural similarity matching operation between the trajectory sets is performed, the node sequence with high trajectory structure continuity in the matching result is taken as the output set, and the core propagation path set is obtained;

[0089] The path frequency statistics submodule: based on the high-impact signal node group, the path node attribution frequency calculation method is used to perform the path segment extraction operation, the path number corresponding to each node is read one by one, the mapping table of nodes and paths is constructed, the number of nodes associated with each path number is accumulated and the node set under the path number is recorded, the number of times the path number appears in all node records is counted as the path frequency value, the number of nodes under the path number is uniquely counted to obtain the path node number value, and the path frequency value and the node number value are paired and recorded for the path number. Set the frequency threshold and the density threshold, filter out the path number whose frequency value is greater than the frequency threshold and whose node number value is greater than the density threshold and mark it as an effective path segment, output all effective path segment numbers and corresponding node sets, and generate the frequency-dense path segment set;

[0090] The time coverage screening submodule: based on the frequency-dense path segment set, the time span comparison screening method is used to perform the time interval processing operation, the time index values of the first node and the tail node in each path segment are read, the time index corresponding to the first and last nodes is combined according to the path segment number, the time index difference value is calculated as the time span record, a fixed time upper limit value is set, the time span of each path segment is compared with the time upper limit one by one, only the path segment number whose time span value is less than or equal to the upper limit value is reserved, the corresponding node sequence set is extracted and outputted according to the path number, and all path segment set numbers that meet the time condition are summarized to generate the time-constrained path segment group;

[0091] The trajectory path extraction submodule: based on the time constraint path segment group, the node sequence extraction operation is performed by using the trajectory sequence structure continuity matching algorithm, the node set under each path segment number is sorted in ascending order according to the time index value to generate a trajectory sequence set, all trajectory sequence sets are paired, the structure similarity of each pair of trajectories is compared, in the structure similarity comparison operation, the two trajectories are compared one by one according to the node number, and the number of the same node number in the corresponding index position is calculated, then the matching number is divided by the length of the shorter trajectory to form a matching proportion value, the trajectory pairs with a matching proportion value higher than a set threshold value in all pairs are recorded as structure continuous pairs, and any trajectory sequence in each structure continuous pair is extracted as a representative trajectory, the node sequence of all trajectory sets satisfying the condition is output, and the core propagation path set is obtained.

[0092] The frequency value and the density value are subjected to double screening operation, and the operation process includes two stages. In the first stage, the frequency value of each path segment appearing in the propagation graph is counted, and a frequency threshold is set. All path segments with a frequency higher than the threshold are retained. In the second stage, the path segment is further judged by density value, and a density lower limit standard is set. The path segment with a density lower than the standard is removed.

[0093] The strength index generation module includes:

[0094] The node index calculation submodule: based on the core propagation path set, the node extraction operation in the path is performed, the product of the bit sequence number of each node in the path and the corresponding propagation strength value is calculated, and the number of repeated nodes is extracted and weighted and accumulated with the product result to generate a node propagation influence value set;

[0095] The platform grouping aggregation submodule: based on the node propagation influence value set, the node platform identification extraction operation is performed, the platform index table is constructed by reading the platform identification content attached to each node and performing classification operation, the platform value summation operation is performed and the corresponding total value is recorded to generate a platform influence value group;

[0096] The regional difference extraction submodule: based on the platform influence value group, the regional identification field extraction operation is performed, the regional classification mapping table is constructed by identifying the platform belonging area label value, the platform influence value is aggregated according to the area and the summary processing is performed to obtain the propagation platform strength index;

[0097] The node index calculation submodule: based on the core propagation path set, the node propagation bit sequence weighting superposition algorithm is used for path internal node extraction operation, the node list under each path set number is numbered in time sequence, the position number of each node in the path is marked and the numbering mapping table is established, the propagation intensity value recorded in the path structure corresponding to each node is extracted, the product calculation is carried out after the node bit sequence number and the corresponding propagation intensity value are read, the product result is taken as the initial influence value of the node, the repeated node numbering in all path sets is executed, the occurrence frequency of each node numbering is taken as the weighted multiple value, the weighted accumulation operation processing is carried out on all product results of the same node numbering, the set of all node numbering and accumulation result value is output, and the node propagation influence value set is generated;

[0098] The platform grouping aggregation submodule: based on the node propagation influence value set, the platform identification field clustering summation method is used for node platform identification extraction operation, the platform identification field content of the combination item of node numbering and influence value is read item by item, the node numbering with the same platform identification field is classified into the same platform classification structure, the index mapping table between platform identification and node numbering is established, the influence value of all node numbering in each platform classification is executed, the summation result is paired with the platform identification, and the platform identification numbering is recorded. Process, summarize the summation value structure under all platforms and output the node influence value result corresponding to each platform, generate the platform influence value group;

[0099] The regional difference extraction submodule: based on the platform influence value group, the platform region aggregation index algorithm is used for regional identification field extraction operation, the regional attribution label attached to each platform identification is read item by item, the platform identification is classified and arranged according to the regional attribution label, and the mapping relationship matrix of regional identification and platform identification is constructed, the corresponding platform influence value in each region is read and added, each platform influence value is set as a single input, the sum of the platform values in the region is calculated, the corresponding accumulation value of each region is combined with the regional label to form a structure list, all regional result record structures are summarized and uniformly coded, and the propagation platform intensity index is obtained.

[0100] Please refer to Figure 2 A propagation effect evaluation method based on multi-platform data fusion, the propagation effect evaluation method based on multi-platform data fusion is executed based on the above-mentioned propagation effect evaluation system based on multi-platform data fusion, including the following steps:

[0101] S1: Based on the multi-platform user behavior record data set, the time index value involved in each record is extracted, the user identification set with record value is selected, the time index order in the record is sorted, whether there is a time adjacent and identification different continuous relationship between each user identification is identified, the appearance frequency value of each user identification in the relationship is counted and the frequency value is normalized, then the normalized frequency value is matched with the position index in the user identification sorting sequence, and the node propagation feature set is generated;

[0102] S2: Based on the node propagation feature set, all user identification pairs are extracted, the time index difference and the difference between the frequency normalization value of each identification pair are obtained, the two difference value domains are linearly combined using a graph neural network, and the position index of the user identification is spliced to form a triple structure, and all triple structures are grouped according to time sequence, each group structure is taken as a path component, and the processing result is output in sequence order, and a path sequence relationship model is obtained;

[0103] S3: Based on the path sequence relationship model, the number of forward connections and the number of backward connections of each user identification in the path are identified, the propagation difference value between the connection numbers is calculated, the user identification with positive propagation difference value is selected, the bit difference value and the appearance frequency value in the path of the identification are read, the product of the two is judged, the user identification with product value higher than the screening standard is screened out, and a high-impact signal node group is constructed;

[0104] S4: Based on the high-impact signal node group, the unique number of the propagation path segment corresponding to each user identification is identified, the appearance frequency value in all propagation paths is counted, the number of user identifications after deduplication in each path segment is read, the frequency value and the number of user identifications are respectively judged with the preset screening standard, only the path segment with frequency value and identification number higher than the standard is retained, and a frequency-intensive path segment set is obtained;

[0105] S5: The time span is calculated based on the time index of the first and last user, the frequency-intensive path segment set is removed based on the time span, and a trajectory sequence is constructed, the trajectory sequence with high structural continuity is selected through structural similarity, a core propagation path set is obtained, and based on the core propagation path set, the node strength and path bit sequence, repeated index weighting are extracted, grouped and aggregated according to the platform, the platform difference value is screened combined with the region identification, and a propagation platform strength index is generated.

[0106] The above is only a preferred embodiment of the present application, not other forms of the present application, any skilled person in the art can use the above disclosed technical content to make changes or modifications as equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application without departing from the technical solution content of the present application still belongs to the protection scope of the present application.

Claims

1. A dissemination effect evaluation system based on multi-platform data fusion, characterized in that: The system comprises: a node behavior decomposition module: based on multi-platform user behavior records, performing time point set extraction, target user sorting and forwarding frequency normalization, mapping the frequency value to the order index to generate a node propagation feature set; a propagation trajectory construction module: based on the node propagation feature set, extracting the forwarding node sequence and interval time, using a graph neural network to construct a propagation path order vector according to the node order difference, combining the frequency difference to perform direction judgment and weight mapping, and establishing a path order relationship model; an influence signal extraction module: based on the path order relationship model, calculating the number of nodes upstream and downstream and the propagation difference, screening nodes with a propagation difference greater than zero, combining the path length difference and the frequency value product to obtain a high-influence signal node group; a propagation core identification module: based on the high-influence signal node group, counting the path segment frequency and judging the density, matching the time coverage and setting the upper limit, extracting the trajectory and performing path distribution aggregation and sequence matching to obtain a core propagation path set; a strength index generation module: based on the core propagation path set, extracting the node strength and path position sequence, repeating index weighting, grouping according to platforms, combining with the region identifier to filter platform difference values, and generating a propagation platform strength index. 2.The multi-platform data fusion based propagation effect evaluation system according to claim 1, wherein, The node behavior decomposition module comprises: a time series extraction submodule: based on multi-platform user behavior records, extracting a set of user forwarding time points, verifying the time format and sorting in ascending order, establishing a time mapping index and binding it with the event number, and generating a time series index set; a user sequence sorting submodule: based on the time series index set, extracting a target user list, corresponding the target identifier of the forwarding node to the index sequence, rearranging in time sequence, and obtaining a target user sequence group; a frequency sequence mapping submodule: based on the target user sequence group, counting the number of repeated occurrences of users and normalizing, establishing a frequency vector, matching the normalized result to the sorting index position, and generating a node propagation feature set. 3.The multi-platform data fusion based propagation effect evaluation system according to claim 1, wherein, The propagation trajectory construction module comprises: a node sequence generation submodule: based on the node propagation feature set, extracting the starting and target nodes, combining the nodes in the user sequence and reading the time index sequence, extracting the index difference and mapping it to the node number position, constructing a structure arranged in time sequence, and obtaining a node position index vector; a propagation path vector construction submodule: based on the node position index vector, extracting the time difference and corresponding frequency difference of each pair of nodes, constructing a composite structure by linear superposition and node position splicing of the two types of difference values through a graph neural network, and grouping and combining the structure according to the path order to output, generating a path order weight structure; a direction weight judgment submodule: based on the path order weight structure, performing node propagation direction comparison operation, judging the path direction by comparing the time difference sign and the bit sequence change value, and classifying the node pairs according to the weight ascending order, combining the path structure information and outputting, and establishing a path order relationship model. 4.The multi-platform data fusion based propagation effect evaluation system according to claim 3, wherein, The graph neural network is according to the formula: ； wherein: represents a node in the propagation effect evaluation system after the first layer graph neural network update, represents a nonlinear activation function, represents the weight matrix of the layer, which is optimized by gradient descent during the training process, represents a node in the layer, represents a concatenation operation for connecting the current node representation with the neighbor aggregated information, represents an aggregation operation on the neighbor node set of the node , represents the representation vector of the neighbor node in the layer, represents the propagation time difference value of the node pair , , represents the propagation frequency difference value of the node pair , , represents the structure tightness coefficient of the node pair , , represents the path bit sequence encoding of the node , , , , represents the weight coefficient of each influence factor. 5.The multi-platform data fusion based propagation effect evaluation system according to claim 3, wherein, The graph neural network firstly constructs a directed graph structure between nodes, each node of the graph corresponds to a user entity, and each edge represents a propagation event, the features of the edge include the time difference value and the propagation frequency difference value between the node pair, and then the features are input into the graph neural network model, the graph convolution is used to aggregate the features of the nodes, the edge features of the adjacent nodes are weighted and calculated, the adjacent node features are superimposed and linearly transformed to obtain the node embedding vector, then the node embedding vector is spliced with the matching position index vector to form a composite representation structure, and a path order division operation is performed on the output layer of the graph neural network, the composite representation is logically combined according to the propagation path by a clustering method, and the embedding representation of each path is output, thereby generating a path order weight structure with propagation intensity and time sequence features. 6.The multi-platform data fusion based propagation effect evaluation system according to claim 1, wherein, The influence signal extraction module comprises: An upstream and downstream quantity extraction submodule: based on the path order relationship model, a node pair set is constructed, the source node and the target node role are counted, the upstream and downstream categories are divided, the quantities are summarized and the quantity difference before and after is calculated, and an upstream and downstream difference set is generated; A propagation intensity screening submodule: based on the upstream and downstream difference set, nodes with a difference greater than zero are screened, the total length of the path and the repetition times are extracted, the product is calculated and compared with a threshold, low-value nodes are removed, and a propagation intensity node set is obtained; An influence node identification submodule: based on the propagation intensity node set, the intersection quantity of nodes that are targets of each other is counted, classified according to partitions, and a node set with the maximum connection strength in each partition is extracted as a high-influence signal node group. 7.The multi-platform data fusion based propagation effect evaluation system according to claim 1, wherein, The propagation core identification module comprises: A path frequency statistical submodule: based on the high-influence signal node group, a path segment extraction operation is performed, the propagation path segment to which each node belongs is obtained, the frequency value of the path segment is counted, the path density value is calculated according to the number of nodes covered by each segment, and a frequency and density double screening operation is performed on the frequency value and the density value to obtain a frequency and density intensive path segment set; A time coverage screening submodule: based on the frequency and density intensive path segment set, a time interval processing operation is performed, the time indexes of the start node and the end node of the path segment are extracted, the span between the indexes is calculated, the span value is compared with a set upper time limit, and an out-of-limit removal operation is performed to obtain a time-constrained path segment group; A trajectory path extraction submodule: based on the time-constrained path segment group, a node sequence extraction operation is performed, a trajectory set is constructed according to the time sequence of the nodes in the path segment, and a structure similarity matching operation is performed between the trajectory sets, the node sequence with high trajectory structure continuity in the matching result is taken as an output set, and a core propagation path set is obtained. 8.The multi-platform data fusion based propagation effect evaluation system according to claim 6, wherein, The frequency and density double screening operation comprises two stages, in the first stage, the frequency value of each path segment in the propagation graph is counted, and a frequency threshold is set, and all path segments with a frequency higher than the threshold are retained, in the second stage, the density value of the path segment is further judged, a density lower limit standard is set, and path segments with a density lower than the standard are removed. 9.The multi-platform data fusion based propagation effect evaluation system according to claim 1, wherein, The strength index generation module comprises: The node index calculation submodule: based on the core propagation path set, extracting the nodes within the path, reading the product of each node bit sequence number and propagation strength value, combining the weighted accumulation of node repetition times, generating the node propagation influence value set; The platform grouping aggregation submodule: based on the node propagation influence value set, extracting the node platform identifier, classifying and constructing the platform index table, calculating the total sum of the node values within the platform, and generating the platform influence value group; The regional difference extraction submodule: based on the platform influence value group, extracting the platform belonging region label, constructing the regional classification mapping table, aggregating the platform influence values within the region and summarizing, and obtaining the propagation platform strength index. 10.A method for evaluating propagation effect based on multi-platform data fusion, characterized in that, The multi-platform data fusion-based propagation effect evaluation system according to any one of claims 1-9, comprising the following steps: S1: Extracting the time index and user identifier of the multi-platform user behavior record, normalizing the user identifier occurrence frequency, and combining the time index to generate the node propagation feature set; S2: Based on the node propagation feature set, extracting the user identifier pair, calculating the time index difference and frequency difference of the user identifier pair, using graph neural network to fuse the two types of difference, grouping to generate path composition sequence and outputting path order relationship model; S3: Based on the path order relationship model, calculating the propagation difference value by the number of connections before and after the user, selecting the user with positive propagation difference value, and combining the product of path bit difference and occurrence frequency to construct high-influence signal node group; S4: Based on the high-influence signal node group, counting the occurrence frequency and user number of the path segment, retaining the path segment with occurrence frequency and user number higher than the set standard, and obtaining the frequency-intensive path segment set; S5: Extracting the time index of the first and last users to calculate the time span, based on the time span, removing the ultra-limit path segment in the frequency-intensive path segment set and constructing the trajectory sequence, selecting the trajectory sequence with high structural continuity through structural similarity, obtaining the core propagation path set, and based on the core propagation path set, extracting the node strength and path bit sequence, weighted repetition index, grouping according to the platform, screening the platform difference value combined with the region identifier, and generating the propagation platform strength index.

Images