E-commerce Online Settlement System Based on Transaction Behavior Analysis

Abstract

The present invention discloses an e-commerce online settlement system based on transaction behavior analysis, which relates to the technical field of online settlement, and includes a preliminary commodity prediction unit, an aggregated feature determination unit, a final commodity prediction unit, a transaction settlement unit, and a false transaction identification unit. Through the preliminary commodity prediction unit, the system can preliminarily predict some of the commodities required by each user in the next round of shopping. In order to further improve the prediction accuracy, the aggregated feature determination unit and the final commodity prediction unit are used to determine the aggregated feature representation and associated feature representation corresponding to each independent sequence set. Since the behaviors of users browsing and collecting commodities are combined in the analysis process, the subsequent prediction of the commodity selection priority according to these feature representations can be more accurate. The system summarizes the commodity numbers predicted twice and pushes the commodity information to the user interface according to the commodity numbers, enhancing the user experience and increasing the probability of transaction completion at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of online settlement, and more specifically, to an e-commerce online settlement system based on transaction behavior analysis. Background Art

[0002] The e-commerce online settlement system based on transaction behavior analysis provides personalized recommendations, risk management, and precision marketing functions by deeply analyzing user behavior, helping merchants improve the user experience, optimize operations, and increase sales. In the invention patent with the application number 202310890337.3, "An online payment optimization system and method based on an e-commerce platform" is disclosed. This invention can monitor the payment delay parameter in real time and send an alarm signal when it exceeds the standard delay threshold. When it does not exceed the standard delay threshold, it will summarize it into a dataset to be evaluated, so as to predict the payment delay parameter of the predicted payment information, and then the optimization steps can be executed before it exceeds the standard delay threshold to ensure the smooth operation of the online payment system. And during the optimization process, not all historical payment information will be uploaded to the cloud storage, but part of the historical payment information will be retained in the payment system for users to view the historical payment information in a timely manner, further enhancing the user experience.

[0003] The above-mentioned prior art solves problems such as the reduction of the user experience caused by an excessive payment delay parameter. However, when the system is running, since it does not analyze the user's browsing records, favorite records, and purchase records, the recommended product information cannot attract users to make purchases, resulting in the system being unable to increase sales for merchants. At the same time, when there are false transactions on the e-commerce platform, the system cannot provide relatively accurate data for supervisors in a short time, resulting in supervisors being unable to discover and take measures to intervene in a timely manner. Summary of the Invention

[0004] The purpose of the present invention is to provide an e-commerce online settlement system based on transaction behavior analysis to solve the problems raised in the above background art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: An e-commerce online settlement system based on transaction behavior analysis, including a false transaction identification unit;

[0006] A preliminary commodity prediction unit, which extracts the purchase records of all users within a specified e-commerce platform, divides all purchase records according to the user numbers and transaction times in the records to obtain multiple groups, constructs multiple independent sequence sets for each group according to the transaction time, statistically analyzes the time feature vectors corresponding to all independent sequence sets in the group, uses Fourier transform to convert them into discrete features, sets a weight matrix, and then uses the weight matrix and discrete features to analyze the frequency feature vectors of each independent sequence set. After converting the frequency feature vectors of all independent sequence sets into corresponding time feature vectors through inverse Fourier transform, uses the FNN model to analyze the feature representations of multiple independent sequence sets corresponding to each user, statistically analyzes the occurrence times of each commodity in all independent sequence sets corresponding to the user, and predicts the commodity numbers corresponding to the independent sequence sets at the next round of transaction for each user by using the occurrence times of the commodity and the feature representations of the independent sequence sets;

[0007] An aggregation feature determination unit, which determines the embedding vectors and interaction label information of the commodities within all independent sequence sets corresponding to each user, combines and analyzes the commodity embedding vectors and interaction label information within the independent sequence sets to determine the interaction time feature vectors corresponding to each independent sequence set at different scales, converts the interaction time feature vectors into discrete feature vectors, and then performs a dot product operation on the discrete feature vectors through the weight matrix to obtain optimized interaction frequency feature vectors, converts the interaction frequency feature vectors into corresponding interaction time feature vectors, and analyzes the aggregation feature representations of multiple independent sequence sets corresponding to each user according to the interaction time feature vectors at different scales;

[0008] A final commodity prediction unit, which receives the commodity numbers and interaction label information within all independent sequence sets corresponding to each user, analyzes them to obtain the connection matrix corresponding to each user, uses a convolutional network to calculate the connection matrix to obtain the correlation feature representations of multiple independent sequence sets corresponding to each user, analyzes the aggregation feature representations and correlation feature representations of each independent sequence set through a fusion feature analysis algorithm to obtain the fusion feature representations of each independent sequence set corresponding to each user, and predicts the commodity numbers corresponding to the independent sequence sets at the next round of transaction for each user according to the fusion features of each independent sequence set corresponding to each user;

[0009] A transaction settlement unit, which extracts all predicted commodity numbers for the next round of transaction of each user, transmits the commodity information to the user recommendation interface according to the predicted commodity numbers, receives the settlement request sent by the user, returns the payment method and process to the user settlement interface, determines the payment completion time, uses it as the transaction time, and generates purchase records according to the user number, commodity number, commodity name, and transaction time.

[0010] Preferably, the preliminary commodity prediction unit includes a purchase record division module, a sequence set generation module, and a time feature analysis module. The purchase record division module extracts the purchase records of all users in a specified e-commerce platform. The purchase records include user numbers, commodity numbers, commodity names, and transaction times. All purchase records are divided according to user numbers and transaction times, and the purchase records with the same user number are stored in the same group. The purchase records in the group are sorted according to the transaction time, so as to obtain multiple sorted groups. The sequence set generation module determines multiple independent sequence sets corresponding to each group according to the transaction time, counts the purchase records with the same transaction time in each group, extracts the commodity numbers in the purchase records with the same transaction time, and then stores them in the corresponding independent sequence sets according to the transaction time. After the time feature analysis module determines the commodity embedding dimension, the total number of different commodities, and the commodity one-hot vector, it calculates the embedding vectors corresponding to the commodities in each independent sequence set according to the embedding dimension, the total number, and the one-hot vector, and performs a max pooling operation on the embedding vectors to obtain the time feature vector of each independent sequence set.

[0011] Preferably, the preliminary commodity prediction unit further includes a frequency feature analysis module, a feature conversion module, and a commodity number prediction module. The frequency feature analysis module counts the time feature vectors corresponding to all independent sequence sets in the group , where is the group number, is the number of independent sequence sets included in the group . Using Fourier transform, the time feature vectors of all independent sequence sets are converted into corresponding discrete features , where is the independent sequence set number. After setting two weight matrices and for the group , and the discrete feature of the th independent sequence set are used to calculate the frequency feature vector of this independent sequence set, where , represents the activation function, represents the adjustment coefficient, Denote the parameter. After the feature transformation module repeats the operation until the frequency feature vectors of each independent sequence set in each group are determined, the frequency feature vectors of the independent sequence sets are analyzed using the sequence feature analysis algorithm to obtain the simplified frequency feature vectors. Through the inverse Fourier transform, the simplified frequency feature vectors of all independent sequence sets are converted into corresponding time feature vectors. The product number prediction module transmits the time feature vectors of each independent sequence set in each group to the FNN model for analysis. After obtaining the feature representations of multiple independent sequence sets corresponding to each user, the occurrence times of each product in the independent sequence sets corresponding to the user are counted. The occurrence times of the products and the feature representations of the independent sequence sets are aggregated and analyzed to obtain the selection priority corresponding to each product. Set the screening number, and predict the product numbers corresponding to the independent sequence sets in the next round of transactions for each user according to the selection priority of the products and the screening number. The sequence feature analysis algorithm is specifically as follows:

[0012] ;

[0013] wherein, denotes the group the frequency feature vector after simplification of the th independent sequence set denotes the group the original frequency feature vector of the th independent sequence set denotes the shrinkage parameter and denotes the weight matrix denotes the group number denotes the number of the independent sequence set denotes the parameter denotes the sign function denotes the maximum value function

[0014] Preferably, the aggregation feature determination unit includes a record extraction module, a tag setting module, an embedding vector analysis module, an interaction feature optimization module, and a feature splicing module. The record extraction module extracts the browsing records and favorite records of all users in a specified e-commerce platform. The browsing records include user numbers, product numbers, product names, and browsing times. The favorite records include user numbers, product numbers, product names, and favorite times. The tag setting module divides the interaction types into three types, namely favorite, browse, and purchase, and determines the interaction tag information of the products in all independent sequence sets corresponding to each user according to the browsing records and favorite records. The interaction tag information includes browse tags and favorite tags. If the transaction time of the product in the independent sequence set is later than the browsing time, the corresponding browse tag is set to 1, otherwise it is set to 0. If the transaction time of the product in the independent sequence set is later than the favorite time, the corresponding browse tag is set to 1, otherwise it is set to 0. The embedding vector analysis module combines and analyzes the product embedding vectors and interaction tag information in all independent sequence sets corresponding to each user to obtain the interaction embedding vectors of the products in the sequence set, and performs a pooling operation on the interaction embedding vectors of each product to determine the interaction time feature vectors corresponding to each independent sequence set at different scales. The interaction feature optimization module uses Fourier transform to convert the interaction time feature vectors of the independent sequence set into discrete feature vectors, and then performs a dot product operation on the discrete feature vectors through a weight matrix to obtain the optimized interaction frequency feature vectors, and uses the inverse Fourier transform to convert the optimized interaction frequency feature vectors into the corresponding interaction time feature vectors. The feature splicing module splices the interaction time feature vectors of each independent sequence set at different scales to obtain the aggregated time feature vectors of the independent sequence set, and transmits them to the PFFN model for analysis to obtain the aggregated feature representations of multiple independent sequence sets corresponding to each user.

[0015] Preferably, the commodity final prediction unit includes a matrix generation module, an associated feature analysis module, a fusion feature analysis module, and a transaction prediction module. After receiving the commodity number and commodity interaction label information in all independent sequence sets corresponding to each user, the matrix generation module uses a matrix generation algorithm to analyze the commodity number and interaction label information in each independent sequence set, and obtains a relationship matrix and an interaction matrix corresponding to each user. The associated feature analysis module splices the relationship matrix and the interaction matrix corresponding to each user to generate a connection matrix, and uses a convolutional network to analyze the connection matrix to obtain the associated feature representation of multiple independent sequence sets corresponding to each user. After receiving the aggregated feature representation and the associated feature representation of multiple independent sequence sets corresponding to each user, the fusion feature analysis module calculates the aggregated feature representation and the associated feature representation respectively, determines the corresponding dynamic weights of the two, and analyzes the dynamic weights, the aggregated feature representation, and the associated feature representation of each independent sequence set through a fusion feature analysis algorithm to obtain the fusion feature representation of each independent sequence set corresponding to each user. The transaction prediction module analyzes the fusion features of each independent sequence set corresponding to each user to obtain the commodity selection priority corresponding to each user, and predicts the commodity numbers corresponding to the independent sequence sets when each user conducts the next round of transactions according to the commodity selection priority and the screening number. The matrix generation algorithm is specifically:

[0016] ;

[0017] ;

[0018] Among them, represents the relationship matrix of user . represents the commodity node in the th independent sequence set, represents the relationship edge between commodities in the th independent sequence set, represents the interaction matrix of user . represents the commodity node with interaction label information in the th independent sequence set, represents the interaction edge between commodities in the th independent sequence set, represents the set of all commodity nodes in the th independent sequence set, represents the set of all commodity nodes with interaction label information in the th independent sequence set, represents a parameter, represents the number of the independent sequence set, represents the A user, represents the user number.

[0019] Preferably, the transaction settlement unit includes a product recommendation module, a price determination module, and a product settlement module. After the product recommendation module extracts all the predicted product numbers for the next round of transactions of each user, it queries the database according to the predicted product numbers, and transmits the obtained product information to the user recommendation interface. When the user browses and collects products, corresponding records are generated and transmitted to the database. After receiving the settlement request sent by the user, the price determination module determines the actual price of the product according to the product number, user number, and request time included in the settlement request, returns it to the user settlement interface, and waits for the user to confirm the information. If the product settlement module does not receive the confirmation information within the preset time, it determines that the current user has not purchased the product, takes the request reception time as the collection time of the product by the current user, generates a collection record according to the user number, product number, product name, and collection time, and stores it in the database. If the confirmation information is received within the preset time, it returns the payment method and process to the user settlement interface, determines the payment completion time, takes it as the transaction time, generates a purchase record according to the user number, product number, product name, and transaction time, and stores it in the database.

[0020] Preferably, the false transaction identification unit includes a comment record screening module, a user node storage module, and a node marking module. The comment record screening module extracts the product comment records of each user within a specified e-commerce platform. The product comment records include user number, product number, comment content, and comment time. According to the user number, it determines the total number of corresponding purchase records. If the total number is lower than the threshold, it eliminates all the product comment records of this user. According to the product number, it determines the total number of comment records. If the total number is lower than the threshold, it eliminates all the comment records corresponding to this product. After the user node storage module determines the uneliminated product comment records and purchase records, it calculates the similarity between users according to the product comment records and purchase records, constructs multiple user nodes according to the total number of users and user numbers, determines the distance value between user nodes through the similarity between users, and stores all user nodes in a set. After setting the first distance value and the second distance value, the node marking module arbitrarily selects an unmarked user node from the set, takes it as the target node and marks it, and counts the distance value between this target node and other nodes. If the distance value is less than the first distance value, it adds the current node to the adjacent set of the target node and marks it. If the distance value is greater than or equal to the first distance value and less than the second distance value, it adds the current node to the adjacent set without marking. If the distance value is greater than the second distance value, it does not add the current node to the adjacent set. Repeat the operation until all nodes in the set are marked.

[0021] Preferably, the false transaction identification unit further includes a group scoring module and a false degree display module. After the group scoring module counts the central points in each adjacent set and uses them as the initial central points, the K-Mean clustering algorithm is used to classify all nodes in the set according to the initial central points, obtaining multiple candidate groups. The transaction evaluation algorithm is used to analyze each candidate group to determine the variance of the transaction time interval, tightness, co-occurrence times of commodities, and repeated purchase times corresponding to each group. Each group is scored according to the variance of the transaction time interval, tightness, co-occurrence times of commodities, and repeated purchase times. The false degree display module selects the group with the highest score, determines the commodity scoring records and purchase records corresponding to all user nodes in this group, analyzes the transaction false degree of the corresponding commodities according to the commodity scoring records and purchase records, and outputs the user's commodity scoring records, purchase records, and the transaction false degree of the commodities through a visual interface.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] Through the commodity preliminary prediction unit, the present invention divides the purchase records of each user into multiple independent sequence sets, which is convenient for combining the number of commodity purchases with the corresponding shopping needs at different transaction times, ensuring that the system can preliminarily predict some of the commodities required by each user in the next round of shopping. And in order to further improve the prediction accuracy, the aggregation feature determination unit and the commodity final prediction unit are used to analyze the multiple independent sequence sets included by each user, obtaining the aggregation feature representation and the association feature representation corresponding to each independent sequence set. Since the behaviors of users browsing and collecting commodities are combined in the analysis process, the subsequent prediction of the commodity selection priority according to these feature representations can be more accurate. The system summarizes the commodity numbers predicted twice and pushes the commodity information to the user interface according to the commodity numbers, improving the user experience and increasing the probability of transaction completion at the same time;

[0024] The present invention receives the settlement request sent by the user through the transaction settlement unit and automatically generates a purchase record when the transaction is completed, which is convenient for predicting the needs of the user's next round of shopping. At the same time, when the transaction fails, a collection record will also be automatically generated according to the user information and commodity information. On the one hand, it is convenient for the user to re-settle, and on the other hand, it also supplements the user behavior data to ensure that the setting of the commodity selection priority can be scientific and reasonable. The system uses the false transaction identification unit to analyze the user's purchase records and comment records to prevent the interference of false transactions on the user's commodity selection, quickly identifying the user numbers and commodity numbers involved in these non-compliant transactions, enabling the platform supervisors to discover and take measures to intervene in a timely manner. Description of the Drawings

[0025] Figure 1Schematic diagram of the overall system process provided by an embodiment of the present invention;

[0026] Figure 2 Internal module block diagram of the preliminary commodity prediction unit provided by an embodiment of the present invention;

[0027] Figure 3 Internal module block diagram of the aggregated feature determination unit provided by an embodiment of the present invention;

[0028] Figure 4 Internal module block diagram of the final commodity prediction unit provided by an embodiment of the present invention;

[0029] Figure 5 Internal module block diagram of the transaction settlement unit provided by an embodiment of the present invention;

[0030] Figure 6 Internal module block diagram of the false transaction identification unit provided by an embodiment of the present invention.

[0031] In the figure: 1. Preliminary commodity prediction unit; 101. Purchase record division module; 102. Sequence set generation module; 103. Time feature analysis module; 104. Frequency feature analysis module; 105. Feature conversion module; 106. Commodity number prediction module; 2. Aggregated feature determination unit; 201. Record extraction module; 202. Label setting module; 203. Embedded vector analysis module; 204. Interaction feature optimization module; 205. Feature splicing module; 3. Final commodity prediction unit; 301. Matrix generation module; 302. Correlation feature analysis module; 303. Fusion feature analysis module; 304. Transaction prediction module; 4. Transaction settlement unit; 401. Commodity recommendation module; 402. Price determination module; 403. Commodity settlement module; 5. False transaction identification unit; 501. Comment record screening module; 502. User node storage module; 503. Node marking module; 504. Group scoring module; 505. False degree display module. Detailed implementation manners

[0032] 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 the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0033] Please refer to Figures 1-6 , the present invention provides a technical solution: an e-commerce online settlement system based on transaction behavior analysis, including a false transaction identification unit 5;

[0034] Commodity Preliminary Prediction Unit 1. Commodity Preliminary Prediction Unit 1 extracts the purchase records of all users in the specified e-commerce platform, divides all purchase records according to the user numbers and transaction times in the records to obtain multiple groups, constructs multiple independent sequence sets for each group according to the transaction time, statistically analyzes the time feature vectors corresponding to all independent sequence sets in the group, uses Fourier transform to convert them into discrete features, sets a weight matrix, and then uses the weight matrix and discrete features to analyze the frequency feature vectors of each independent sequence set. After converting the frequency feature vectors of all independent sequence sets into corresponding time feature vectors through inverse Fourier transform, after using the FNN model to analyze the feature representations of multiple independent sequence sets corresponding to each user, statistically analyzes the occurrence times of each commodity in all independent sequence sets corresponding to the user, and uses the occurrence times of the commodity and the feature representations of the independent sequence sets to predict the commodity numbers corresponding to the independent sequence sets when each user makes the next transaction;

[0035] Aggregate Feature Determination Unit 2. Aggregate Feature Determination Unit 2 determines the embedding vectors and interaction label information of the commodities in all independent sequence sets corresponding to each user, combines and analyzes the commodity embedding vectors and interaction label information in the independent sequence sets, so as to determine the interaction time feature vectors corresponding to each independent sequence set at different scales. After converting the interaction time feature vectors into discrete feature vectors, performs a dot product operation on the discrete feature vectors through the weight matrix to obtain the optimized interaction frequency feature vectors, converts the interaction frequency feature vectors into corresponding interaction time feature vectors, and analyzes the aggregate feature representations of multiple independent sequence sets corresponding to each user according to the interaction time feature vectors at different scales;

[0036] Commodity Final Prediction Unit 3. After receiving the commodity numbers and interaction label information in all independent sequence sets corresponding to each user, Commodity Final Prediction Unit 3 analyzes them to obtain the connection matrix corresponding to each user, uses a convolutional network to calculate the connection matrix to obtain the associated feature representations of multiple independent sequence sets corresponding to each user, analyzes the aggregate feature representations and associated feature representations of each independent sequence set through a fusion feature analysis algorithm to obtain the fusion feature representations of each independent sequence set corresponding to each user, and predicts the commodity numbers corresponding to the independent sequence sets when each user makes the next transaction according to the fusion features of each independent sequence set corresponding to each user;

[0037] Transaction Settlement Unit 4. After Transaction Settlement Unit 4 extracts all the predicted commodity numbers for the next transaction of each user, transmits the commodity information to the user recommendation interface according to the predicted commodity numbers, receives the settlement request sent by the user, returns the payment method and process to the user settlement interface, determines the payment completion time and uses it as the transaction time, and generates purchase records according to the user number, commodity number, commodity name and transaction time.

[0038] The initial commodity prediction unit 1 includes a purchase record division module 101, a sequence set generation module 102, and a time feature analysis module 103. The purchase record division module 101 extracts the purchase records of all users within a specified e-commerce platform. The purchase records include user IDs, commodity IDs, commodity names, and transaction times. The purchase records are divided according to the user ID and the transaction time, and the purchase records with the same user ID are stored in the same group. The purchase records within the group are sorted according to the transaction time, so as to obtain multiple sorted groups. The sequence set generation module 102 determines multiple independent sequence sets corresponding to each group according to the transaction time, counts the purchase records with the same transaction time in each group, extracts the commodity IDs in the purchase records with the same transaction time, and then stores them in the corresponding independent sequence sets according to the transaction time. After the time feature analysis module 103 determines the commodity embedding dimension, the total number of different commodities, and the commodity one-hot vector, it calculates the embedding vectors corresponding to the commodities in each independent sequence set according to the embedding dimension, the total number, and the one-hot vector, and performs a max pooling operation on the embedding vectors to obtain the time feature vector of each independent sequence set;

[0039] The initial commodity prediction unit 1 further includes a frequency feature analysis module 104, a feature conversion module 105, and a commodity ID prediction module 106. The frequency feature analysis module 104 counts the time feature vectors corresponding to all independent sequence sets in the group , where is the group number, is the number of independent sequence sets included in the group , and uses the Fourier transform to convert the time feature vectors of all independent sequence sets into the corresponding discrete features , where is the independent sequence set number. After setting two weight matrices and for the group , use , and the discrete feature of the th independent sequence set to calculate the frequency feature vector of this independent sequence set, where , represents the activation function, represents the adjustment coefficient, Denote the parameter. After the feature transformation module 105 repeats the operation until the frequency feature vectors of each independent sequence set in each group are determined, the frequency feature vectors of the independent sequence sets are analyzed using the sequence feature analysis algorithm to obtain the simplified frequency feature vectors. Through the inverse Fourier transform, the simplified frequency feature vectors of all independent sequence sets are converted into corresponding time feature vectors. The product number prediction module 106 transmits the time feature vectors of each independent sequence set in each group to the FNN model for analysis. After obtaining the feature representations of multiple independent sequence sets corresponding to each user, the occurrence times of each product in the independent sequence sets corresponding to the user are counted, and the occurrence times of the product and the feature representations of the independent sequence sets are aggregated and analyzed to obtain the selection priority corresponding to each product. Set the screening number, and predict the product numbers corresponding to the independent sequence sets for each user's next round of transactions according to the selection priority and screening number of the product. The sequence feature analysis algorithm is specifically as follows:

[0040] ;

[0041] Among them, denotes the group the frequency feature vector after simplification of the nth independent sequence set, denotes the group the original frequency feature vector of the nth independent sequence set, denotes the shrinkage parameter, denotes the weight matrix, denotes the group number, denotes the number of the independent sequence set, denotes the parameter, denotes the sign function, denotes the maximum value function;

[0042] The aggregation feature determination unit 2 includes a record extraction module 201, a label setting module 202, an embedded vector analysis module 203, an interaction feature optimization module 204, and a feature splicing module 205. The record extraction module 201 extracts the browsing records and favorite records of all users within a specified e-commerce platform. The browsing records include user numbers, product numbers, product names, and browsing times. The favorite records include user numbers, product numbers, product names, and favorite times. The label setting module 202 divides the interaction types into three types, namely favorite, browse, and purchase, and determines the interaction label information of the products within all independent sequence sets corresponding to each user according to the browsing records and favorite records. The interaction label information includes browse labels and favorite labels. If the transaction time of the product in the independent sequence set is later than the browsing time, the corresponding browse label is set to 1, otherwise it is set to 0. If the transaction time of the product in the independent sequence set is later than the favorite time, the corresponding browse label is set to 1, otherwise it is set to 0. The embedded vector analysis module 203 combines and analyzes the product embedded vectors and interaction label information within all independent sequence sets corresponding to each user to obtain the interaction embedded vectors of the products within the sequence sets, and performs a pooling operation on the interaction embedded vectors of each product to determine the interaction time feature vectors corresponding to each independent sequence set at different scales. The interaction feature optimization module 204 uses Fourier transform to convert the interaction time feature vectors of the independent sequence sets into discrete feature vectors, and then performs a dot product operation on the discrete feature vectors through a weight matrix to obtain the optimized interaction frequency feature vectors, and uses the inverse Fourier transform to convert the optimized interaction frequency feature vectors into the corresponding interaction time feature vectors. The feature splicing module 205 splices the interaction time feature vectors at different scales of each independent sequence set to obtain the aggregation time feature vector of the independent sequence set, and transmits it to the PFFN model for analysis to obtain the aggregation feature representation of multiple independent sequence sets corresponding to each user;

[0043] The final product prediction unit 3 includes a matrix generation module 301, a correlation feature analysis module 302, a fusion feature analysis module 303, and a transaction prediction module 304. After receiving the product number and product interaction label information in all independent sequence sets corresponding to each user, the matrix generation module 301 analyzes the product number and interaction label information in each independent sequence set using a matrix generation algorithm to obtain a relationship matrix and an interaction matrix corresponding to each user. The correlation feature analysis module 302 splices the relationship matrix and the interaction matrix corresponding to each user to generate a connection matrix, and analyzes the connection matrix using a convolutional network to obtain the correlation feature representation of multiple independent sequence sets corresponding to each user. After receiving the aggregated feature representation and the correlation feature representation of multiple independent sequence sets corresponding to each user, the fusion feature analysis module 303 calculates the aggregated feature representation and the correlation feature representation respectively, determines the corresponding dynamic weights of the two, and analyzes the dynamic weights, the aggregated feature representation, and the correlation feature representation of each independent sequence set using a fusion feature analysis algorithm to obtain the fusion feature representation of each independent sequence set corresponding to each user. The transaction prediction module 304 analyzes the product selection priority corresponding to each user based on the fusion feature of each independent sequence set corresponding to each user, and predicts the product number corresponding to the independent sequence set when each user conducts the next round of transactions according to the product selection priority and the screening number. The matrix generation algorithm is specifically:

[0044] ;

[0045] ;

[0046] Among them, represents the relationship matrix of user . represents the product node in the th independent sequence set, represents the relationship edge between products in the th independent sequence set, represents the interaction matrix of user , represents the product node with interaction label information in the th independent sequence set, represents the interaction edge between products in the th independent sequence set, represents the set of all product nodes in the th independent sequence set, represents the set of all product nodes with interaction label information in the th independent sequence set, represents a parameter, represents the number of the independent sequence set, represents the a user, indicating the user ID;

[0047] The fusion feature analysis algorithm is specifically as follows:

[0048] ;

[0049] Among them, represents the fusion feature representation of the th independent sequence set of the user, represents the training matrix, represents the transposed training matrix, represents the parameter matrix, represents the aggregation feature representation of the th independent sequence set of the user, represents the correlation feature representation of the th independent sequence set of the user, represents the user ID, represents the independent sequence set ID, represents the th user;

[0050] The transaction settlement unit 4 includes a commodity recommendation module 401, a price determination module 402, and a commodity settlement module 403. After the commodity recommendation module 401 extracts all the predicted commodity numbers for the next round of transactions of each user, it queries the database according to the predicted commodity numbers, and transmits the obtained commodity information to the user recommendation interface. When the user browses and collects commodities, corresponding records are generated and transmitted to the database. After the price determination module 402 receives the settlement request sent by the user, it determines the actual price of the commodity according to the commodity number, user number, and request time included in the settlement request, returns it to the user settlement interface, and waits for the user to confirm the information. If the commodity settlement module 403 does not receive the confirmation information within the preset time, it determines that the current user has not purchased the commodity, takes the request reception time as the collection time of the commodity by the current user, generates a collection record according to the user number, commodity number, commodity name, and collection time, and stores it in the database. If the confirmation information is received within the preset time, it returns the payment method and process to the user settlement interface, determines the payment completion time, takes it as the transaction time, generates a purchase record according to the user number, commodity number, commodity name, and transaction time, and stores it in the database;

[0051] The false transaction identification unit 5 includes a comment record screening module 501, a user node storage module 502, and a node marking module 503. The comment record screening module 501 extracts the commodity comment records of each user within a specified e-commerce platform. The commodity comment records include user numbers, commodity numbers, comment contents, and comment times. The total number of corresponding purchase records is determined according to the user numbers. If the total number is lower than the threshold, all the commodity comment records of this user are excluded. The total number of comment records is determined according to the commodity numbers. If the total number is lower than the threshold, all the comment records corresponding to this commodity are excluded. After the user node storage module 502 determines the unexcluded commodity comment records and purchase records, the similarity between users is calculated according to the commodity comment records and purchase records. Multiple user nodes are constructed according to the total number of users and user numbers. The distance value between user nodes is determined through the similarity between users. All user nodes are stored in a set. After the node marking module 503 sets the first distance value and the second distance value, an unmarked user node is randomly selected from the set and used as the target node and marked. The distance values between the target node and other nodes are counted. If the distance value is less than the first distance value, the current node is added to the adjacent set of the target node and marked. If the distance value is greater than or equal to the first distance value and less than the second distance value, the current node is added to the adjacent set without being marked. If the distance value is greater than the second distance value, the current node is not added to the adjacent set. The operation is repeated until all nodes in the set are marked;

[0052] The false transaction identification unit 5 further includes a group scoring module 504 and a false degree display module 505. After the group scoring module 504 counts the central points in each adjacent set and uses them as the initial central points, the K-Mean clustering algorithm is used to classify all nodes in the set according to the initial central points, and multiple candidate groups are obtained. The transaction evaluation algorithm is used to analyze each candidate group to determine the transaction time interval variance, compactness, commodity co-occurrence times, and repeated purchase times corresponding to each group. Each group is scored according to the transaction time interval variance, compactness, commodity co-occurrence times, and repeated purchase times. The false degree display module 505 selects the group with the highest score. After determining the commodity scoring records and purchase records corresponding to all user nodes in this group, the transaction false degree of the corresponding commodity is analyzed according to the commodity scoring records and purchase records, and the commodity scoring records, purchase records of users, and the transaction false degree of the commodity are output through a visualization interface. The transaction evaluation algorithm is specifically:

[0053] ;

[0054] Among them, represents the maximum commodity co-occurrence times in the th group, represents the The average number of repeat purchases in a group, Indicates the transaction interval time corresponding to the user, Indicates the average value of the transaction interval time, Indicates the number of user nodes in the group, Indicates the Total number of comment records in the th group, Indicates the Cartesian product of users and products in the th group, Indicates the Co-occurrence times of the th product, Indicates the Number of repeat purchases of the th product, and Indicates the serial number.

[0055] Working principle: In the present invention, the purchase record division module 101 in the preliminary commodity prediction unit 1 obtains multiple sorted groups, the sequence set generation module 102 determines multiple independent sequence sets corresponding to each group, the time feature analysis module 103 obtains the time feature vectors of each independent sequence set, the frequency feature analysis module 104 calculates the frequency feature vectors of the independent sequence sets, the feature conversion module 105 converts the simplified frequency feature vectors into corresponding time feature vectors, the commodity number prediction module 106 predicts the commodity numbers corresponding to the independent sequence sets when each user makes the next transaction. The record extraction module 201 in the aggregation feature determination unit 2 extracts the browsing records and favorite records of all users within the specified e-commerce platform, the label setting module 202 determines the interaction label information of the commodities within all independent sequence sets corresponding to each user, the embedded vector analysis module 203 obtains the interaction time feature vectors of each independent sequence set at different scales, the interaction feature optimization module 204 converts the optimized interaction frequency feature vectors into interaction time feature vectors, the feature splicing module 205 obtains the aggregation feature representation of multiple independent sequence sets corresponding to each user, the matrix generation module 301 in the final commodity prediction unit 3 outputs the relationship matrix and interaction matrix of each user, the associated feature analysis module 302 obtains the associated feature representation of multiple independent sequence sets corresponding to each user, the fusion feature analysis module 303 analyzes the aggregation feature representation and the associated feature representation of each independent sequence set to obtain the fusion feature representation of each independent sequence set corresponding to each user, the transaction prediction module 304 predicts the commodity numbers corresponding to the independent sequence sets when each user makes the next transaction, the commodity recommendation module 401 in the transaction settlement unit 4 queries in the database according to the predicted commodity numbers, and transmits the obtained commodity information to the user recommendation interface, the price determination module 402 determines the actual price of the commodity and returns it to the user settlement interface, the commodity settlement module 403 records the payment completion time. The comment record screening module 501 in the false transaction identification unit 5 extracts the commodity comment records of each user within the specified e-commerce platform, the user node storage module 502 constructs multiple user nodes according to the total number of users and user numbers, determines the distance values between user nodes through the similarity between users, the node marking module 503 constructs a corresponding adjacent set for each target node, the group scoring module 504 scores each group according to the variance of the transaction time interval, compactness, commodity co-occurrence times, and repeated purchase times, and the false degree display module 505 outputs the commodity scoring records, purchase records of the user, and the transaction false degree of the commodity through a visualization interface.

[0056] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device.

[0057] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An e-commerce online settlement system based on transaction behavior analysis, comprising a false transaction identification unit (5), characterized in that: A commodity preliminary prediction unit (1), wherein the commodity preliminary prediction unit (1) extracts purchase records of all users in a designated e-commerce platform, divides all purchase records according to user numbers and transaction times in the records to obtain multiple groups, constructs multiple independent sequence sets for each group according to the transaction time, counts time feature vectors corresponding to all independent sequence sets in the group, converts them into discrete features using Fourier transform, sets a weight matrix, analyzes frequency feature vectors of each independent sequence set using the weight matrix and discrete features, converts the frequency feature vectors of all independent sequence sets into corresponding time feature vectors using inverse Fourier transform, analyzes feature representations of multiple independent sequence sets corresponding to each user using an FNN model, counts the number of occurrences of each commodity in all independent sequence sets corresponding to the user, and predicts the commodity number corresponding to the independent sequence set of each user in the next round of transactions using the number of occurrences of the commodity and the feature representation of the independent sequence set; An aggregate feature determination unit (2), wherein the aggregate feature determination unit (2) determines the embedding vectors and interaction label information of the products in all independent sequence sets corresponding to each user, combines and analyzes the embedding vectors and interaction label information of the products in the independent sequence sets, thereby determining the interaction time feature vectors corresponding to each independent sequence set at different scales, converts the interaction time feature vectors into discrete feature vectors, performs a dot multiplication operation on the discrete feature vectors through a weight matrix, obtains an optimized interaction frequency feature vector, converts the interaction frequency feature vectors into corresponding interaction time feature vectors, and analyzes the aggregate feature representations of multiple independent sequence sets corresponding to each user based on the interaction time feature vectors at different scales; The commodity final prediction unit (3) comprises a matrix generation module (301), an associated feature analysis module (302), a fusion feature analysis module (303) and a transaction prediction module (304). After receiving the commodity numbers and commodity interaction tag information in all independent sequence sets corresponding to each user, the matrix generation module (301) uses a matrix generation algorithm to analyze the commodity numbers and interaction tag information in each independent sequence set to obtain a relationship matrix and an interaction matrix corresponding to each user. The associated feature analysis module (302) performs a splicing operation on the relationship matrix and the interaction matrix corresponding to each user to generate a connection matrix. The connection matrix is ​​analyzed using a convolutional network to obtain a plurality of independent sequence sets corresponding to each user. The fusion feature analysis module (303) receives the aggregate feature representation and the associated feature representation of the multiple independent sequence sets corresponding to each user, and then respectively calculates the aggregate feature representation and the associated feature representation, and after determining the dynamic weights corresponding to the two, analyzes the dynamic weight, the aggregate feature representation and the associated feature representation of each independent sequence set through the fusion feature analysis algorithm to obtain the fusion feature representation of each independent sequence set corresponding to each user. The transaction prediction module (304) analyzes the commodity selection priority corresponding to each user according to the fusion features of each independent sequence set corresponding to each user, and predicts the commodity number corresponding to the independent sequence set of each user in the next round of transactions according to the commodity selection priority and the screening number. The matrix generation algorithm is specifically as follows: ； ； in, Indicates user The relationship matrix, Indicates The product nodes in the independent sequence set, Indicates The relationship edge between items in the independent sequence set, Indicates user The interaction matrix, Indicates The commodity nodes with interactive label information in the independent sequence set, Indicates The interaction edges between the products in the independent sequence sets, Indicates The set of all commodity nodes in an independent sequence set, Indicates The set of all commodity nodes with interactive label information in an independent sequence set, Indicates the parameters, represents the number of the independent sequence set, Indicates Users, Indicates the user's ID; The fusion feature analysis algorithm is specifically as follows: ； in, Indicates user No. The fusion feature representation of independent sequence sets is represents the training matrix, represents the transposed training matrix, represents the parameter matrix, Indicates user No. The aggregate feature representation of a set of independent sequences, Indicates user No. The associated feature representation of a set of independent sequences, Indicates the user's ID. represents the number of the independent sequence set, Indicates Users; A transaction settlement unit (4), wherein the transaction settlement unit (4) extracts all predicted product numbers for the next round of transactions of each user, transmits product information to the user recommendation interface according to the predicted product numbers, receives a settlement request from the user, returns the payment method and process to the user settlement interface, determines the payment completion time, uses it as the transaction time, and generates a purchase record based on the user number, product number, product name and transaction time.

2. The e-commerce online settlement system based on transaction behavior analysis according to claim 1, characterized in that: The commodity preliminary prediction unit (1) comprises a purchase record division module (101), a sequence set generation module (102) and a time feature analysis module (103). The purchase record division module (101) extracts purchase records of all users in a specified e-commerce platform, wherein the purchase records include a user number, a commodity number, a commodity name and a transaction time, divides all purchase records according to the user number and the transaction time, stores purchase records of the same user number in the same group, and sorts the purchase records in the group according to the transaction time, thereby obtaining a plurality of sorted groups. The sequence set generation module (102) determines a plurality of independent sequence sets corresponding to each group according to the transaction time, counts purchase records with the same transaction time in each group, extracts commodity numbers in purchase records with the same transaction time, and stores them in the corresponding independent sequence sets according to the transaction time. The time feature analysis module (103) determines the commodity embedding dimension, the total number of different commodities and the commodity unique hot vector, and then calculates the embedding vector corresponding to the commodity in each independent sequence set according to the embedding dimension, the total number and the unique hot vector, performs a maximum pooling operation on the embedding vector, and obtains the time feature vector of each independent sequence set.

3. The e-commerce online settlement system based on transaction behavior analysis according to claim 2 is characterized by: The commodity preliminary prediction unit (1) further comprises a frequency feature analysis module (104), a feature conversion module (105) and a commodity number prediction module (106). The frequency feature analysis module (104) performs statistical grouping. The time feature vectors corresponding to all independent sequence sets in ,in is the group number, For grouping The number of independent sequence sets contained in , using Fourier transform to convert the time feature vectors of all independent sequence sets Convert to corresponding discrete features ,in Number independent sequence sets and set groups The two weight matrices and After that, use , and Discrete features of a set of independent sequences Calculate the frequency feature vector of the independent sequence set ,in , represents the activation function, represents the adjustment factor, The feature conversion module (105) repeats the operation until the frequency feature vectors of each independent sequence set in each group are determined, and then uses a sequence feature analysis algorithm to analyze the frequency feature vectors of the independent sequence set to obtain a simplified frequency feature vector. The simplified frequency feature vectors of all independent sequence sets are converted into corresponding time feature vectors by inverse Fourier transform. The product number prediction module (106) transmits the time feature vectors of each independent sequence set in each group to the FNN model for analysis. After obtaining the feature representations of multiple independent sequence sets corresponding to each user, the number of occurrences of each product in the independent sequence set corresponding to the user is counted, and the number of occurrences of the product and the feature representation of the independent sequence set are aggregated and analyzed to obtain the selection priority corresponding to each product, set the number of screening, and predict the product number corresponding to the independent sequence set of each user in the next round of transactions according to the selection priority and the number of screening.

4. The e-commerce online settlement system based on transaction behavior analysis according to claim 1, characterized in that: The aggregate feature determination unit (2) comprises a record extraction module (201), a label setting module (202), an embedded vector analysis module (203), an interactive feature optimization module (204) and a feature splicing module (205). The record extraction module (201) extracts browsing records and collection records of all users in a specified e-commerce platform. The browsing records include a user number, a product number, a product name and a browsing time. The collection records include a user number, a product number, a product name and a collection time. The label setting module (202) divides the interaction types into three types, namely collection, browsing and purchase. The interactive label information of the products in all independent sequence sets corresponding to each user is determined based on the browsing records and the collection records. The interactive label information includes a browsing label and a collection label. If the transaction time of the product in the independent sequence set is later than the browsing time, the corresponding browsing label is set to 1, otherwise it is set to 0. If the transaction time of the product in the independent sequence set is later than the collection time, the corresponding collection label is set to 1. Otherwise, it is set to 0. The embedding vector analysis module (203) combines and analyzes the product embedding vectors and interaction label information in all independent sequence sets corresponding to each user to obtain the interaction embedding vectors of the products in the sequence set, and performs a pooling operation on the interaction embedding vectors of each product to determine the interaction time feature vectors corresponding to each independent sequence set at different scales. The interaction feature optimization module (204) uses Fourier transform to convert the interaction time feature vector of the independent sequence set into a discrete feature vector, and then performs a dot multiplication operation on the discrete feature vector through a weight matrix to obtain an optimized interaction frequency feature vector. The optimized interaction frequency feature vector is converted into a corresponding interaction time feature vector using an inverse Fourier transform. The feature splicing module (205) performs a splicing operation on the interaction time feature vectors at different scales of each independent sequence set to obtain an aggregated time feature vector of the independent sequence set, which is transmitted to the PFFN model for analysis to obtain an aggregated feature representation of multiple independent sequence sets corresponding to each user.

5. The e-commerce online settlement system based on transaction behavior analysis according to claim 1, characterized in that: The transaction settlement unit (4) includes a product recommendation module (401), a price determination module (402) and a product settlement module (403). After the product recommendation module (401) extracts all predicted product numbers of each user's next round of transactions, it searches the database according to the predicted product numbers, and transmits the obtained product information to the user recommendation interface. When the user browses and collects products, a corresponding record is generated and transmitted to the database. After the price determination module (402) receives a settlement request from the user, it determines the actual price of the product according to the product number, user number and request time contained in the settlement request, and sends it to the database. Return to the user settlement interface and wait for user confirmation information. If the product settlement module (403) does not receive confirmation information within the preset time, it is determined that the current user has not purchased the product, and the request reception time is used as the current user's collection time for the product. A collection record is generated according to the user number, product number, product name and collection time, and stored in the database. If a confirmation message is received within the preset time, the payment method and process are returned to the user settlement interface, the payment completion time is determined, and it is used as the transaction time. A purchase record is generated according to the user number, product number, product name and transaction time, and stored in the database.

6. The e-commerce online settlement system based on transaction behavior analysis according to claim 1, characterized in that: The false transaction identification unit (5) comprises a comment record screening module (501), a user node storage module (502) and a node marking module (503). The comment record screening module (501) extracts the product comment record of each user in the specified e-commerce platform, wherein the product comment record includes the user number, the product number, the comment content and the comment time, determines the total number of corresponding purchase records according to the user number, and if the total number is lower than a threshold, all the product comment records of the user are removed, determines the total number of comment records according to the product number, and if the total number is lower than the threshold, all the comment records corresponding to the product are removed, and after the user node storage module (502) determines the product comment records and purchase records that have not been removed, calculates the user's purchase history according to the product comment records and purchase records. The similarity between the two nodes is determined, and multiple user nodes are constructed according to the total number of users and the user numbers. The distance values ​​between the user nodes are determined according to the similarity between the users, and all the user nodes are stored in a set. After the node marking module (503) sets the first distance value and the second distance value, an unmarked user node is randomly selected from the set, and the node is used as a target node and marked. The distance values ​​between the target node and other nodes are counted. If the distance value is less than the first distance value, the current node is added to the adjacent set of the target node and marked. If the distance value is greater than or equal to the first distance value and less than the second distance value, the current node is added to the adjacent set without marking. If the distance value is greater than the second distance value, the current node is not added to the adjacent set. The operation is repeated until all nodes in the set are marked.

7. The e-commerce online settlement system based on transaction behavior analysis according to claim 6, characterized in that: The false transaction identification unit (5) further comprises a group scoring module (504) and a false degree display module (505). The group scoring module (504) counts the center points in each adjacent set and uses the center points as the initial center points. The K-Mean clustering algorithm is used to classify all nodes in the set according to the initial center points to obtain multiple candidate groups. The transaction evaluation algorithm is used to analyze each candidate group to determine the transaction time interval variance, closeness, commodity co-occurrence times and repeated purchase times corresponding to each group. Each group is scored according to the transaction time interval variance, closeness, commodity co-occurrence times and repeated purchase times. The false degree display module (505) selects the group with the highest score, determines the commodity scoring records and purchase records corresponding to all user nodes in the group, analyzes the transaction false degree of the corresponding commodity according to the commodity scoring records and purchase records, and outputs the user's commodity scoring records, purchase records and commodity transaction false degree through a visual interface.

Images