E-commerce transaction anomaly detection big data model system based on time sequence analysis

Abstract

The invention relates to the technical field of e-commerce transactions, in particular to a time sequence analysis-based e-commerce transaction anomaly detection big data model system, which comprises an e-commerce data acquisition module, a multi-dimensional user portrait generation module, a time sequence weighted portrait fusion module and an abnormal order judgment module, and the multi-dimensional user portrait generation module constructs a plurality of user portraits for each user according to the e-commerce data between each transaction order of the user, so that the one-sidedness of a single portrait is avoided, a rich data basis is provided for subsequent accurate analysis, and the user experience is improved. Then, the time sequence weighted portrait fusion module sets different portrait weights according to the sequential time sequence of the orders, finally, a current user portrait is constructed by adopting a weighted fusion algorithm, and the weighted fusion algorithm enables the portrait weights corresponding to the orders closer to the current time to be higher; the recent behavior can better reflect the current real demand and behavior mode of the user, so that the description accuracy of the portrait on the current state of the user is improved.

Description

Technical Field

[0001] The present invention relates to the field of e-commerce transaction technology, and in particular to an e-commerce transaction anomaly detection big data model system based on time series analysis. Background Art

[0002] Driven by the wave of digitalization, e-commerce platforms have flourished, with explosive growth in the number of users and transaction volume. However, this is accompanied by an increase in abnormal transaction behaviors, which not only damage the tangible interests of consumers and merchants, but also pose a serious threat to the reputation and healthy development of e-commerce platforms.

[0003] At present, most e-commerce transaction anomaly detection systems integrate users' e-commerce data, build user portraits, and then set user portrait thresholds to determine whether there are anomalies when users place new orders. However, in the actual shopping process, as time goes by, users' behaviors on e-commerce platforms continue to change, and the generated e-commerce data is also dynamically updated. This makes the user portrait always in a state of continuous evolution. During this change process, if the user portrait cannot be adjusted in a timely and effective manner, the threshold set based on the old portrait will be difficult to accurately adapt to the new user behavior pattern, resulting in deviations in the judgment of abnormal orders, which will allow some abnormal transaction behaviors to escape monitoring and bring potential risks to the platform and users. In view of this, we propose an e-commerce transaction anomaly detection big data model system based on time series analysis. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of not being able to adjust user portraits in time.

[0005] To achieve the above-mentioned purpose, the present invention provides an e-commerce transaction anomaly detection big data model system based on time series analysis, including an e-commerce data collection module, a multi-dimensional user portrait generation module, a time series weighted portrait fusion module and an abnormal order judgment module;

[0006] The e-commerce data acquisition module uses the access token to send a connection request to the e-commerce database to establish a connection with the e-commerce database, and obtains the e-commerce data of each user in the e-commerce database. The multi-dimensional user portrait generation module uses the K-Means clustering algorithm and multiple orders of each user in the e-commerce data to construct multiple user portraits corresponding to the user, and generates a user portrait set according to the order time sequence of the multiple user portraits corresponding to each user;

[0007] The time-series weighted portrait fusion module uses a linear decay method to generate and set different portrait weights. The linear decay method sets the portrait weight according to the order time sequence, sorts the order time in order, and the portrait constructed by the e-commerce data corresponding to the order closer to the current time is given a higher weight;

[0008] When a new order is added, the abnormal order judgment module senses the user portrait corresponding to the new order in the multi-dimensional user portrait generation module, defines it as a judgment portrait, calculates the similarity between the current user portrait and the judgment portrait, and sets the user portrait threshold interval by the threshold setting method. If the similarity is not within the user portrait threshold interval, the new transaction order is determined to be an abnormal order, and the abnormal order is fed back to the staff;

[0009] When judging whether a new order is an abnormal order, the abnormal order judgment module matches and calculates whether there is a product corresponding to the new order in the products corresponding to the historical orders. If there is a corresponding product, the browsing record of the product corresponding to the new order is called out and defined as a browsing record collection. The most comprehensive browsing record of the e-commerce data in the browsing record collection is matched, and the most comprehensive browsing record is replaced with the browsing record corresponding to the new order, and then the current user portrait is calculated through the time series weighted portrait fusion module.

[0010] As a further improvement of the technical solution, the working principle of establishing a connection with the e-commerce database in the e-commerce data collection module is as follows:

[0011] The e-commerce data collection module completes the registration process with the open platform of the e-commerce platform, submits detailed corporate or personal information, and clarifies the purpose and scope of obtaining e-commerce data. After the open platform reviews and approves it, a unique access token is allocated to the e-commerce data collection module;

[0012] The e-commerce data collection module uses the network protocol to send the access token to the e-commerce database server in the form of a data packet. During the transmission process, the data packet passes through multiple network nodes and finally reaches the target server;

[0013] After receiving the access token, the e-commerce database server calls out the pre-stored verification information to verify the access token. If the verification information is inconsistent with the access token, the server determines that the verification has failed and returns the verification failure information to the e-commerce data collection module; if the two are consistent, the verification is successful, and the server returns the verification success information to the e-commerce data collection module, and the e-commerce data collection module establishes a connection with the e-commerce database.

[0014] As a further improvement of this technical solution, the working principle of the e-commerce data collection module to obtain the e-commerce data of each user in the e-commerce database is as follows:

[0015] The e-commerce data collection module constructs a query statement according to the demand for obtaining e-commerce data, and sends the query statement to the database engine through the established connection. The database engine searches the database according to the query statement, retrieves the e-commerce data that meets the conditions, and outputs the e-commerce data to the e-commerce data collection module through the established connection.

[0016] As a further improvement of the present technical solution, when the multi-dimensional user portrait generation module uses the K-Means algorithm to construct the user portrait system, users with similar behavioral characteristics are grouped into one category. The specific working steps are as follows:

[0017] Step 1: Preset different numbers of clusters , number of clusters The number of different types of user portraits;

[0018] Step 2: Randomly select from e-commerce protective gear E-commerce data is used as the initial centroid, and each centroid is a vector with the values ​​of these three dimensions;

[0019] Step 3: For each data in the e-commerce data, calculate the Euclidean distance between each data and The distance between the centroids is used to assign the e-commerce data to the cluster represented by the nearest centroid.

[0020] Step 4: After allocating all e-commerce data to corresponding clusters, recalculate the centroid position of each cluster;

[0021] For a cluster , its centroid The calculation method is to take the mean of the characteristic values ​​of each dimension of all e-commerce data in the cluster. There are E-commerce data , each e-commerce data , then the centroid No. Dimensional coordinates for: ;

[0022] Step 5: Repeat steps 3 and 4, continuously assigning e-commerce data to clusters and updating the centroid position until the centroid position no longer changes significantly.

[0023] As a further improvement of this technical solution, the linear attenuation method in the time series weighted portrait fusion module perceives the total number of e-commerce data of the same user. Order No. The portrait weight corresponding to each order The calculation formula is: ;

[0024] Perceive the weight of the portrait and merge multiple user portraits with different weights through the weighted fusion method to generate the current user portrait;

[0025] Weighted fusion method perceives the same user's shared Different user profiles , the corresponding weights calculated by the linear attenuation method are ,and , profile each user Represented as a feature vector ,in is the dimension of the feature;

[0026] Current user portrait generated by weighted fusion method Also a The eigenvector of dimension characteristic The calculation formula is: .

[0027] As a further improvement of the technical solution, the abnormal order judgment module calculates the similarity between the current user portrait and the judgment portrait as follows:

[0028] Perceive the current user profile , the image is determined to be , then the similarity between the two is: .

[0029] As a further improvement of the technical solution, the threshold setting method in the abnormal order judgment module sets the user portrait threshold interval according to the change rule between multiple user portraits of the same user in the e-commerce data collection module. The specific calculation formula is as follows:

[0030] Perception User Profile Collection , which contains User portraits, each with feature dimension, for the Feature Dimensions , and its corresponding eigenvalue set is ;

[0031] Calculate the mean: The mean of the feature dimensions The calculation formula is ;

[0032] Calculate the standard deviation: The standard deviation of the feature dimension The calculation formula is ;

[0033] The first The user portrait threshold range of each feature dimension is: ,in is an initial constant determined by cross-validation method.

[0034] As a further improvement of the technical solution, the cross-validation method senses the current user portraits of different users and the current user portrait of the same user in the time-series weighted portrait fusion module, defines the current user portraits of different users as abnormal orders, and defines the user portraits of the same user as normal orders;

[0035] By randomly setting the initial constant When judging that the current user profiles of different users are normal orders, the step length Adjust initial constants Similarly, if the initial constant When the user profile of the same user is judged to be an abnormal order, the synchronization step Adjust initial constants ;

[0036] Perception unadjusted initial constant When the user portrait threshold interval is used to judge abnormal orders, the accuracy is , through the step length Adjustment constant Accuracy of determining abnormal orders ,like , then again through the step size Adjust initial constants Otherwise, change the adjustment direction until the accuracy cannot be improved.

[0037] As a further improvement of the technical solution, the working principle of the abnormal order judgment module to match and calculate whether there is a new order corresponding product in the historical order corresponding products is as follows:

[0038] The e-commerce data of each user in the e-commerce data collection module contains product numbers corresponding to different order products and browsing records corresponding to different products. The product numbers of the products corresponding to the newly added orders are compared with the product numbers of the products corresponding to the historical orders one by one. When the product numbers are exactly the same, the two products are the same products. Similarly, the most comprehensive browsing records in the browsing record collection are compared in the same way.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] In the e-commerce transaction anomaly detection big data model system based on time series analysis, the multi-dimensional user portrait generation module constructs multiple user portraits for each user based on the e-commerce data between each transaction order of the user. Constructing multiple user portraits can comprehensively capture user behavior characteristics from different transaction stages and different business scenarios, avoid the one-sidedness of a single portrait, and provide a rich data basis for subsequent precise analysis. Then the time series weighted portrait fusion module sets different portrait weights according to the time series of the orders, and finally uses the weighted fusion algorithm to construct the current user portrait. The weighted fusion algorithm sets a higher portrait weight corresponding to the order closer to the current time, because recent behavior can better reflect the user's current real needs and behavior patterns, thereby improving the accuracy of the portrait's description of the user's current status;

[0041] The abnormal order judgment module sets the user portrait interval to determine whether the new order is an abnormal order. Before judging the abnormal order, the abnormal order judgment module calculates whether there is a product corresponding to the new order in the products corresponding to the historical orders through matching. If there is a corresponding product, the most comprehensive browsing record in the historical orders is adjusted to the browsing record of the new order. By comparing the products of historical orders and new orders, the correlation and consistency of user purchasing behavior can be further explored, providing more dimensional references for judging anomalies. On the other hand, adjusting the most comprehensive browsing records to new orders can enrich the information dimension of new orders, so that when judging abnormal orders in the future, analysis can be performed based on more complete user behavior data, thereby improving the accuracy and reliability of anomaly detection and reducing misjudgments and missed judgments. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the overall module principle diagram of the present invention;

[0043] Figure 2 This is a schematic diagram of a multi-dimensional user portrait generation module of the present invention;

[0044] Figure 3 This is a schematic diagram of the time-series weighted image fusion module of the present invention;

[0045] Figure 4 It is a schematic diagram of the abnormal order judgment module of the present invention.

[0046] The meaning of each number in the figure is:

[0047] 100. E-commerce data collection module; 200. Multi-dimensional user portrait generation module; 300. Time series weighted portrait fusion module; 400. Abnormal order judgment module. DETAILED DESCRIPTION

[0048] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] The e-commerce transaction anomaly detection big data model system based on time series analysis includes an e-commerce data collection module 100, a multi-dimensional user portrait generation module 200, a time series weighted portrait fusion module 300 and an abnormal order judgment module 400;

[0050] During the operation of the e-commerce platform, in order to realize various business functions such as order processing, inventory management, user services, etc., it is necessary to record the e-commerce data of each user in the e-commerce platform database. Specifically, from the user registration, the e-commerce platform will collect the user's basic information, such as user name, password, contact information, etc. When the user performs shopping operations, it will also record the order number, product information, transaction amount, transaction time and other data;

[0051] The e-commerce data collection module 100 uses the access token to send a connection request to the e-commerce database to establish a connection with the e-commerce database, and obtains the e-commerce data of each user from the e-commerce database;

[0052] The e-commerce data collection module 100 completes the registration process with the open platform of the e-commerce platform, submits detailed corporate or personal information, and clarifies the purpose and scope of obtaining e-commerce data. After the open platform reviews and approves it, it allocates a unique access token to the e-commerce data collection module 100;

[0053] The e-commerce data collection module 100 uses the network protocol to send the access token to the e-commerce database server in the form of a data packet. During the transmission process, the data packet passes through multiple network nodes and finally reaches the target server;

[0054] After receiving the access token, the e-commerce database server calls out the pre-stored verification information to verify the access token. If the verification information is inconsistent with the access token, the server determines that the verification has failed, and returns the verification failure information to the e-commerce data collection module 100; if the two are consistent, the verification is successful, and the server returns the verification success information to the e-commerce data collection module 100, and the e-commerce data collection module 100 establishes a connection with the e-commerce database;

[0055] The working principle of the e-commerce data collection module 100 to obtain the e-commerce data of each user in the e-commerce database is as follows:

[0056] The e-commerce data collection module 100 constructs a query statement according to the need to obtain e-commerce data, and sends the query statement to the database engine through the established connection (the database engine is the core component of the database management system, responsible for receiving, parsing and executing query statements). The database engine searches the database according to the query statement, calls out the e-commerce data that meets the conditions, and outputs the e-commerce data to the e-commerce data collection module 100 through the established connection. The above-mentioned query statement construction is reminiscent of those skilled in the art and will not be elaborated here.

[0057] The multi-dimensional user portrait generation module 200 uses the K-Means clustering algorithm and multiple orders of each user in the e-commerce data to construct multiple user portraits corresponding to the user, and generates a user portrait set according to the order time sequence of the multiple user portraits corresponding to each user;

[0058] The multi-dimensional user portrait generation module 200 uses the K-Means clustering algorithm to group users with similar behavioral characteristics into one category. For example, users with similar characteristics such as the types of goods browsed, purchase frequency, and average order value are grouped into one cluster, and each cluster represents a type of user portrait, such as high-spending fashion users and cost-effective users.

[0059] When the multi-dimensional user portrait generation module 200 uses the K-Means algorithm to construct the user portrait system, users with similar behavioral characteristics are clustered into one category, which is the number of clusters in the K-Means algorithm. For example, users with similar characteristics such as browsing product categories, purchase frequency, and customer unit price in e-commerce data are divided into a cluster, and each cluster is recorded as ;

[0060] Step 1: Preset different numbers of clusters , number of clusters Represents the number of different types of user portraits, such as high-consumption fashion user portraits, cost-effective user portraits, etc.

[0061] Step 2: Randomly select from e-commerce protective gear The e-commerce data is used as the initial centroid. Assuming that the user data is represented by a three-dimensional vector, which is the number of browsed product types, purchase frequency, and average order value, then each centroid is a vector with these three dimensional values.

[0062] Step 3: For each data in the e-commerce data, calculate its Euclidean distance The distance between the centroids is used to assign the e-commerce data to the cluster represented by the nearest centroid.

[0063] Euclidean distance is used to calculate the distance between e-commerce data and the centroid. Assume there are two Wei's e-commerce data and , and The Euclidean distance between for:

[0064] ;

[0065] In the user portrait scenario, if we take the three features of browsed product types, purchase frequency, and average order value as examples, and These are two different user e-commerce data. They represent the number of products browsed by the user, the purchase frequency, and the average order value. Similarly, the user The corresponding features of the above can be used to calculate the distance between two users, and then determine the distance between the user and the center of mass.

[0066] Step 4: After all e-commerce data are assigned to corresponding clusters, the centroid position of each cluster is recalculated. The new centroid is the mean of all e-commerce data in the cluster.

[0067] For a cluster , its centroid The calculation method is to take the mean of the characteristic values ​​of each dimension of all e-commerce data in the cluster. Let cluster There are E-commerce data , each e-commerce data , then the centroid No. Dimensional coordinates for:

[0068] ;

[0069] For example, for user data in a cluster, the value of the browsed product category dimension of its new centroid is calculated by adding the number of browsed product categories of all users in the cluster and dividing it by the number of users;

[0070] Step 5: Repeat steps 3 and 4, continuously assigning e-commerce data to clusters and updating the centroid position until the centroid position no longer changes significantly or the preset number of iterations is reached. At this point, the clustering results are relatively stable, the algorithm converges, and each cluster contains e-commerce data with similar characteristics, thereby constructing multiple user portraits corresponding to each user.

[0071] In order to make the constructed portrait more in line with the user's real-time behavioral changes, the time-series weighted portrait fusion module 300 uses a linear decay method to generate different portrait weights. The linear decay method sets the portrait weight according to the time series of the orders, and sorts the order times in chronological order. The portrait constructed by the e-commerce data corresponding to the orders closer to the current time is assigned a higher weight, because recent behavior can better reflect the user's current real needs and behavior patterns; and the portrait weight corresponding to the orders with a longer time is relatively low.

[0072] The linear attenuation method perceives the total number of e-commerce data of the same user Order No. orders (counting from the order closest to the current time, ) Corresponding image weight The calculation formula is: ;

[0073] The time-series weighted portrait fusion module 300 senses the portrait weight, merges multiple user portraits with different weights through a weighted fusion method, and generates a current user portrait. The current user portrait provides strong data support for personalized recommendations, precision marketing and other businesses of the e-commerce platform;

[0074] Weighted fusion method perceives the same user's shared Different user profiles , the corresponding weights calculated by the linear attenuation method are ,and , profile each user Represented as a feature vector ,in is the dimension of the feature;

[0075] Current user portrait generated by weighted fusion method Also a The eigenvector of dimension characteristic The calculation formula is: .

[0076] The abnormal order judgment module 400 is used to perceive the newly added transaction orders in the e-commerce data collection module 100 and the user portraits corresponding to the newly added orders in the multi-dimensional user portrait generation module 200, and define them as judgment portraits;

[0077] Perceiving the current user portrait constructed in the temporal weighted portrait fusion module 300, and calculating the similarity between the current user portrait and the judgment portrait;

[0078] Perceive the current user profile , the image is determined to be , then the similarity between the two is: ;

[0079] The user portrait threshold interval is set according to the threshold setting method. If the similarity is not within the user portrait threshold interval, the new transaction order is determined to be an abnormal order, and the abnormal order is fed back to the staff.

[0080] The threshold setting method sets the user portrait threshold interval according to the change rules between multiple user portraits of the same user in the e-commerce data collection module 100;

[0081] The calculation formula for setting the user portrait threshold interval by the threshold setting method is as follows:

[0082] Perception User Profile Collection , which contains User portraits, each with feature dimension, for the Feature Dimensions , and its corresponding eigenvalue set is ;

[0083] Calculate the mean: The mean of the feature dimensions The calculation formula is .

[0084] Calculate the standard deviation: The standard deviation of the feature dimension The calculation formula is ;

[0085] The first The user portrait threshold range of each feature dimension is: ,in is an initial constant, determined by cross-validation method;

[0086] The cross-validation method perceives the current user portraits of different users in the time-series weighted portrait fusion module 300 and the current user portrait of the same user, and defines the current user portraits of different users as abnormal orders, and the user portraits of the same user as normal orders;

[0087] By randomly setting the initial constant When judging that the current user profiles of different users are normal orders, the step length Adjust initial constants Similarly, if the initial constant When the user profile of the same user is judged to be an abnormal order, the synchronization step Adjust initial constants ;

[0088] Perception unadjusted initial constant When the user portrait threshold interval is used to judge abnormal orders, the accuracy is , through the step length Adjustment constant Accuracy of determining abnormal orders ,like , then again through the step size Adjust initial constants Otherwise, change the adjustment direction until the accuracy cannot be improved.

[0089] When constructing the current user portrait, if there is a product corresponding to the newly added order in the historical orders, it means that the user has known about the product before the shopping. At this time, the e-commerce data corresponding to the newly added order has particularity. If it cannot be reasonably adjusted, the time-series weighted portrait fusion module 300 uses a linear attenuation method to generate different portrait weights. If the e-commerce data corresponding to the newly added order product is selected, it will cause the time-series weighted portrait fusion module 300 to deviate when constructing the current user portrait.

[0090] When the abnormal order judgment module 400 judges whether the newly added order is an abnormal order through the user portrait threshold interval, it matches and calculates whether there is a product corresponding to the newly added order in the products corresponding to the historical orders. If there is a corresponding product, it calls out the browsing record of the product corresponding to the newly added order, which is defined as a browsing record collection, matches the most comprehensive browsing record of e-commerce data in the browsing record collection, replaces the most comprehensive browsing record with the browsing record corresponding to the newly added order, and then calculates the current user portrait through the time series weighted portrait fusion module 300;

[0091] In the e-commerce data collection module 100, each user's e-commerce data contains product numbers corresponding to different order products and browsing records corresponding to different products. The product numbers of the products corresponding to the newly added orders are compared with the product numbers of the products corresponding to the historical orders. If the product numbers are exactly the same, the two products are the same product.

[0092] Compare browsing records synchronously one by one, including how long the user stayed on the product details page, how long they watched each picture or video, etc.; whether the user's scrolling behavior during browsing is recorded, such as the scrolling distance and number of times; whether the subsequent operations of the user after adding to favorites or shopping carts are recorded, such as whether they viewed it again, whether they finally purchased it, etc.

[0093] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An e-commerce transaction anomaly detection big data model system based on time series analysis, characterized by: It comprises an e-commerce data collection module (100), a multi-dimensional user portrait generation module (200), a time series weighted portrait fusion module (300) and an abnormal order judgment module (400); The e-commerce data collection module (100) uses an access token to send a connection request to the e-commerce database to establish a connection with the e-commerce database, and obtains the e-commerce data of each user in the e-commerce database. The multi-dimensional user portrait generation module (200) uses a K-Means clustering algorithm to construct multiple user portraits for each user. The multiple user portraits are e-commerce data corresponding to each user in the e-commerce data corresponding to an order. The multiple user portraits corresponding to each user are combined to generate a user portrait set according to the order time sequence. The time-series weighted portrait fusion module (300) uses a linear decay method to generate and set different portrait weights. The linear decay method sets the portrait weights according to the order time sequence, sorts the order time in order, and the portrait constructed by the e-commerce data corresponding to the order closer to the current time is assigned a higher weight; When a new order is added, the abnormal order judgment module (400) senses the user portrait corresponding to the new order in the multi-dimensional user portrait generation module (200), defines it as a judgment portrait, calculates the similarity between the current user portrait and the judgment portrait, sets the user portrait threshold interval by a threshold setting method, and if the similarity is not within the user portrait threshold interval, determines that the new transaction order is an abnormal order, and feeds back the abnormal order to the staff; When judging whether a newly added order is an abnormal order, the abnormal order judgment module (400) matches and calculates whether there is a product corresponding to the newly added order in the products corresponding to the historical orders. If there is a corresponding product, the browsing record of the product corresponding to the newly added order is retrieved and defined as a browsing record collection. The most comprehensive browsing record of the e-commerce data in the browsing record collection is matched, and the most comprehensive browsing record is replaced with the browsing record corresponding to the newly added order. Then, the current user portrait is calculated through the time-series weighted portrait fusion module (300).

2. The big data model system for e-commerce transaction anomaly detection based on time series analysis according to claim 1 is characterized by: The working principle of establishing a connection with the e-commerce database in the e-commerce data collection module (100) is as follows: The e-commerce data collection module (100) completes the registration process on the open platform of the e-commerce platform, submits detailed corporate or personal information, and clarifies the purpose and scope of obtaining e-commerce data. After the open platform reviews and approves the registration, a unique access token is allocated to the e-commerce data collection module (100); The e-commerce data collection module (100) uses a network protocol to send the access token in the form of a data packet to the e-commerce database server. During the transmission process, the data packet passes through multiple network nodes and finally reaches the target server; After receiving the access token, the e-commerce database server retrieves pre-stored verification information to verify the access token. If the verification information is inconsistent with the access token, the server determines that the verification has failed and returns verification failure information to the e-commerce data collection module (100); if the two are consistent, the verification is successful, and the server returns verification success information to the e-commerce data collection module (100), and the e-commerce data collection module (100) establishes a connection with the e-commerce database.

3. The big data model system for e-commerce transaction anomaly detection based on time series analysis according to claim 2 is characterized by: The working principle of the e-commerce data collection module (100) for acquiring the e-commerce data of each user in the e-commerce database is as follows: The e-commerce data collection module (100) constructs a query statement according to the need to obtain e-commerce data, and sends the query statement to the database engine through the established connection. The database engine searches the database according to the query statement, retrieves the e-commerce data that meets the conditions, and outputs the e-commerce data to the e-commerce data collection module (100) through the established connection.

4. The big data model system for e-commerce transaction anomaly detection based on time series analysis according to claim 1 is characterized by: When the multi-dimensional user portrait generation module (200) uses the K-Means algorithm to construct a user portrait system, users with similar behavioral characteristics are grouped into one category. The specific working steps are as follows: Step 1: Preset different numbers of clusters , number of clusters The number of different types of user portraits; Step 2: Randomly select from e-commerce protective gear E-commerce data is used as the initial centroid, and each centroid is a vector with the values ​​of these three dimensions; Step 3: For each data in the e-commerce data, calculate the Euclidean distance between each data and The distance between the centroids is used to assign the e-commerce data to the cluster represented by the nearest centroid. Step 4: After allocating all e-commerce data to corresponding clusters, recalculate the centroid position of each cluster; For a cluster , its centroid The calculation method is to take the mean of the characteristic values ​​of each dimension of all e-commerce data in the cluster. There are E-commerce data , each e-commerce data , then the centroid No. Dimensional coordinates for: ; Step 5: Repeat steps 3 and 4, continuously assigning e-commerce data to clusters and updating the centroid position until the centroid position no longer changes significantly.

5. The big data model system for e-commerce transaction anomaly detection based on time series analysis according to claim 4 is characterized by: The linear attenuation method in the time series weighted portrait fusion module (300) perceives the total number of e-commerce data of the same user Order No. The portrait weight corresponding to each order The calculation formula is: ; Perceive the weight of the portrait and merge multiple user portraits with different weights through the weighted fusion method to generate the current user portrait; Weighted fusion method perceives the same user's shared Different user profiles , the corresponding weights calculated by the linear attenuation method are ,and , profile each user Represented as a feature vector ,in is the dimension of the feature; Current user portrait generated by weighted fusion method Also a The eigenvector of dimension characteristic The calculation formula is: .

6. The big data model system for e-commerce transaction anomaly detection based on time series analysis according to claim 5 is characterized by: The abnormal order judgment module (400) calculates the similarity between the current user portrait and the judgment portrait as follows: Perceive the current user profile , the image is determined to be , then the similarity between the two is: .

7. The big data model system for e-commerce transaction anomaly detection based on time series analysis according to claim 1 is characterized by: The threshold setting method in the abnormal order judgment module (400) sets the user portrait threshold interval according to the change rules between multiple user portraits of the same user in the e-commerce data collection module (100). The specific calculation formula is as follows: Perception User Profile Collection , which contains User portraits, each with feature dimension, for the Feature Dimensions , and its corresponding eigenvalue set is ; Calculate the mean: The mean of the feature dimensions The calculation formula is ; Calculate the standard deviation: The standard deviation of the feature dimension The calculation formula is ; The first The user portrait threshold range of each feature dimension is: ,in is an initial constant determined by cross-validation method.

8. The big data model system for e-commerce transaction anomaly detection based on time series analysis according to claim 7 is characterized by: The cross-validation method senses the current user portraits of different users and the current user portrait of the same user in the time-series weighted portrait fusion module (300), defines the current user portraits of different users as abnormal orders, and defines the user portraits of the same user as normal orders; By randomly setting the initial constant When judging that the current user profiles of different users are normal orders, the step length Adjust initial constants Similarly, if the initial constant When the user profile of the same user is judged to be an abnormal order, the synchronization step Adjust initial constants ; Perception unadjusted initial constant When the user portrait threshold interval is used to judge abnormal orders, the accuracy is , through the step length Adjustment constant Accuracy of determining abnormal orders ,like , then again through the step size Adjust initial constants Otherwise, change the adjustment direction until the accuracy cannot be improved.

9. The big data model system for e-commerce transaction anomaly detection based on time series analysis according to claim 1 is characterized by: The working principle of the abnormal order judgment module (400) for matching and calculating whether there is a product corresponding to the newly added order in the products corresponding to the historical orders is as follows: The e-commerce data of each user in the e-commerce data collection module (100) contains product numbers corresponding to different order products and browsing records corresponding to different products. The product numbers of the products corresponding to the newly added orders are compared with the product numbers of the products corresponding to the historical orders. When the product numbers are exactly the same, the two products are the same. Similarly, the most comprehensive browsing records in the browsing record collection are compared in the same way.

Images