A multi-platform e-commerce operation data analysis system and method based on federated learning

Abstract

This application provides a multi-platform e-commerce operation data analysis system and method based on federated learning. The central server receives encrypted quantile values ​​uploaded by clients from various platforms, calculates global feature quantile values ​​in the encrypted domain, and generates feature binning boundaries for distribution. Based on these boundaries, user behavior features of the platform are locally binned, and the sample count and positive sample count for each bin interval are statistically analyzed. Secret sharing is used to split all sample counts and positive sample counts into shares, and global sample count slices and global positive sample count slices are calculated. Based on these slices, evidence weight values ​​for each bin interval are calculated. Based on these evidence weight values, the information values ​​of each user behavior feature are federated and aggregated to output an operational feature profile of the shared user group across platforms. This scheme enables localized binning calculation of lightweight statistical summaries in the encrypted domain.

Description

Technical Field

[0001] This application relates to the field of federated learning technology, and more specifically, to a multi-platform e-commerce operation data analysis system and method based on federated learning. Background Technology

[0002] Multi-platform e-commerce operation refers to merchants or brands simultaneously joining multiple mainstream e-commerce platforms, maximizing user coverage and diversifying sales channels through unified product management, differentiated marketing strategies, and cross-platform data integration.

[0003] In the traditional federated learning framework, multi-platform e-commerce operational feature analysis generally faces bottlenecks in communication and computation efficiency. Existing solutions mostly employ iterative gradient transmission or batch ciphertext operations under homomorphic encryption. Each round of interaction requires the transmission of high-dimensional feature vectors or large encryption matrices, causing communication overhead to increase linearly with the number of platforms and feature dimensions. This easily leads to high latency and packet loss retransmission in wide area network environments. At the same time, each platform needs to frequently perform time-consuming homomorphic operations such as relinearization and modulus switching, resulting in a significant imbalance in local computing load. In particular, small and medium-sized platforms with limited computing resources cannot handle high-frequency ciphertext aggregation tasks. Most solutions place feature binning and statistical counting on a central server for plaintext aggregation. This not only forces the central node to bear the pressure of sorting and quantizing massive amounts of data, but also requires multiple adjustments to the binning strategy due to the heterogeneous feature distribution of each platform, exacerbating computational oscillations and communication round trips in cross-platform collaboration. Meanwhile, existing methods often require each platform to upload complete histograms or probability density estimates when evaluating feature effectiveness after binning. The size of intermediate results expands quadratically with the number of bins, causing a non-linear increase in communication load as feature granularity refines. This severely restricts real-time performance and horizontal scalability in large-scale platform clusters and fine-grained feature engineering scenarios. Therefore, how to achieve localized binning computation of lightweight statistical summaries in the encrypted domain, thereby improving the overall communication efficiency of federated feature engineering in large-scale multi-platform scenarios, has become a challenge for the industry. Summary of the Invention

[0004] This application provides a multi-platform e-commerce operation data analysis system and method based on federated learning, which can realize localized binning calculation of lightweight statistical summaries in encrypted fields, thereby improving the overall communication efficiency of federated feature engineering in large-scale multi-platform scenarios.

[0005] Firstly, this application provides a multi-platform e-commerce operation data analysis method based on federated learning, including:

[0006] After the central server sends the feature quantile query command to each platform client, the central server receives the encrypted quantile values ​​uploaded by each platform client, calculates the global feature quantile value in the ciphertext domain and generates the feature bin boundary and sends it out.

[0007] Each platform client performs partial binning of user behavior features on its own platform based on the binning boundaries, and then counts the sample count and positive sample count for each bin interval. Secret sharing is used to split all the sample count and positive sample count into shares and exchange them between platform clients.

[0008] Each platform client aggregates its share, then calculates the global sample count slice and the global positive sample count slice, and calculates the evidence weight value slice for each bin interval based on the global sample count slice and the global positive sample count slice and uploads it to the central server;

[0009] The central server performs federated aggregation of information values ​​of each user's behavioral characteristics based on the weight values ​​of each piece of evidence, and outputs an operational characteristic profile of the user group shared by all platforms.

[0010] In some embodiments, the central server receives encrypted quantile values ​​uploaded by clients from various platforms, calculates global feature quantile values ​​in the ciphertext domain, and generates feature bin boundaries for distribution. Specifically, this includes:

[0011] The central server generates homomorphic encryption key pairs and sends the homomorphic encryption public key to each platform client along with the feature quantile query command;

[0012] Each platform client sorts the specified user behavior feature values ​​locally, extracts the feature values ​​corresponding to the feature quantiles, encrypts the extracted feature values ​​using the homomorphic encryption public key, obtains the encrypted quantile values, and uploads them to the central server.

[0013] The central server performs a homomorphic addition and averaging operation on all encrypted quantile values ​​in the ciphertext field to obtain the global feature quantile value of the ciphertext state. The global feature quantile value of the ciphertext state is then homomorphically decrypted to obtain the global feature quantile value of the plaintext state.

[0014] The central server uses the global feature quantile value of the plaintext state as the interval breakpoint to generate feature bin boundaries and distribute them to each platform client.

[0015] In some embodiments, each platform client performs partial binning of its user behavior features based on the feature binning boundaries, and then counts the sample count and positive sample count for each binning interval. Specifically, this includes:

[0016] Each platform client receives the feature binning boundary, compares the feature value of the target feature in each user behavior record stored on the platform with the interval breakpoints in the feature binning boundary, and determines the binning interval to which each user behavior record belongs.

[0017] Each platform client iterates through all user behavior records on its platform, accumulates the total number of user behavior records falling into each bin interval, and obtains the sample count for each bin interval;

[0018] Within each binning interval, count the number of records in the user behavior records whose behavior labels are positive sample labels, and obtain the positive sample count.

[0019] In some embodiments, using secret sharing to split all sample counts and positive sample counts into shares and exchange them between platform clients specifically includes:

[0020] Each platform client obtains the total number of platform clients participating in the current calculation. For each binning interval, the sample count and positive sample count calculated by this platform are used as the secret value to be split.

[0021] Each platform client uses an additive secret sharing protocol to split each secret value to be split into secret shares equal to the total number of platform clients, with each secret share corresponding to one platform client;

[0022] Each platform client sends its secret share to the corresponding platform client, while keeping its own secret share locally, thus completing the share exchange.

[0023] In some embodiments, the aggregation of the shares held by each platform client, and the subsequent calculation of the global sample count shard and the global positive sample count shard, specifically includes:

[0024] Each platform client receives the secret share corresponding to each bin interval from other platform clients, and adds the received secret share to its own locally retained secret share according to the bin interval;

[0025] For each bin interval, the secret share of the sample count split and exchanged by each platform client is accumulated to obtain the global sample count fragment of the bin interval;

[0026] For each bin interval, the secret share of positive sample count split and exchanged by each platform client is accumulated to obtain the global positive sample count fragment of the bin interval;

[0027] This leads to the global sample count slice and the global positive sample count slice for each bin interval.

[0028] In some embodiments, calculating the evidence weight value fragment for each bin interval based on the global sample count fragment and the global positive sample count fragment, and uploading it to the central server specifically includes:

[0029] For each binning interval, each platform client obtains the global sample count slice and the global positive sample count slice of the binning interval, and determines the global negative sample count slice based on the difference between the global sample count slice and the global positive sample count slice.

[0030] Each platform client calculates the positive sample percentage slice and negative sample percentage slice of the binning interval based on the global positive sample count slice and the global negative sample count slice;

[0031] Each platform client compares the positive sample proportion slice with the negative sample proportion slice and takes the natural logarithm to obtain the evidence weight value slice for each bin interval, and then obtains the evidence weight value slice for each bin interval.

[0032] Each platform client uploads the evidence weight value of each bin segment to the central server.

[0033] In some embodiments, the central server performs federated aggregation of the information values ​​of each user behavior feature based on the evidence weight values ​​of each shard, and outputs an operational feature profile of the common user group across platforms, specifically including:

[0034] The central server obtains the evidence weight value fragments of each bin interval from each platform client, sums and reconstructs the evidence weight value fragments of each bin interval to obtain the evidence weight value of each bin interval.

[0035] The central server calculates the information value of each user behavior feature based on the evidence weight value of each bin interval, and filters out the target features that reach the preset information value threshold from the full set of user behavior features according to each information value. The target features are used as features for cross-platform analysis and are notified to each platform client.

[0036] Each platform client extracts the feature values ​​corresponding to the target features locally and uploads them to the central server;

[0037] The central server performs federated average aggregation of the feature values ​​of clients from various platforms, provides statistical descriptions of the aggregation results, and outputs an operational feature profile of the common user group across all platforms.

[0038] Secondly, this application provides a multi-platform e-commerce operation data analysis system based on federated learning, including:

[0039] The query module is used to receive encrypted quantile values ​​uploaded by each platform client after the central server sends the feature quantile query command to each platform client, calculate the global feature quantile value in the ciphertext domain and generate the feature bin boundary for distribution.

[0040] The processing module is used to partially bin the user behavior features of each platform client according to the feature binning boundary, and then count the sample count and positive sample count of each binning interval. The module uses secret sharing to split all the sample count and positive sample count into shares and exchange them between platform clients.

[0041] The processing module is also used to aggregate the share held by each platform client, and then calculate the global sample count slice and the global positive sample count slice. Based on the global sample count slice and the global positive sample count slice, the evidence weight value slice for each bin interval is calculated and uploaded to the central server.

[0042] The execution module is used by the central server to perform federated aggregation of information values ​​of each user's behavioral characteristics based on the weight values ​​of each piece of evidence, and output an operational characteristic profile of the user group shared by various platforms.

[0043] Thirdly, this application provides a computer device, which includes a memory and a processor. The memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the computer device executes the above-mentioned multi-platform e-commerce operation data analysis method based on federated learning.

[0044] Fourthly, this application provides a computer-readable storage medium storing instructions or code that, when executed on a computer, cause the computer to implement the aforementioned multi-platform e-commerce operation data analysis method based on federated learning.

[0045] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0046] In the multi-platform e-commerce operation data analysis system and method based on federated learning provided in this application, after the central server issues feature quantile query instructions to each platform client, the central server receives the encrypted quantile values ​​uploaded by each platform client, calculates the global feature quantile value in the encrypted domain, and generates feature bin boundaries for distribution; each platform client performs local binning of user behavior features on its own platform according to the feature bin boundaries, and then counts the sample count and positive sample count of each bin interval. Secret sharing is used to split all sample counts and positive sample counts into shares and exchange them among the platform clients; each platform client aggregates its shares and then calculates the global sample count slice and the global positive sample count slice. Based on the global sample count slice and the global positive sample count slice, the evidence weight value slice of each bin interval is calculated and uploaded to the central server; the central server performs federated aggregation of the information values ​​of each user behavior feature based on each evidence weight value slice, and outputs an operational feature profile of the common user group across all platforms.

[0047] Therefore, in this application, the central server performs federated aggregation of the information values ​​of each user behavior feature based on the weight values ​​of each piece of evidence, and outputs the operational feature profile of the user group shared by various platforms. First, by determining the feature binning boundary, a unified quantization scale in the ciphertext domain can be obtained, thereby avoiding the iterative alignment overhead caused by the heterogeneity of features across multiple platforms. The central server calculates the global feature quantile value in the ciphertext domain and generates a unified binning boundary applicable to all platforms at once, so that each platform can complete the local feature discretization at the same quantization granularity without exchanging the original features or local statistical histograms. This boundary possesses the mathematical closure derived from the encrypted domain, ensuring that binning breakpoints do not leak sample density information within specific numerical intervals. It also eliminates the need for normalization preprocessing due to inconsistencies in feature scales between platforms. The pre-coordination stage of feature engineering is compressed into a single downlink command, significantly reducing the frequency of cross-platform control flow interactions. This prevents communication overhead from linearly increasing with the number of platforms or feature dimensions, establishing a unified numerical mapping foundation for efficient collaborative analysis in large-scale platform clusters. Then, determining the evidence weight values ​​for partitioning yields a compact, fractionalized representation of global statistical information, thereby reducing the communication load expansion caused by the original histogram transmission. Secret sharing is employed to synchronize the sample counts of each platform with the positive... Sample counts are split into shares and exchanged between platforms. Each platform only needs to aggregate its own share to obtain global sample count fragments and global positive sample count fragments, and then calculate evidence weight value fragments. This reduces the single transmission load without changing the number of communication rounds. At the same time, the evidence weight value fragments themselves are already intermediate results of feature validity evaluation and can be directly used for information value aggregation, avoiding the need for additional transmission of auxiliary statistics used to calculate evidence weights. This optimizes communication efficiency in the global statistical feature extraction stage. In summary, based on the above scheme, localized binning calculation of lightweight statistical summaries in the encrypted field can be achieved, thereby improving the overall communication efficiency of federated feature engineering in large-scale multi-platform scenarios. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is an exemplary flowchart of a multi-platform e-commerce operation data analysis method based on federated learning, as shown in some embodiments of this application;

[0050] Figure 2 This is a flowchart illustrating the process of determining evidence weight values ​​and segmenting them according to some embodiments of this application;

[0051] Figure 3 This is a schematic diagram of the structure of a multi-platform e-commerce operation data analysis system based on federated learning, as shown in some embodiments of this application;

[0052] Figure 4 This is a schematic diagram of the structure of a computer device that implements a multi-platform e-commerce operation data analysis method based on federated learning, according to some embodiments of this application. Detailed Implementation

[0053] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] refer to Figure 1 The figure is an exemplary flowchart of a multi-platform e-commerce operation data analysis method based on federated learning, according to some embodiments of this application. The multi-platform e-commerce operation data analysis method based on federated learning mainly includes the following steps:

[0055] In step 101, after the central server sends the feature quantile query instruction to each platform client, the central server receives the encrypted quantile values ​​uploaded by each platform client, calculates the global feature quantile value in the ciphertext domain, and generates and sends the feature bin boundaries.

[0056] It should be noted that in this application, the feature percentile query instruction is an instruction that instructs each platform client to sort the specified user behavior feature values ​​locally and extract the feature values ​​corresponding to the specified percentiles. In specific implementation, the central server presets a percentile list, which contains multiple percentile ratio values, each percentile ratio value being a value between 0 and 1; the central server generates a homomorphic encryption key pair, which contains a homomorphic encryption public key and a homomorphic encryption private key; the central server encapsulates the percentile list and the homomorphic encryption public key into a feature percentile query instruction, and the central server issues the feature percentile query instruction to each platform client.

[0057] In some embodiments, the central server receives encrypted quantile values ​​uploaded by clients from various platforms, calculates global feature quantile values ​​in the ciphertext domain, and generates feature bin boundaries for distribution. This can be achieved through the following steps:

[0058] The central server generates homomorphic encryption key pairs and sends the homomorphic encryption public key to each platform client along with the feature quantile query command;

[0059] Each platform client sorts the specified user behavior feature values ​​locally, extracts the feature values ​​corresponding to the feature quantiles, encrypts the extracted feature values ​​using the homomorphic encryption public key, obtains the encrypted quantile values, and uploads them to the central server.

[0060] The central server performs a homomorphic addition and averaging operation on all encrypted quantile values ​​in the ciphertext field to obtain the global feature quantile value of the ciphertext state. The global feature quantile value of the ciphertext state is then homomorphically decrypted to obtain the global feature quantile value of the plaintext state.

[0061] The central server uses the global feature quantile value of the plaintext state as the interval breakpoint to generate feature bin boundaries and distribute them to each platform client.

[0062] It should be noted that in this application, the homomorphic encryption public key is the encryption key used by each platform client to homomorphically encrypt the extracted feature values ​​to generate encrypted quantile values; the encrypted quantile values ​​are homomorphic ciphertext data used by the central server to calculate the global feature quantile values ​​in the ciphertext domain; the global feature quantile values ​​are plaintext values ​​reflecting the overall distribution level of cross-platform user behavior feature values ​​at each quantile point; and the feature bin boundaries are the set of interval breakpoints used to divide continuous user behavior feature values ​​into multiple discrete intervals.

[0063] In specific implementation, firstly, the central server presets a quantile list, which contains multiple quantile ratio values, each between 0 and 1. The central server generates a homomorphic encryption key pair, containing a homomorphic encryption public key and a homomorphic encryption private key. The central server encapsulates the quantile list and the homomorphic encryption public key into a feature quantile query command, and sends the command and the public key to each platform client. Secondly, each platform client receives the feature quantile query command and the public key, and parses the quantile list from the command. Each platform client reads the user behavior feature values ​​stored on its local machine, sorts the user behavior feature values ​​to obtain a sorted feature value sequence. Based on each quantile ratio value in the quantile list, each platform client locates the position corresponding to that quantile ratio value in the sorted feature value sequence and extracts the feature value at that position as the feature value corresponding to that quantile. Each platform client uses homomorphic encryption... The public key is used to homomorphically encrypt the feature value corresponding to each quantile point, and the resulting ciphertext is used as the encrypted quantile value. Each platform client uploads the encrypted quantile value to the central server. The central server then receives the uploaded quantile values ​​from each platform client, performs homomorphic addition on all encrypted quantile values ​​corresponding to the same quantile point in the ciphertext field, and divides the addition result by the number of platform clients to obtain the global feature quantile value of the ciphertext state. The central server then uses the homomorphic encryption private key to homomorphically decrypt the global feature quantile value of the ciphertext state, and uses the decrypted plaintext as the global feature quantile value of the plaintext state. Finally, the central server sorts the global feature quantile values ​​of the plaintext state corresponding to each quantile point in ascending order of value, uses the sorted global feature quantile values ​​of each plaintext state as interval breakpoints, and forms a binning interval by adjacent interval breakpoints. The set of interval breakpoints corresponding to all binning intervals is used as the feature binning boundary. The central server then distributes the feature binning boundary to each platform client.

[0064] In step 102, each platform client performs partial binning of the user behavior features of its platform according to the feature binning boundary, and then counts the sample count and positive sample count of each binning interval. The secret sharing method is used to split all the sample count and positive sample count into shares and exchange them between platform clients.

[0065] In some embodiments, the following steps can be taken to partially bin the user behavior features of each platform client based on the feature binning boundaries, and then count the sample count and positive sample count for each binning interval:

[0066] Each platform client receives the feature binning boundary, compares the feature value of the target feature in each user behavior record stored on the platform with the interval breakpoints in the feature binning boundary, and determines the binning interval to which each user behavior record belongs.

[0067] Each platform client iterates through all user behavior records on its platform, accumulates the total number of user behavior records falling into each bin interval, and obtains the sample count for each bin interval;

[0068] Within each binning interval, count the number of records in the user behavior records whose behavior labels are positive sample labels, and obtain the positive sample count.

[0069] It should be noted that, in this application, the binning interval is a numerical range used to classify continuous user behavior feature values ​​into discrete intervals; the sample count represents the statistical value of the total number of user behavior records falling into a single binning interval; and the positive sample count represents the statistical value of the number of user behavior records falling into a single binning interval and whose behavior label is a positive sample label.

[0070] In practice, firstly, each platform client receives the feature binning boundary from the central server. The feature binning boundary contains multiple interval breakpoints arranged in ascending order of numerical value. Each platform client reads user behavior records stored on its own platform. Each user behavior record contains the feature value of the target feature and a behavior label. For each user behavior record, each platform client compares the feature value of the target feature with the interval breakpoints in the feature binning boundary sequentially, determining the numerical range formed by two adjacent interval breakpoints. This numerical range is then used as the binning interval to which the user behavior record belongs, and each user behavior record is assigned to the corresponding binning interval. Next, each platform client initializes the sample count for each binning interval to 0. Each platform client then iterates through all the data on its platform. For each user behavior record, the sample count corresponding to that bin is incremented by 1. After all user behavior records have been traversed, each platform client obtains the sample count for each bin. Finally, each platform client initializes the positive sample count for each bin to 0. Each platform client traverses all bins on its platform, and for each user behavior record within each bin, determines whether the behavior label of that user behavior record is a positive sample label. If the behavior label is a positive sample label, the positive sample count corresponding to that bin is incremented by 1. After all user behavior records within that bin have been determined, each platform client obtains the positive sample count for that bin. Finally, each platform client obtains the positive sample count for each bin.

[0071] In some embodiments, using secret sharing to split all sample counts and positive sample counts into shares and exchange them between platform clients can be achieved through the following steps:

[0072] Each platform client obtains the total number of platform clients participating in the current calculation. For each binning interval, the sample count and positive sample count calculated by this platform are used as the secret value to be split.

[0073] Each platform client uses an additive secret sharing protocol to split each secret value to be split into secret shares equal to the total number of platform clients, with each secret share corresponding to one platform client;

[0074] Each platform client sends its secret share to the corresponding platform client, while keeping its own secret share locally, thus completing the share exchange.

[0075] It should be noted that in this application, the secret value to be split is the original statistical value that is input to the additive secret sharing protocol and needs to be split into multiple secret shares; the secret share is a value fragment held by multiple participants without disclosing the original statistical value; the share exchange is the process by which each platform client holds secret shares from all platform clients corresponding to the same bin interval, so as to subsequently aggregate and calculate the global statistical fragments.

[0076] In practice, firstly, each platform client obtains the total number of platform clients participating in the current calculation. This total number is an integer pre-configured in the central server and distributed to each platform client during the initialization phase. For each bin interval, each platform client uses its calculated sample count as the first secret value to be split and its calculated positive sample count as the second secret value to be split. Then, using an additive secret sharing protocol, each platform client generates a first set of secret shares equal to the total number of platform clients for the first secret value to be split, with the total number of platform clients as the number of shares. Each first set of secret shares corresponds to one platform client. For the second secret value to be split... A second set of secret shares is generated, equal in number to the total number of platform clients. Each second set of secret shares corresponds to one platform client. The first and second sets of secret shares are collectively referred to as secret shares. Finally, for each bin segment, each platform client sends the secret shares corresponding to other platform clients in the first set of secret shares to the corresponding platform clients, and retains the secret shares corresponding to itself in the first set of secret shares locally. Similarly, each platform client sends the secret shares corresponding to other platform clients in the second set of secret shares to the corresponding platform clients, and retains the secret shares corresponding to itself in the second set of secret shares locally. After each platform client completes the sending and retention of secret shares for all bin segments, the share exchange is completed.

[0077] In step 103, each platform client aggregates its share and then calculates the global sample count slice and the global positive sample count slice. Based on the global sample count slice and the global positive sample count slice, the evidence weight value slice for each bin interval is calculated and uploaded to the central server.

[0078] In some embodiments, the aggregation of the shares held by each platform client to calculate the global sample count slice and the global positive sample count slice can be achieved through the following steps:

[0079] Each platform client receives the secret share corresponding to each bin interval from other platform clients, and adds the received secret share to its own locally retained secret share according to the bin interval;

[0080] For each bin interval, the secret share of the sample count split and exchanged by each platform client is accumulated to obtain the global sample count fragment of the bin interval;

[0081] For each bin interval, the secret share of positive sample count split and exchanged by each platform client is accumulated to obtain the global positive sample count fragment of the bin interval;

[0082] This leads to the global sample count slice and the global positive sample count slice for each bin interval.

[0083] It should be noted that in this application, the global sample count slice is an intermediate statistic slice that participates in the calculation of the evidence weight value without restoring the original sample counts of each platform; the global positive sample count slice is an intermediate statistic slice that participates in the calculation of the evidence weight value without restoring the original positive sample counts of each platform.

[0084] In practice, firstly, each platform client receives secret shares corresponding to each bin interval from other platform clients. These secret shares include those generated by other platform clients for sample count splitting and those generated by other platform clients for positive sample count splitting. Each platform client then aggregates the received secret shares from other platform clients along with its own locally retained secret shares, categorized by bin interval. Secondly, for each bin interval, each platform client sums the aggregated secret shares generated by all platform clients within that bin interval for sample count splitting and exchanging. The sum is used as the global sample count fragment for that bin interval. Then, for each bin interval, each platform client sums the aggregated secret shares generated by all platform clients within that bin interval for positive sample count splitting and exchanging. The sum is used as the global positive sample count fragment for that bin interval. Finally, the global sample count fragment and the global positive sample count fragment for each bin interval are obtained through the above methods.

[0085] In some embodiments, the evidence weight value fragment for each bin interval is calculated based on the global sample count fragment and the global positive sample count fragment, and then uploaded to the central server, for reference. Figure 2The diagram is a flowchart illustrating the process of determining evidence weight value segments in some embodiments of this application. In this embodiment, determining evidence weight value segments can be achieved using the following steps:

[0086] In step 1031, for each bin interval, each platform client obtains the global sample count slice and the global positive sample count slice of the bin interval, and determines the global negative sample count slice based on the difference between the global sample count slice and the global positive sample count slice.

[0087] In step 1032, each platform client calculates the positive sample percentage slice and negative sample percentage slice of the binning interval based on the global positive sample count slice and the global negative sample count slice;

[0088] In step 1033, each platform client compares the positive sample proportion slice and the negative sample proportion slice and takes the natural logarithm to obtain the evidence weight value slice of the bin interval, and then obtains the evidence weight value slice of each bin interval.

[0089] In step 1034, each platform client uploads the evidence weight value of each bin section to the central server in segments.

[0090] It should be noted that in this application, the global negative sample count slice is an intermediate statistical slice used in the calculation of the proportion of positive and negative samples without restoring the original negative sample counts of each platform; the positive sample proportion slice is a numerical slice reflecting the proportion of positive samples in all samples within a single binning interval; the negative sample proportion slice is a numerical slice reflecting the proportion of negative samples in all samples within a single binning interval; and the evidence weight value slice is a numerical slice reflecting the ability of feature values ​​to distinguish between positive and negative samples within a single binning interval without restoring the original data of each platform.

[0091] In specific implementation, firstly, each platform client obtains the global sample count slice and the global positive sample count slice for each bin interval; each platform client subtracts the global positive sample count slice from the global sample count slice, and uses the result as the global negative sample count slice for that bin interval; secondly, each platform client divides the global positive sample count slice by the global sample count slice for each bin interval, and uses the result as the positive sample percentage slice for that bin interval; and divides the global negative sample count slice by the global sample count slice, and uses the result as the positive sample percentage slice for that bin interval. The result of the division is used as the negative sample proportion segment for that binning interval. Then, for each binning interval, each platform client divides the positive sample proportion segment by the negative sample proportion segment, takes the natural logarithm of the division result, and uses the result after taking the natural logarithm as the evidence weight value segment for that binning interval. Each platform client performs the above calculation on all binning intervals to obtain the evidence weight value segment for each binning interval. Finally, each platform client packages the calculated evidence weight value segments for each binning interval and uploads the packaged evidence weight value segments to the central server.

[0092] In step 104, the central server performs federated aggregation of the information values ​​of each user behavior feature based on the weight values ​​of each piece of evidence, and outputs an operational feature profile of the user group shared by all platforms.

[0093] In some embodiments, the central server performs federated aggregation of the information values ​​of each user behavior feature based on each evidence weight value shard, and outputs an operational feature profile of the common user group across platforms. This can be achieved through the following steps:

[0094] The central server obtains the evidence weight value fragments of each bin interval from each platform client, sums and reconstructs the evidence weight value fragments of each bin interval to obtain the evidence weight value of each bin interval.

[0095] The central server calculates the information value of each user behavior feature based on the evidence weight value of each bin interval, and filters out the target features that reach the preset information value threshold from the full set of user behavior features according to each information value. The target features are used as features for cross-platform analysis and are notified to each platform client.

[0096] Each platform client extracts the feature values ​​corresponding to the target features locally and uploads them to the central server;

[0097] The central server performs federated average aggregation of the feature values ​​of clients from various platforms, provides statistical descriptions of the aggregation results, and outputs an operational feature profile of the common user group across all platforms.

[0098] It should be noted that in this application, the evidence weight value is a complete plaintext numerical value that reflects the ability of the feature value within the binning interval to distinguish between positive and negative samples; the information value is an evaluation index that measures the overall ability of user behavior features to distinguish between positive and negative samples; the target feature is the user behavior feature of cross-platform operation analysis; and the operation feature profile is a set of data descriptions that reflect the distribution and preference features of the behavior features of the common user group.

[0099] In specific implementation, firstly, the central server receives evidence weight value fragments for each bin space uploaded by clients from various platforms. For each bin space, it sums all received evidence weight value fragments and uses the sum as the evidence weight value for that bin space. The central server performs summation and reconstruction on all bin spaces to obtain the evidence weight value for each bin space. Secondly, the central server presets an information value threshold, which is a value greater than 0. For each user behavior feature, the central server substitutes the evidence weight values ​​of all bin spaces corresponding to that user behavior feature into the information value calculation formula to obtain the information value for that user behavior feature. The central server compares the information value of each user behavior feature with the preset information value threshold one by one, filtering out target features whose information values ​​are greater than or equal to the preset information value threshold. The central server then selects the target features. The target feature is used for cross-platform analysis. The identifier of the target feature is notified to each platform client. Then, each platform client receives the target feature identifier from the central server, reads the user behavior records stored on its local machine, and extracts the corresponding feature value from the user behavior records. Each platform client uploads the extracted feature value to the central server. Finally, the central server receives the feature values ​​uploaded by each platform client, performs federated average aggregation on all feature values ​​of the same target feature, and the federated average aggregation is the sum of the feature values ​​uploaded by each platform client divided by the number of platform clients to obtain the aggregation result. The central server provides a statistical description of the aggregation result, including calculating the mean, distribution range, and central tendency indicators of the aggregation result. The results of the statistical description are used as an operational characteristic profile of the common user group across all platforms.

[0100] Furthermore, in another aspect of this application, in some embodiments, this application provides a multi-platform e-commerce operation data analysis system based on federated learning, referencing... Figure 3 The figure is a schematic diagram of the structure of a multi-platform e-commerce operation data analysis system based on federated learning, according to some embodiments of this application. This system includes a query module 201, a processing module 202, and an execution module 203, which are described below:

[0101] The query module 201 in this application is mainly used to receive the encrypted quantile values ​​uploaded by each platform client after the central server sends the feature quantile query instruction to each platform client, calculate the global feature quantile value in the ciphertext domain and generate the feature bin boundary for distribution.

[0102] Processing module 202, in this application, is used by each platform client to partially bin the user behavior features of the platform according to the feature binning boundary, and then count the sample count and positive sample count of each binning interval. It uses secret sharing to split all the sample count and positive sample count into shares and exchange them between platform clients.

[0103] It should be noted that the processing module 202 is also used to aggregate the share held by each platform client, and then calculate the global sample count slice and the global positive sample count slice, and calculate the evidence weight value slice for each bin interval based on the global sample count slice and the global positive sample count slice and upload it to the central server.

[0104] The execution module 203 in this application is mainly used by the central server to perform federated aggregation of the information values ​​of each user's behavioral characteristics based on the weight values ​​of each piece of evidence, and output the operational characteristic profile of the user group shared by various platforms.

[0105] The foregoing detailed examples of a multi-platform e-commerce operation data analysis system and method based on federated learning provided in this application. It is understood that the corresponding apparatus, in order to achieve the above functions, includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0106] In some embodiments, this application also provides a computer device, the computer device including a memory and a processor, the memory for storing a computer program, and the processor for calling and running the computer program from the memory, so that the computer device performs the above-described multi-platform e-commerce operation data analysis method based on federated learning.

[0107] In some embodiments, reference Figure 4The dashed lines in the figure indicate that the unit or module is optional. This figure is a schematic diagram of the structure of a computer device implementing a multi-platform e-commerce operation data analysis method based on federated learning, according to an embodiment of this application. The multi-platform e-commerce operation data analysis method based on federated learning described in the above embodiments can... Figure 4 The computer device shown is used to implement this, and the computer device includes at least one processor 301, a memory 302 and at least one communication unit 305. The computer device may be a terminal device, a server or a chip.

[0108] Processor 301 can be a general-purpose processor or a special-purpose processor. For example, processor 301 can be a central processing unit (CPU), which can be used to control computer devices, execute software programs, and process data from software programs. The computer device may also include a communication unit 305 for inputting (receiving) and outputting (transmitting) signals.

[0109] For example, the computer device may be a chip, and the communication unit 305 may be the input and / or output circuit of the chip, or the communication unit 305 may be the communication interface of the chip, which may be a component of a terminal device, network device or other device.

[0110] For example, the computer device may be a terminal device or a server, and the communication unit 305 may be a transceiver of the terminal device or the server, or the communication unit 305 may be a transceiver circuit of the terminal device or the server.

[0111] The computer device may include one or more memories 302 storing a program 304. The program 304 can be executed by a processor 301 to generate instructions 303, causing the processor 301 to execute the method described in the above method embodiments according to the instructions 303. Optionally, the memory 302 may also store data (such as a target audit model). Optionally, the processor 301 may also read data stored in the memory 302, which may be stored at the same storage address as the program 304, or it may be stored at a different storage address than the program 304.

[0112] The processor 301 and memory 302 can be configured separately or integrated together, for example, integrated on the system on chip (SOC) of the terminal device.

[0113] It should be understood that each step of the above method embodiment can be completed by hardware logic circuits or software instructions in the processor 301. The processor 301 can be a CPU, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, such as discrete gate, transistor logic devices, or discrete hardware components.

[0114] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0115] For example, in some embodiments, this application also provides a computer-readable storage medium storing instructions or code that, when executed on a computer, cause the computer to implement the above-described multi-platform e-commerce operation data analysis method based on federated learning.

[0116] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0117] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A multi-platform e-commerce operation data analysis method based on federated learning, characterized in that, Includes the following steps: After the central server sends the feature quantile query command to each platform client, the central server receives the encrypted quantile values ​​uploaded by each platform client, calculates the global feature quantile value in the ciphertext domain and generates the feature bin boundary and sends it out. Each platform client performs partial binning of user behavior features on its own platform based on the binning boundaries, and then counts the sample count and positive sample count for each bin interval. All sample counts and positive sample counts are split into shares using secret sharing and exchanged between platform clients. Each platform client aggregates its share, then calculates the global sample count slice and the global positive sample count slice, and calculates the evidence weight value slice for each bin interval based on the global sample count slice and the global positive sample count slice and uploads it to the central server; The central server performs federated aggregation of information values ​​of each user's behavioral characteristics based on the weight values ​​of each piece of evidence, and outputs an operational characteristic profile of the user group shared by all platforms.

2. The method as described in claim 1, characterized in that, The central server receives encrypted quantile values ​​uploaded by clients from various platforms, calculates global feature quantile values ​​in the ciphertext domain, and generates feature bin boundaries for distribution. Specifically, this includes: The central server generates homomorphic encryption key pairs and sends the homomorphic encryption public key to each platform client along with the feature quantile query command; Each platform client sorts the specified user behavior feature values ​​locally, extracts the feature values ​​corresponding to the feature quantiles, encrypts the extracted feature values ​​using the homomorphic encryption public key, obtains the encrypted quantile values, and uploads them to the central server. The central server performs a homomorphic addition and averaging operation on all encrypted quantile values ​​in the ciphertext field to obtain the global feature quantile value of the ciphertext state. The global feature quantile value of the ciphertext state is then homomorphically decrypted to obtain the global feature quantile value of the plaintext state. The central server uses the global feature quantile value of the plaintext state as the interval breakpoint to generate feature bin boundaries and distribute them to each platform client.

3. The method as described in claim 1, characterized in that, Each platform client performs local binning of user behavior features based on the aforementioned binning boundaries, and then counts the sample count and positive sample count for each binning interval. Specifically, this includes: Each platform client receives the feature binning boundary, compares the feature value of the target feature in each user behavior record stored on the platform with the interval breakpoints in the feature binning boundary, and determines the binning interval to which each user behavior record belongs. Each platform client iterates through all user behavior records on its platform, accumulates the total number of user behavior records falling into each bin interval, and obtains the sample count for each bin interval; Within each binning interval, count the number of records in the user behavior records whose behavior labels are positive sample labels, and obtain the positive sample count.

4. The method as described in claim 1, characterized in that, The secret sharing mechanism splits all sample counts and positive sample counts into shares and exchanges them between platform clients. Specifically, this includes: Each platform client obtains the total number of platform clients participating in the current calculation. For each binning interval, the sample count and positive sample count calculated by this platform are used as the secret value to be split. Each platform client uses an additive secret sharing protocol to split each secret value to be split into secret shares equal to the total number of platform clients, with each secret share corresponding to one platform client; Each platform client sends its secret share to the corresponding platform client, while keeping its own secret share locally, thus completing the share exchange.

5. The method as described in claim 1, characterized in that, The calculation of global sample count shards and global positive sample count shards involves aggregating the share held by clients on various platforms. Each platform client receives the secret share corresponding to each bin interval from other platform clients, and adds the received secret share to its own locally retained secret share according to the bin interval; For each bin interval, the secret share of the sample count split and exchanged by each platform client is accumulated to obtain the global sample count fragment of the bin interval; For each bin interval, the secret share of positive sample count split and exchanged by each platform client is accumulated to obtain the global positive sample count fragment of the bin interval; This leads to the global sample count slice and the global positive sample count slice for each bin interval.

6. The method as described in claim 1, characterized in that, The calculation of evidence weight value segments for each bin interval based on the global sample count segments and the global positive sample count segments, and the uploading of these segments to the central server, specifically includes: For each binning interval, each platform client obtains the global sample count slice and the global positive sample count slice of the binning interval, and determines the global negative sample count slice based on the difference between the global sample count slice and the global positive sample count slice. Each platform client calculates the positive sample percentage slice and negative sample percentage slice of the binning interval based on the global positive sample count slice and the global negative sample count slice; Each platform client compares the positive sample proportion slice with the negative sample proportion slice and takes the natural logarithm to obtain the evidence weight value slice for each bin interval, and then obtains the evidence weight value slice for each bin interval. Each platform client uploads the evidence weight value of each bin segment to the central server.

7. The method as described in claim 1, characterized in that, The central server performs federated aggregation of information values ​​for each user's behavioral characteristics based on the weight values ​​of each piece of evidence, and outputs an operational characteristic profile of the user group shared by all platforms, specifically including: The central server obtains the evidence weight value fragments of each bin interval from each platform client, sums and reconstructs the evidence weight value fragments of each bin interval to obtain the evidence weight value of each bin interval. The central server calculates the information value of each user behavior feature based on the evidence weight value of each bin interval, and filters out the target features that reach the preset information value threshold from the full set of user behavior features according to each information value. The target features are used as features for cross-platform analysis and are notified to each platform client. Each platform client extracts the feature values ​​corresponding to the target features locally and uploads them to the central server; The central server performs federated average aggregation of the feature values ​​of clients from various platforms, provides statistical descriptions of the aggregation results, and outputs an operational feature profile of the common user group across all platforms.

8. A multi-platform e-commerce operation data analysis system based on federated learning, characterized in that, include: The query module is used to receive encrypted quantile values ​​uploaded by each platform client after the central server sends the feature quantile query command to each platform client, calculate the global feature quantile value in the ciphertext domain and generate the feature bin boundary for distribution. The processing module is used to partially bin the user behavior features of each platform client according to the feature binning boundary, and then count the sample count and positive sample count of each binning interval. The module uses secret sharing to split all the sample count and positive sample count into shares and exchange them between platform clients. The processing module is also used to aggregate the share held by each platform client, and then calculate the global sample count slice and the global positive sample count slice. Based on the global sample count slice and the global positive sample count slice, the evidence weight value slice for each bin interval is calculated and uploaded to the central server. The execution module is used by the central server to perform federated aggregation of information values ​​of each user's behavioral characteristics based on the weight values ​​of each piece of evidence, and output an operational characteristic profile of the user group shared by various platforms.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to call and run the computer programs from the memory, so that the computer device performs the multi-platform e-commerce operation data analysis method based on federated learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions or code that, when executed on a computer, cause the computer to implement the multi-platform e-commerce operation data analysis method based on federated learning as described in any one of claims 1 to 7.

Images