A customer value evaluation method for cross-border e-commerce platforms based on artificial intelligence

Abstract

The present invention discloses a customer value evaluation method for a cross-border e-commerce platform based on artificial intelligence, which relates to the technical field of customer value evaluation. The present invention adopts digital twin technology to fuse customer behavior time series data and multimodal features into virtual images through nonlinear mapping, so that discrete data is converted into continuous image display, thereby intuitively presenting the evolution process of customer behavior; the time series data mean is aggregated and mapped together with the multimodal features, and the influence of each data is adjusted using a weight matrix and bias to achieve effective docking between data, and intuitively display the correspondence between customer status and behavior; the customer behavior data is encoded and decoded through a deep autoencoder, and the reconstruction error is calculated, which is converted into anomaly scores through a scoring mapping function, and then combined with the isolation forest algorithm for anomaly detection, thereby generating early warning signals, and timely identifying the risks of surge in returns and abnormal transactions, thereby making up for the shortcomings of traditional methods in real-time risk monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of customer value evaluation, and in particular to a customer value evaluation method for a cross-border e-commerce platform based on artificial intelligence. Background Art

[0002] At present, most cross-border e-commerce platforms evaluate customer value by building customer value models based on historical orders, reviews, and logistics data, and using fixed rules to divide customers into different levels / tiers.

[0003] However, in actual operations, the information collected by the platform mostly comes from multiple countries and channels, the data is not updated in a timely manner, and the format is not uniform, which cannot effectively capture the instantaneous changes in customer behavior. In international trade, different countries have large differences in language, culture and payment habits, which can easily lead to inaccurate customer portraits. In addition, when faced with a surge in order returns or abnormal transactions caused by sudden regional events, holidays, etc., the platform can only discover problems through statistics after the fact, and lacks real-time monitoring and early warning. In response to this, some platforms have set dynamic thresholds and adjustment rules, but it is still difficult to cope with the ever-changing risk patterns in the market environment. There are lags and noise in data collection, which makes the risk identification model less sensitive to abnormal behavior, affecting the accuracy of customer value evaluation. At the same time, static models cannot dynamically track the entire life cycle of customers, resulting in it being difficult for the platform to intervene in advance to adjust operating strategies when customer behavior fluctuates.

[0004] This situation not only weakens the platform's ability to prevent and control customer churn risks and fraud, but also makes it easy to miss sudden order returns or abnormal payment behaviors. Such sudden order returns or abnormal payments have a certain probability of fraud risks or system problems, and the value evaluation process only regards them as low-value or invalid data. However, in fact, this part of the data has been evaluated and analyzed, but has not been effectively identified. It can be seen that traditional evaluation methods have obvious shortcomings in risk warning. Therefore, there is an urgent need for a customer value evaluation method for cross-border e-commerce platforms based on artificial intelligence to solve such problems. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides an artificial intelligence-based customer value evaluation method for a cross-border e-commerce platform to solve the problem that traditional value evaluation methods have limited analysis of cross-border return rates and abnormal transactions, cannot identify return anomalies in a timely manner, and cannot discover the risk in a timely manner.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] The embodiment of the present invention provides a method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence, which includes:

[0009] Step S1, collect order, evaluation, return and channel information, and standardize the collected data to obtain pre-processed data;

[0010] Step S2: Process the preprocessed data using natural language processing, image recognition, and speech analysis technologies to extract multimodal features, and use a deep learning model to generate dynamic customer tags;

[0011] Step S3, performing time series analysis on the dynamic customer tags to obtain customer behavior time series data;

[0012] Step S4: Using digital twin technology to virtually map the customer behavior time series data and multimodal features to form a customer digital twin image;

[0013] Step S5: Input the customer behavior time series data into a deep autoencoder and anomaly detection algorithm to detect abnormal returns and fraud risk behaviors and generate early warning signals;

[0014] Step S6: updating the customer value evaluation index according to the customer behavior time series data and the early warning signal, and stratifying the customers.

[0015] As a preferred solution of the cross-border e-commerce platform customer value evaluation method based on artificial intelligence described in the present invention, wherein: in step S2, the step of using natural language processing, image recognition and speech analysis technology to process preprocessed data to extract multimodal features, and using a deep learning model to generate dynamic customer tags is as follows:

[0016] Based on the preprocessed data, embedding methods are used for text, image, and speech respectively to calculate the multimodal feature vector. The definition formula is:

[0017] F=σ(W1e1+W2e2+W3e3+b),

[0018] Where F represents the multimodal feature vector, e1 represents the text embedding vector, e2 represents the image embedding vector, e3 represents the speech embedding vector, W1 represents the weight matrix corresponding to the text embedding, W2 represents the weight matrix corresponding to the image embedding, W3 represents the weight matrix corresponding to the speech embedding, b represents the bias vector, and σ represents the activation function;

[0019] The multimodal features are mapped to the label space using the fully connected layer, which is defined as:

[0020] z=UF+c,L=softmax(z),

[0021] Where z represents the fully connected layer output vector, U represents the mapping weight matrix, F represents the aforementioned multimodal feature vector, c represents the mapping bias vector, L represents the dynamic customer label vector, and softmax represents the normalization function.

[0022] As a preferred solution of the cross-border e-commerce platform customer value evaluation method based on artificial intelligence described in the present invention, the multimodal feature extraction step also includes processing social interaction data to generate dynamic customer tags containing regional, cultural and payment habit information.

[0023] As a preferred solution of the cross-border e-commerce platform customer value evaluation method based on artificial intelligence described in the present invention, the digital twin image is used to reflect the correspondence between customer behavior time series data and multimodal features.

[0024] As a preferred solution of the cross-border e-commerce platform customer value evaluation method based on artificial intelligence described in the present invention, wherein: in step S4, the step of using digital twin technology to virtually map the customer behavior time series data and multimodal features to form a customer digital twin image is as follows:

[0025] For customer behavior time series data {t1, t2,…, t n} for mean aggregation, the calculation formula is:

[0026]

[0027] Where T′ represents the aggregated time series data vector, t i represents the eigenvector of the i-th moment, and n represents the number of moments;

[0028] The aggregated time series data vector T ′ It is fused with the multimodal feature vector F and the customer digital twin image is constructed through a nonlinear mapping function. The formula is:

[0029] D=φ(W t T′+W f F+d),

[0030] Where D represents the customer’s digital twin image, W t Represents the time series data conversion weight matrix, T ′ Represents the aggregated time series data vector, W f represents the multimodal feature conversion weight matrix, F represents the multimodal feature vector, d represents the bias vector, and φ represents the nonlinear mapping function.

[0031] As a preferred solution of the cross-border e-commerce platform customer value evaluation method based on artificial intelligence described in the present invention, the anomaly detection step in step S5 is detected by combining the isolation forest algorithm with a deep autoencoder.

[0032] As a preferred solution of the cross-border e-commerce platform customer value evaluation method based on artificial intelligence described in the present invention, wherein: in step S5, the customer behavior time series data is input into the deep autoencoder and the anomaly detection algorithm to detect abnormal returns and fraud risk behaviors, and the step of generating an early warning signal is as follows:

[0033] For customer behavior time series data x, or its aggregated representation, a deep autoencoder is used for encoding. The encoding process is defined as:

[0034] h=f(x)=σ(W e x+b e ),

[0035] Where h represents the encoded feature vector, f(x) represents the encoding function, x represents the input customer behavior time series data vector, and W e represents the encoding weight matrix, b e represents the encoding bias vector, σ represents the activation function;

[0036] The decoding process reconstructs the encoded vector h into the output vector, the formula is:

[0037]

[0038] in, represents the reconstructed vector, g(h) represents the decoding function, W d represents the decoding weight matrix, b d represents the decoding bias vector, and σ represents the activation function.

[0039] As a preferred embodiment of the method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence described in the present invention, in step S5, the step of inputting the customer behavior time series data into a deep autoencoder and an anomaly detection algorithm to detect abnormal returns and fraud risk behaviors and generate an early warning signal also includes:

[0040] The reconstruction error is calculated using the Euclidean distance, and the formula is:

[0041]

[0042] Where E represents the reconstruction error, ‖·‖2 represents the Euclidean distance;

[0043] The reconstruction error E is mapped to anomaly score as follows:

[0044] s=ψ(E),

[0045] Where s represents the anomaly score, ψ represents the score mapping function;

[0046] After setting the threshold τ, the indicator function is used to convert the abnormal score into a warning signal. The formula is:

[0047] P=1 {s>τ} ,

[0048] Among them, P represents the warning signal, 1 {s>τ} It means that when s is greater than the threshold τ, it takes 1, otherwise it takes 0, and τ represents the detection threshold.

[0049] As a preferred solution of the cross-border e-commerce platform customer value evaluation method based on artificial intelligence described in the present invention, wherein: in step S6, the step of updating the customer value evaluation index according to the customer behavior time series data and the early warning signal and stratifying the customers is,

[0050] Combine the historical customer value index V0 with the current behavior time series data aggregation vector T agg And the early warning signal P, the weighted update formula is used to calculate the updated customer value index vector V:

[0051] V=ηV0+(1-η)(T agg +δP),

[0052] Where V represents the updated customer value indicator vector, V0 represents the previous customer value indicator vector, η represents the smoothing coefficient, T agg represents the behavioral time series data aggregation vector, δ represents the warning signal weight, and P represents the warning signal vector.

[0053] As a preferred solution of the cross-border e-commerce platform customer value evaluation method based on artificial intelligence described in the present invention, in step S6, the step of updating the customer value evaluation index according to the customer behavior time series data and the early warning signal and stratifying the customers further includes:

[0054] The indicator vectors of all customers are stratified and the clustering method is used to solve the grouping. The mathematical expression is:

[0055]

[0056] Among them, V i represents the value indicator vector of the i-th customer, μ k(i) represents the cluster center to which the i-th customer belongs, N represents the total number of customers, |·| 2 represents the squared Euclidean distance.

[0057] The beneficial effects of the present invention are as follows: the present invention adopts digital twin technology to fuse customer behavior time series data and multimodal features into virtual images through nonlinear mapping, so that discrete data is converted into continuous image display, thereby intuitively presenting the evolution process of customer behavior; the time series data mean is aggregated and mapped together with the multimodal features, and the weight matrix and bias are used to adjust the influence of each data to achieve effective docking between data, and intuitively display the correspondence between customer status and behavior.

[0058] The present invention uses a deep autoencoder to encode and decode customer behavior data, calculates the reconstruction error, converts it into anomaly scores through a scoring mapping function, and then combines it with the isolation forest algorithm for anomaly detection, thereby generating early warning signals and promptly identifying the risks of surges in returns and abnormal transactions, thus making up for the shortcomings of traditional methods in real-time risk monitoring.

[0059] In addition, the present invention uses a weighted update formula to integrate historical customer value indicators, current time series data and early warning signals, and then adopts a clustering algorithm to stratify customers, thereby achieving dynamic updating and scientific division of customer value indicators.

[0060] In summary, the present invention solves problems such as multi-channel data integration, insufficient expression of multi-dimensional customer characteristics, and delayed risk warning. It has the characteristics of unified parameter definition and mathematical model description, and provides a scientific and continuous customer stratification management method for cross-border e-commerce platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0062] Figure 1 This is a flow chart of the customer value evaluation method for a cross-border e-commerce platform based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0063] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0064] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0065] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0066] Example 1, reference Figure 1 This embodiment provides a method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence, comprising the following steps:

[0067] Step S1, collect order, evaluation, return and channel information, and standardize the collected data to obtain pre-processed data;

[0068] Step S2: Process the preprocessed data using natural language processing, image recognition, and speech analysis technologies to extract multimodal features, and use a deep learning model to generate dynamic customer tags;

[0069] In step S2, the steps of using natural language processing, image recognition and speech analysis technology to process preprocessed data to extract multimodal features and using deep learning model to generate dynamic customer tags are as follows:

[0070] Based on the preprocessed data, embedding methods are used for text, image, and speech respectively to calculate the multimodal feature vector. The definition formula is:

[0071] F=σ(W1e1+W2e2+W3e3+b),

[0072] Where F represents the multimodal feature vector, e1 represents the text embedding vector, e2 represents the image embedding vector, e3 represents the speech embedding vector, W1 represents the weight matrix corresponding to the text embedding, W2 represents the weight matrix corresponding to the image embedding, W3 represents the weight matrix corresponding to the speech embedding, b represents the bias vector, and σ represents the activation function;

[0073] The multimodal features are mapped to the label space using the fully connected layer, which is defined as:

[0074] z=UF+c,L=softmax(z),

[0075] Where z represents the fully connected layer output vector, U represents the mapping weight matrix, F represents the aforementioned multimodal feature vector, c represents the mapping bias vector, L represents the dynamic customer label vector, and softmax represents the normalization function;

[0076] Specifically, the preprocessed data is embedded with text, image, and speech features, and a unified multimodal feature vector is formed using linear combination and activation function. This feature vector is then mapped to the label space using a fully connected layer, and dynamic customer labels are generated using a normalization function. This process integrates multimodal information so that the labels reflect the multidimensional behavioral characteristics of customers in cross-border e-commerce, thereby generating dynamic labels.

[0077] The multimodal feature extraction step also includes processing social interaction data to generate dynamic customer tags containing regional, cultural, and payment habit information;

[0078] Step S3, performing time series analysis on dynamic customer tags to obtain customer behavior time series data;

[0079] Step S4: Using digital twin technology to virtually map customer behavior time series data and multimodal features to form a digital twin image of the customer;

[0080] Digital twin images are used to reflect the correspondence between customer behavior time series data and multimodal features;

[0081] In step S4, the digital twin technology is used to virtually map the customer behavior time series data and multimodal features to form the customer digital twin image.

[0082] For customer behavior time series data {t1, t2,…, t n} for mean aggregation, the calculation formula is:

[0083]

[0084] Among them, T ′ Represents the aggregated time series data vector, t i represents the eigenvector of the i-th moment, and n represents the number of moments;

[0085] The aggregated time series data vector T ′ It is fused with the multimodal feature vector F and the customer digital twin image is constructed through a nonlinear mapping function. The formula is:

[0086] D=φ(W t T ′ +W f F+d),

[0087] Where D represents the customer’s digital twin image, W t Represents the time series data conversion weight matrix, T ′ Represents the aggregated time series data vector, W f represents the multimodal feature conversion weight matrix, F represents the multimodal feature vector, d represents the bias vector, and φ represents the nonlinear mapping function;

[0088] Specifically, the customer behavior time series data is mean-aggregated to obtain a representative vector of the overall time series characteristics. This aggregated vector and the multimodal feature vector extracted in step S2 are then input into a nonlinear mapping function. The influence of different data is adjusted through the weight matrix and bias to form a customer digital twin image. The image reflects the mapping relationship between the customer's time series behavior and multimodal characteristics.

[0089] The digital twin approach helps transform discrete time-series data and multimodal information into continuous image representations, making it easier to observe changes in customer behavior;

[0090] Step S5: Input the customer behavior time series data into the deep autoencoder and anomaly detection algorithm to detect abnormal returns and fraud risk behaviors and generate early warning signals;

[0091] The anomaly detection step in step S5 is performed by combining the isolation forest algorithm with a deep autoencoder;

[0092] In step S5, the customer behavior time series data is input into the deep autoencoder and the anomaly detection algorithm to detect abnormal returns and fraud risk behaviors and generate early warning signals.

[0093] For customer behavior time series data x, or its aggregated representation, a deep autoencoder is used for encoding. The encoding process is defined as:

[0094] h=f(x)=σ(W e x+b e ),

[0095] Where h represents the encoded feature vector, f(x) represents the encoding function, x represents the input customer behavior time series data vector, and W e represents the encoding weight matrix, b e represents the encoding bias vector, σ represents the activation function;

[0096] The decoding process reconstructs the encoded vector h into the output vector, the formula is:

[0097]

[0098] in, represents the reconstructed vector, g(h) represents the decoding function, W d represents the decoding weight matrix, b d represents the decoding bias vector, σ represents the activation function;

[0099] In step S5, the customer behavior time series data is input into the deep autoencoder and the anomaly detection algorithm to detect abnormal returns and fraud risk behaviors, and the step of generating an early warning signal also includes:

[0100] The reconstruction error is calculated using the Euclidean distance, and the formula is:

[0101]

[0102] Where E represents the reconstruction error, ‖·‖2 represents the Euclidean distance;

[0103] The reconstruction error E is mapped to anomaly score as follows:

[0104] s=ψ(E),

[0105] Where s represents the anomaly score, ψ represents the score mapping function;

[0106] After setting the threshold τ, the indicator function is used to convert the abnormal score into a warning signal. The formula is:

[0107] p=1 {s>τ} ,

[0108] Among them, P represents the warning signal, 1 {s>τ} It means that when s is greater than the threshold τ, it takes 1, otherwise it takes 0, and τ represents the detection threshold;

[0109] Specifically, a deep autoencoder is used to encode and decode the input customer behavior time series data to generate a reconstruction vector. The reconstruction error is calculated using the Euclidean distance. The reconstruction error is used as the basis for anomaly detection and converted into an anomaly score through a scoring mapping function. Using a set threshold and indicator function, the anomaly score is binarized to generate an early warning signal, which is used to identify abnormal returns and fraud risk behaviors. The reconstruction error reflects data anomalies.

[0110] Step S6: updating the customer value evaluation index based on the customer behavior time series data and early warning signals, and stratifying the customers;

[0111] In step S6, the customer value evaluation index is updated according to the customer behavior time series data and the early warning signal, and the customer is stratified.

[0112] Combine the historical customer value index V0 with the current behavior time series data aggregation vector T agg And the early warning signal p, the weighted update formula is used to calculate the updated customer value index vector V:

[0113] V=ηV0+(1-η)(T agg +δP),

[0114] Where V represents the updated customer value indicator vector, V0 represents the previous customer value indicator vector, η represents the smoothing coefficient, T agg represents the behavioral time series data aggregation vector, δ represents the warning signal weight, and P represents the warning signal vector;

[0115] In step S6, the steps of updating the customer value evaluation index based on the customer behavior time series data and the early warning signal and stratifying the customers also include:

[0116] The indicator vectors of all customers are stratified and the clustering method is used to solve the grouping. The mathematical expression is:

[0117]

[0118] Among them, V i represents the value indicator vector of the i-th customer, μ k(i) represents the cluster center to which the i-th customer belongs, N represents the total number of customers, |·| 2 represents the squared Euclidean distance;

[0119] Specifically, this step integrates historical customer value data with current customer behavior information and early warning signals, and uses weighted updates to generate new customer value indicator vectors. During the update process, the smoothing coefficient adjusts the ratio of historical data to new input data, and the early warning signal intervenes with appropriate weights to reflect risk information; then a clustering method is used to group each customer indicator vector, and the Euclidean distance is used to measure the similarity between customers, thereby achieving hierarchical management.

[0120] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A customer value evaluation method for a cross-border e-commerce platform based on artificial intelligence, characterized by: include, Step S1, collect order, evaluation, return and channel information, and standardize the collected data to obtain pre-processed data; Step S2: Process the preprocessed data using natural language processing, image recognition, and speech analysis technologies to extract multimodal features, and use a deep learning model to generate dynamic customer tags; Step S3, performing time series analysis on the dynamic customer tags to obtain customer behavior time series data; Step S4: Using digital twin technology to virtually map the customer behavior time series data and multimodal features to form a customer digital twin image; Step S5: Input the customer behavior time series data into a deep autoencoder and anomaly detection algorithm to detect abnormal returns and fraud risk behaviors and generate early warning signals; The deep autoencoder encodes and decodes the input customer behavior time series data to generate a reconstruction vector. The reconstruction error is calculated using the Euclidean distance. The reconstruction error is used as the basis for anomaly detection and converted into an anomaly score through a scoring mapping function. The anomaly score is binarized using a set threshold and indicator function to generate an early warning signal for identifying abnormal returns and fraud risk behaviors. The reconstruction error reflects data anomalies. Step S6: updating the customer value evaluation index based on the customer behavior time series data and the early warning signal, and stratifying the customers; The indicator vectors of all customers are stratified and solved using clustering method.

2. The method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence according to claim 1, characterized in that: In step S2, the steps of using natural language processing, image recognition and speech analysis technology to process preprocessed data to extract multimodal features and using a deep learning model to generate dynamic customer tags are as follows: Based on the preprocessed data, embedding methods are used for text, image, and speech respectively to calculate the multimodal feature vector. The definition formula is: F=σ(W1e1+W2e2+W3e3+b), Where F represents the multimodal feature vector, e1 represents the text embedding vector, e2 represents the image embedding vector, e3 represents the speech embedding vector, W1 represents the weight matrix corresponding to the text embedding, W2 represents the weight matrix corresponding to the image embedding, W3 represents the weight matrix corresponding to the speech embedding, b represents the bias vector, and σ represents the activation function; The multimodal features are mapped to the label space using the fully connected layer, which is defined as: z=UF+c,L=softmax(z), Where z represents the fully connected layer output vector, U represents the mapping weight matrix, F represents the aforementioned multimodal feature vector, c represents the mapping bias vector, L represents the dynamic customer label vector, and softmax represents the normalization function.

3. The method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence according to claim 2, characterized in that: The multimodal feature extraction step also includes processing social interaction data to generate dynamic customer tags containing regional, cultural and payment habit information.

4. The method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence according to claim 3, characterized in that: The digital twin image is used to reflect the correspondence between customer behavior time series data and multimodal features.

5. The method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence according to claim 4, characterized in that: In step S4, the step of using digital twin technology to virtually map the customer behavior time series data and multimodal features to form a customer digital twin image is as follows: For customer behavior time series data {t1, t2,…, t n } for mean aggregation, the calculation formula is: Among them, T' represents the aggregated time series data vector, t i represents the eigenvector of the i-th moment, and n represents the number of moments; The aggregated time series data vector T' is fused with the multimodal feature vector F, and the customer digital twin image is constructed through a nonlinear mapping function. The formula is: D=φ(W t T'+W f F+d), Where D represents the customer’s digital twin image, W t represents the time series data conversion weight matrix, T' represents the aggregated time series data vector, W f represents the multimodal feature conversion weight matrix, F represents the multimodal feature vector, d represents the bias vector, and φ represents the nonlinear mapping function.

6. The method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence according to claim 5, characterized in that: The anomaly detection step in step S5 is performed by combining the isolation forest algorithm with a deep autoencoder.

7. The method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence according to claim 6, characterized in that: In step S5, the customer behavior time series data is input into a deep autoencoder and anomaly detection algorithm to detect abnormal returns and fraud risk behaviors and generate early warning signals. For customer behavior time series data x, a deep autoencoder is used for encoding. The encoding process is defined as: h=f(x)=σ(W e x+b e ), Where h represents the encoded feature vector, f(x) represents the encoding function, x represents the input customer behavior time series data vector, and W e represents the encoding weight matrix, b e represents the encoding bias vector, σ represents the activation function; The decoding process reconstructs the encoded vector h into the output vector, the formula is: in, represents the reconstructed vector, g(h) represents the decoding function, W d represents the decoding weight matrix, b d represents the decoding bias vector, and σ represents the activation function.

8. The method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence according to claim 7, characterized in that: In step S5, the customer behavior time series data is input into a deep autoencoder and anomaly detection algorithm to detect abnormal returns and fraud risk behaviors, and the step of generating an early warning signal also includes: The reconstruction error is calculated using the Euclidean distance, and the formula is: Where E represents the reconstruction error, ||·||2 represents the Euclidean distance; The reconstruction error E is mapped to anomaly score as follows: s=ψ(E), Where s represents the anomaly score, ψ represents the score mapping function; After setting the threshold τ, the indicator function is used to convert the abnormal score into a warning signal, the formula is P = 1 {s>τ} , Among them, P represents the warning signal, 1 {s>τ} It means that when s is greater than the threshold τ, it takes 1, otherwise it takes 0, and τ represents the detection threshold.

9. The method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence according to claim 8, characterized in that: In step S6, the step of updating the customer value evaluation index according to the customer behavior time series data and the early warning signal and stratifying the customers is as follows: Combine the historical customer value index V0 with the current behavior time series data aggregation vector T agg And the early warning signal P, the weighted update formula is used to calculate the updated customer value index vector V: V=ηV0+(1-η)(T agg +δP), Where V represents the updated customer value indicator vector, V0 represents the previous customer value indicator vector, η represents the smoothing coefficient, T agg represents the behavioral time series data aggregation vector, δ represents the warning signal weight, and P represents the warning signal vector.

10. The method for evaluating customer value of a cross-border e-commerce platform based on artificial intelligence according to claim 9, characterized in that: In step S6, the step of updating the customer value evaluation index according to the customer behavior time series data and the early warning signal and stratifying the customers further includes: The indicator vectors of all customers are stratified and the clustering method is used to solve the grouping. The mathematical expression is: Among them, V i Represents the value indicator vector of the i-th customer, μ k(i) represents the cluster center to which the i-th customer belongs, N represents the total number of customers, |·| 2 represents the squared Euclidean distance.

Images