A customer churn early warning and dynamic value evaluation method based on a graph neural network

Abstract

The application discloses a kind of customer loss early warning and dynamic value evaluation method based on graph neural network includes, construct dynamic heterogeneous customer interaction graph with time effectiveness attenuation, utilize time series graph neural network to extract dynamic node embedding features including multidimensional evolution track;Combining Monte Carlo sampling and graph isomorphism agent model, approximate calculation is based on the customer node network contribution value of collaborative game;Individual characteristics, node embedding and network contribution feature fusion, use time series convolution network to predict customer loss probability;Linkage loss probability and network contribution value, output dynamic network adjustment value.The application breaks the evaluation limitation of traditional isolated data, realizes the linkage of risk early warning and whole network overflow value, provides intelligent retention decision support with high precision and high timeliness for enterprise.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for customer churn early warning and dynamic value assessment based on graph neural networks. Background Technology

[0002] With the rapid development of the digital economy, enterprise customer relationship management (CRM) has shifted from acquiring new customers to the refined operation of existing customers. Accurate customer churn prediction and reasonable customer lifetime value assessment have become core elements for enterprises to improve operational efficiency. Currently, the customer data accumulated by enterprises in omnichannel operations is characterized by strong heterogeneity, high temporal dynamics, and complex network relationships, posing a fundamental challenge to traditional customer analysis techniques. Traditional customer churn prediction and value assessment methods often treat individual customers as independent analysis units, relying on RFM models and traditional statistical learning methods. These methods can only extract isolated individual characteristics of customers and cannot effectively utilize the topological structure information and correlation transmission effects in massive interaction data. They generally suffer from problems such as delayed churn prediction and large discrepancies between customer value assessment and actual business contribution. In recent years, Graph Neural Networks (GNNs) have become a cutting-edge technology in this field due to their powerful modeling and high-order feature extraction capabilities for graph-structured data. Existing solutions often construct static graphs or discrete-time snapshot graphs of customer interactions, using graph convolution and graph attention networks to generate implicit representations of customer nodes, thereby completing churn probability prediction and customer value estimation. This approach, to some extent, breaks through the limitations of traditional methods and improves the analysis accuracy in complex relational scenarios.

[0003] However, existing GNN-based solutions still suffer from key technical bottlenecks: First, they lack sufficient modeling of the heterogeneity and temporal dynamics of customer interactions, failing to accurately capture the temporal evolution of customer behavior and the heterogeneous semantics of various interaction types; second, customer value assessment focuses only on the individual dimension, completely ignoring the network spillover effect and synergistic value of customers, failing to quantify the chain reaction risk of core customer churn, and resulting in severely distorted assessment results; third, while the classic Shapley value can quantify the global contribution of nodes, it suffers from combinatorial explosion in large-scale customer graphs, with time complexity increasing exponentially, and there is currently no efficient engineering solution adapted to business scenarios; fourth, churn warning and value assessment are disconnected, failing to achieve dynamic and forward-looking assessment, and making it difficult to provide effective decision support for refined enterprise operations. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a customer churn early warning and dynamic value assessment method based on graph neural networks to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a customer churn early warning and dynamic value assessment method based on graph neural networks, comprising:

[0007] Construct a dynamic heterogeneous customer interaction graph, which includes customer nodes, non-customer nodes related to customers, and edges with temporal attributes that record the interaction types and interaction times between nodes;

[0008] Based on the dynamic heterogeneous customer interaction graph, a time-series graph neural network model is used to learn and generate a dynamic node embedding vector for each customer node at different time points.

[0009] The entire set of customer nodes in the dynamic heterogeneous customer interaction graph is defined as a set of participants in a collaborative game. The characteristic function of the collaborative game is approximated by a proxy model based on a graph neural network, and the network contribution value of each customer node is calculated based on the approximated characteristic function.

[0010] By integrating the individual characteristics of each customer node, the dynamic node embedding vector, and the network contribution value, a fused feature is constructed, and the future churn risk of the customer node is predicted based on the fused feature.

[0011] By combining the future transaction value, network contribution value, and future churn risk of each customer node, the dynamic network adjustment value of each customer node is assessed.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0013] 1. Existing technologies typically evaluate customers as independent individuals, thus ignoring the social and transactional network effects among customers. This invention constructs a dynamic heterogeneous customer interaction graph with time-dependent decay and utilizes a temporal graph neural network with an attention mechanism for feature extraction. This not only accurately captures the complex heterogeneous topological relationships between nodes but also adaptively integrates the temporal evolution patterns of customer interaction trajectories, providing high-quality data for downstream risk warning and value assessment.

[0014] 2. This invention introduces Shapley value from cooperative game theory into graph topology networks, scientifically and fairly quantifying the spillover effects and associated contributions of each customer node in the entire business interaction network, enabling customer value assessment to be upgraded from a single-point static to a global dynamic.

[0015] 3. This invention employs a proxy model with graph isomorphic networks and a set-to-set readout mechanism, combined with Monte Carlo marginal contribution sampling. It transforms the originally unsolvable global traversal computation into a low-latency model forward inference process, achieving online computation of massive industrial-grade data while ensuring the mathematical rigor of value allocation. Furthermore, this invention multiplicatively couples churn prediction probability with dynamic network adjustment value. This guides the enterprise decision engine to precisely allocate limited retention resources (such as marketing budgets and human customer service) to high-risk, high-network-adjustment-value core customers, thereby improving the ROI of enterprise customer relationship management. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0017] Figure 1 This is a flowchart illustrating the overall process of a customer churn early warning and dynamic value assessment method based on graph neural networks, as described in one embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, the term "one embodiment" or "embodiment" as used 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 different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0021] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0022] Example 1

[0023] Reference Figure 1 This is the first embodiment of the present invention, which provides a customer churn early warning and dynamic value assessment method based on graph neural networks, including:

[0024] S1. Construct a dynamic heterogeneous customer interaction graph, which includes customer nodes, non-customer nodes related to customers, and edges with temporal attributes that record the interaction type and interaction time between nodes.

[0025] It's important to note that traditional customer churn prediction often abstracts customers into isolated data rows (e.g., RFM statistics), severing the connection between the customer's networked existence in the real business environment. To ensure the underlying data structure accurately reflects the complex business ecosystem and meets the extraction requirements of downstream graph neural networks (GNNs) for high-order topological features, this invention extracts and cleans massive amounts of heterogeneous data from enterprise omnichannel business systems (e.g., CRM, ERP, event logs) to construct a dynamic, heterogeneous customer interaction graph with a time scale, denoted as […]. .

[0026] Furthermore, the set of nodes in the graph Defined as the union of multiple entity types, i.e. .in, This represents a set of customer nodes, whose initial individual characteristics encompass demographic attributes and static account information. This represents a set of product or service nodes, with features including product category, unit price, and tags. This represents a set of marketing campaign nodes, with features including non-customer nodes such as campaign budget and campaign duration. Represents a set of heterogeneous interaction relationship types (e.g., edges of different types such as fund transfer, social invitation, group buying and sharing); This represents the set of dynamic weights for the aforementioned edges.

[0027] It should be noted that by introducing non-customer nodes, the aim is to break the information limitations of a single customer graph, enabling the model to implicitly learn the potential causes of customer churn through customers' preferences for specific products or activities (for example, a group of customers who purchased a batch of products with high negative reviews have a higher commonality of churn).

[0028] Furthermore, the set of edges in the graph Defined as heterogeneous interaction relationships between nodes. For any edge Represent it as a quintuple, denoted as .in, For the source and target nodes of the interaction; For the predefined interaction types, this embodiment specifically includes: social relationships between customers (e.g., inviting registration, group buying), purchase or evaluation relationships between customers and products (e.g., browsing, favorites, placing orders, negative reviews), and participation relationships between customers and marketing activities (e.g., receiving coupons, redeeming coupons). This is the timestamp of the interaction. The raw attribute values ​​that describe the intensity of the interaction (e.g., transaction amount, page dwell time, comment word count).

[0029] It should be explained that different types of interaction behaviors have drastically different representations of churn (e.g., submitting a negative review often carries a stronger churn risk signal than daily browsing). By classifying heterogeneous edges, we can better provide a semantic basis for graph attention mechanisms.

[0030] Furthermore, in real-world business scenarios, customer transaction intentions and churn tendencies evolve dynamically over time. Therefore, to overcome the deficiency of static graph models where ancient and recent interactions are weighted equally, this invention applies an exponential decay function based on time half-life to the weight of each edge before inputting the graph data into the graph neural network. It is important to emphasize that applying this exponential decay function to the edge weights not only conforms to the physical laws of human memory and the forgetting curve of interests, but also forces the model to focus on key signals that reflect recent changes in customer status (e.g., a sharp drop in recent interactions).

[0031] Specifically, for any edge At the current assessment time Dynamic weights The calculation formula is as follows:

[0032]

[0033] in, This represents the dynamic effective weight of the edge in the current system time. This value will be used as the graph adjacency matrix element value that the temporal graph neural network and surrogate model rely on when performing information aggregation. This represents the initial strength value when interactions occur between nodes. This serves as the current time reference point for the system to perform churn warnings or value assessments. This is a timestamp recording the historical interactions that occurred on the edge. For the corresponding interaction type The preset half-life parameter should be noted that this parameter has different decay rates for different business actions. For example, the half-life of browsing products may be set to only 3 days, while the half-life of purchasing large items may be set to 90 days.

[0034] It should be explained that by introducing a parameter that distinguishes between preset half-lives, this formula can dynamically calculate the impact of historical behavior on the present.

[0035] Furthermore, in order to meet the timeliness requirements of the downstream customer model, before proceeding to the next step of generating dynamic node embedding vectors, it is necessary to perform an edge weight preprocessing step on the dynamic heterogeneous customer interaction graph. This step aims to reconstruct the global edge weights by applying the exponential decay function based on the time half-life mentioned above.

[0036] Furthermore, after the above processing, the weights of recent interactions will approach the initial strength values ​​when interactions occur between nodes, while the weights of long-term interactions will decay exponentially or even approach zero (manifested as the dynamic disappearance of edges in the graph topology). Finally, it is only necessary to convert the physical business data into a multidimensional dynamic graph adjacency tensor flow, and feed it, along with the node feature matrix, as structured input data directly into the temporal graph neural network model.

[0037] S2. Based on the dynamic heterogeneous customer interaction graph, a time-series graph neural network model is adopted to learn and generate a dynamic node embedding vector for each customer node at different time points.

[0038] It should be noted that although the constructed dynamic heterogeneous customer interaction graph fully preserves the original correlation and temporal characteristics of the data, it belongs to a discrete high-dimensional symbolic space and cannot be directly used for downstream numerical prediction and value game calculation. Therefore, the present invention designs a spatiotemporally decoupled and re-fused temporal graph neural network model, mapping the complex heterogeneous network topology and temporal evolution laws into a low-dimensional dense continuous vector space. Specifically, the feature extraction process of this temporal graph neural network model mainly covers the heterogeneous semantic aggregation in the spatial dimension and the state evolution update in the temporal dimension.

[0039] Furthermore, multiple heterogeneous meta-paths are predefined to capture different semantic neighborhood relationships.

[0040] Specifically, in real-world business scenarios, customer churn is often driven by various motivations. For example, some customers churn because they purchased inferior products, while others churn due to reduced marketing subsidies. To enable the model to accurately distinguish these structural features, this invention predefines a set on a dynamic heterogeneous customer interaction graph. Multiple metapaths .

[0041] In one feasible implementation of this embodiment, the meta-path is set to 3, as follows:

[0042] Path 1: "Customer-Buy-Product-Buy-Customer" path (representing customer communities with similar preferences).

[0043] Path 2: "Customer-Entry-Marketing Campaign" path (indicating customer sensitivity to platform promotions).

[0044] Path 3: "Customer-Submission-Negative Review-Product" path (indicating high-risk transmission of negative emotions).

[0045] It should be noted that by using metapaths as information filters, the model can be guided to focus on interaction trajectories with specific business logic in a complex graph, thereby avoiding the problems of excessive smoothing of node information and semantic confusion in traditional isomorphic graph networks.

[0046] Furthermore, for each meta-path, a weighted aggregation of the neighbor node information of the target customer node is performed through a graph attention mechanism that integrates time coding.

[0047] Specifically, for each meta-path Target customer nodes At the present moment The set of neighboring nodes, denoted as Considering that interactions between the same neighbor at different times have different impacts on the current state, this invention maps time intervals to continuous temporal encoding vectors and injects them into the attention calculation. The specific calculation formula is as follows:

[0048]

[0049] in, In the metapath Next, neighboring nodes For target customer nodes At any moment Attention weights. Target customer node The hidden state vector at the previous time step; Neighboring nodes At the moment of interaction eigenvectors. It is specific to metapath The characteristic linear transformation weight matrix; It is the parameter vector of the attention mechanism; This indicates a vector concatenation operation. It is a time coding function (e.g., time harmonic coding based on cosine functions) used to convert time intervals Convert it into a vector with the same dimension as the node features.

[0050] It should be explained that by introducing a temporal encoding function to transform the time interval, the attention mechanism can not only evaluate nodes... The model can assess not only the content and intensity of the interaction, but also the timing of the interaction. For example, even if the content of a negative review from a month ago is the same as that of a negative review from yesterday, the model will automatically assign a higher attention weight to yesterday's negative review through time coding adjustments.

[0051] Furthermore, when performing neighbor information aggregation, in order for the model to accurately map the heterogeneous attributes and time-sensitive attributes in the graph, it is necessary to call the set of interaction relationship types and the set of dynamic edge weights.

[0052] Specifically, based on the calculated attention weights and combined with the dynamic weights extracted from the dynamic edge weight set, the neighbor information under the meta-path composed of specific heterogeneous interaction relationships is weighted and aggregated to obtain the spatial local aggregation features of the client node under the specific meta-path. :

[0053]

[0054] in, It is a non-linear activation function. The target customer node extracted from the dynamic edge weight set. with neighboring nodes The decay weight at the current moment.

[0055] Furthermore, by semantically fusing features from multiple heterogeneous meta-paths defined by sets of interaction relationship types (e.g., through weighted summation or concatenation), the client node can be obtained. At any moment The comprehensive spatial structure characteristics are denoted as .

[0056] Furthermore, a recurrent neural network unit is used to fuse the neighborhood information aggregated at the current time step with the node embedding vector at the previous time step, and update the dynamic node embedding vector at the current time step.

[0057] Specifically, since customer loyalty and value are a continuously evolving long-term state, simply aggregating the current comprehensive spatial structure features is insufficient. Therefore, this embodiment uses a gated recurrent unit (GRU) as the state update module in the time dimension. It takes the comprehensive spatial structure features as input to the current moment and uses the hidden state vector of the customer node from the previous moment as its hidden state memory. The update process is as follows:

[0058]

[0059] in, Target customer node The hidden state vector at the current moment.

[0060] Specifically, the GRU unit contains update and reset gates. The update gate primarily controls how much of the hidden state vector (historical value memory) from the previous time step is retained in the current time step; while the reset gate primarily controls the extent to which the comprehensive spatial structure features (new interaction events) occurring in the current time step change the existing customer state. For example, when a long-term high-net-worth active customer (historical value memory indicates high loyalty) suddenly experiences consecutive returns and complaints in the current time step (new interaction events manifesting as negative semantics), the GRU's reset gate will be activated. The model will then quickly forget some of its loyalty features, and this will be reflected in the final output. China has significantly raised its risk profile.

[0061] It should be noted that, through the above processing, for each client node in the dynamic graph, the model will not only output a static feature, but also a sequence of dynamic node embedding vector trajectories that evolves over time. It is important to note that these dynamic node embedding vectors are generated by... This network model is composed of [various components]. Furthermore, the training of this network model employs an end-to-end approach, achieving automatic optimization of the weight matrices in each layer through joint backpropagation of subsequent churn classification cross-entropy loss and value regression loss.

[0062] S3. Define all customer nodes in the dynamic heterogeneous customer interaction graph as a set of participants in a collaborative game. Approximate the feature function of the collaborative game using a proxy model based on a graph neural network, and calculate the network contribution value of each customer node based on the approximate feature function.

[0063] It's important to note that in real-world business ecosystems, customer value isn't simply the sum of individual transaction amounts; rather, it exhibits a significant network spillover effect. For instance, the loss of a network hub (KOL, community core) can trigger a chain reaction of customer attrition. Therefore, it's essential to introduce Shapley's value from collaborative game theory to accurately quantify each customer's global network contribution. However, calculating the classic Shapley value requires traversing all possible customer subsets, often resulting in exponential time complexity. When faced with a customer interaction graph containing a massive number of nodes, combinatorial explosion can occur, making it difficult to implement in engineering. To address this, this invention employs an approximate estimation architecture based on a graph neural network (GNN) surrogate model combined with Monte Carlo sampling, which reduces computational complexity while ensuring mathematical rigor.

[0064] Furthermore, multiple customer subsets are randomly sampled from the dynamic heterogeneous customer interaction graph, and the corresponding real business value is calculated for each customer subset as training data.

[0065] Specifically, within the set of customer nodes, a large number of customer subsets of varying sizes are generated using a biased random walk algorithm (e.g., a variant of Node2Vec) or completely random sampling, denoted as . (in, For each sampled customer subset, extract the actual comprehensive business revenue generated by customers within that subset within a specific time window from the enterprise's historical database (e.g., the customer's own transaction volume plus the first-order conversion amount of new customers brought in through their social relationship chains). Define this as the collaborative game in the subset. The true characteristic function value on is denoted as This leads to the construction of a large amount of supervised data. ,in, The subgraph is induced from a subset of customers in a dynamic heterogeneous graph. Then, a graph neural network agent model is trained so that it can take a subgraph consisting of a subset of customers as input and output a predicted value for the business of that subgraph.

[0066] Furthermore, in order to enable the proxy model to accurately capture the impact of the extremely complex topological structure within the subgraph on its overall business value, a specific network structure is designed for the graph neural network proxy model in this embodiment. This network structure includes a graph isomorphic network layer and a set-to-set readout function.

[0067] Specifically, for the graph isomorphic network layer, this layer mainly uses the dynamic node embedding vectors in the dynamic node embedding vector sequence generated in S2 as the initial node input features of the graph isomorphic network layer, denoted as... The graph isomorphic network layer has been mathematically proven to possess expressive power equivalent to the Weisfeiler-Lehman graph isomorphism test, thus maximizing the differentiation of different subgraph topologies (e.g., the value difference between star communities and chain communities). Based on this, its... The formula for updating the node features of a layer is as follows:

[0068]

[0069] in, Represents a node In the Implicit structured representation of layers; It is a multilayer perceptron; It is a learnable scalar parameter used to adjust the weight ratio between the node's own features and the aggregated features of its neighbors; Indicates the current subgraph internal nodes The set of neighboring nodes.

[0070] Specifically, for this read function, after... After extracting the layers of the isomorphic graph network, the structured representation set of all nodes in the subgraph can be obtained, denoted as . However, since the value of collaborative game theory depends only on the nodes contained in the set and not on their order, this embodiment employs a set-to-set readout function. Inside this readout function, a multi-step iterative decoding process is performed using an attention-enabled Long Short-Term Memory (LSTM) network. Each decoding step calculates the attention weights between the query vector and the node feature set, adaptively focusing on the core nodes in the subgraph that contribute the most to business value. Ultimately, the unordered node feature set is mapped to a fixed-dimensional graph-level global representation vector. Finally, a fully connected layer outputs the predicted business value of the subgraph, denoted as […]. It is important to emphasize that, in order to integrate the model with specific business scenarios of enterprises, the logic for obtaining the true feature function values ​​in the training data is as follows: using the set of nodes in the subgraph as the target audience, the direct consumption amount generated by this audience within the sampling observation window is summarized from the enterprise's order database and traffic logs, along with the first order conversion amount generated by new customers who are guided to register through social invitation links or sharing behavior generated by this audience. The sum of the two constitutes a numerical label.

[0071] It should be noted that this surrogate model uses the mean squared error (MSE) loss function and employs end-to-end training with the true feature function values ​​of the cooperative game on a subset as labels. Once trained, this surrogate model becomes a value oracle with extremely low time complexity, capable of operating at near-zero time complexity. It allows for rapid prediction of the synergistic value of any customer portfolio within a given timeframe.

[0072] Furthermore, the network contribution value of the target customer nodes is calculated.

[0073] Specifically, after the agent model is trained, the Monte Carlo sampling method is used to approximate the Shapley value. Specifically, this is done by performing sampling on all client nodes... A completely randomized sort is generated. There are 3 different permutations. In the 1st... Randomly sorted sequence In this context, for a target customer node, the set of all nodes preceding it in the permutation is defined as the predecessor node set, denoted as . Then, using the trained graph neural network surrogate model, the predictive value of the predecessor node set is quickly inferred. And the predictive value after the target customer node is added to the set. Calculate the predicted business value increment (i.e., marginal contribution) brought about by the addition of the target customer node, denoted as . The formula is as follows:

[0074]

[0075] Finally, this target customer node will be on all The marginal contributions generated in each random sorting are averaged to determine the final network contribution value (an approximation of the Shapley value), denoted as . :

[0076]

[0077] It should be explained that in the above formula, if the network contribution value is much greater than the target customer node's own historical transaction volume, it means that the customer is a typical network amplifier (e.g., someone who gives key opinions), and their presence greatly boosts the activity of surrounding network nodes; conversely, if the network contribution value is extremely low or even negative, it means that the customer is on the edge of the network and may be a bad node that spreads negative emotions (e.g., someone who posts a lot of bad reviews).

[0078] S4. Integrate the individual characteristics of each customer node, the dynamic node embedding vector, and the network contribution value to construct a fused feature, and predict the future churn risk of customer nodes based on the fused feature.

[0079] It is important to emphasize that in this invention, customer churn decisions are influenced not only by individual traits but also by the dynamic evolution of their position within a complex interaction graph and the fluctuations in their global network influence. Based on this, this invention constructs a multi-dimensional, cross-functional fusion feature sequence and introduces a temporal convolutional network with a self-attention mechanism to achieve forward-looking prediction of future customer churn risk.

[0080] Furthermore, a fused feature is constructed by integrating the individual characteristics of each customer node, the dynamic node embedding vector, and the network contribution value.

[0081] Specifically, for any target customer node at a specific moment, its inherent static individual characteristics (such as registration location, membership level, age group, etc.) are extracted and encoded into an individual feature vector, denoted as . Simultaneously, the dynamic node embedding vectors and the network contribution value calculated by the surrogate model are obtained. It is important to emphasize that, due to the slow evolution of the network structure, the network contribution value is assumed to be locally stationary within the feature fusion window of this embodiment. Therefore, the network contribution value is broadcast and replicated along the time dimension to align it with the time series step size.

[0082] Furthermore, to eliminate the differences in dimensions and distributions of features from different sources, this embodiment uses a multilayer perceptron (MLP) to map these three parts of information onto the same dense latent semantic space for nonlinear fusion, constructing a time-series... fused feature vector The calculation formula is as follows:

[0083]

[0084] in, A linearly scaled weight matrix that contributes value to the network is used to expand the network's contribution value into high-dimensional features that match the graph representation. This represents the nonlinear fusion operation of a multilayer perceptron.

[0085] It should be noted that this fusion operation represents business interpretability. For example, when a customer's dynamic node embedding shows a declining activity trend, if their network contribution value is extremely high (i.e., the customer is a core member of the community), then during the fusion operation, forward propagation will output a fusion feature representing high-risk systemic churn due to the cross-activation of high-weight features; conversely, if the network contribution value is very low (peripheral independent customer), then a normal individual churn feature will be output. Therefore, this fusion ensures that the predictive model can not only assess who will churn, but also implicitly assess the potential destructiveness of the churn event.

[0086] Furthermore, the fused features of each customer node at multiple consecutive time points are used to form a time-series feature sequence, which is then input into a time-series prediction model to output its churn probability at future time points.

[0087] Specifically, because customer churn is a gradual evolutionary process (e.g., from active participation to browsing without purchasing and then to complete silence), snapshots of a single point in time cannot characterize this decline trajectory. Therefore, this invention extracts customer nodes in chronological order. In the past continuous The fused features within a single observation window (e.g., daily status over the past 30 days) are used to construct a temporal feature sequence matrix, which contains multiple fused feature vectors. This temporal feature sequence matrix is ​​then input into a time series prediction model composed of a temporal convolutional network with self-attention. The model's structure comprises two core modules:

[0088] The first part utilizes Temporal Convolutional Networks (TCNs) to capture local and long-term dependency patterns in temporal feature sequences. In this embodiment, the TCN employs a causal dilated convolutional architecture. It's important to explain that causal convolution ensures the model strictly adheres to the chronological order during prediction, preventing the leakage of future information; while the dilation mechanism allows the model to exponentially expand its receptive field without increasing the number of parameters, thereby capturing long-term behavioral dependencies. The TCN's... The hidden state update formula for the layer is:

[0089]

[0090] in, Indicates at time No. Layer output representation (initial input) ); For size One-dimensional convolution kernel weights; The coefficient of thermal expansion (increases with the number of layers by 1) is the coefficient of thermal expansion. (Incremental).

[0091] It should be noted that by stacking multiple layers of TCN, the model can identify long-term and short-term decline patterns, such as a step-like decline in activity over two consecutive weeks.

[0092] The second part utilizes a self-attention mechanism to identify key time points in the temporal feature sequence that play a decisive role in churn prediction. While TCN excels at extracting trends, in real-world business scenarios, customer churn is often directly caused by one or more specific triggering events (e.g., experiencing extremely poor customer service on a particular day, or participating in a poorly conducted marketing campaign). To enable the model to accurately pinpoint these key nodes, this embodiment adds a time-dimensional self-attention mechanism to the TCN output, calculated as follows:

[0093]

[0094]

[0095] in, For the linear projection matrix of the query and the key; This is the scaling factor; For a moment The attention weight of historical states on the current final predictive decision.

[0096] Furthermore, after attention-weighted summation, the comprehensive temporal representation vector of the customer can be obtained. .

[0097] It should be noted that the introduction of self-attention mechanism can make the model interpretable. It allows the model to transcend long time spans and directly focus on abnormal interactions that occurred on a specific day in the past, assigning them extremely high attention weights, thereby preventing key loss signals from being diluted by daily behavioral data during long convolutional pooling processes.

[0098] Furthermore, the probability of churn at future time points is calculated and the model is trained.

[0099] Specifically, the comprehensive temporal representation vector output by the self-attention mechanism is fed into a classifier consisting of a fully connected layer and a sigmoid activation function, which outputs the target customer node. Probability of churn within a future forecast window (e.g., the next 7 days) :

[0100]

[0101] in, and These are the weights and biases for the output layer.

[0102] Furthermore, during the model training phase, the actual customer retention / churn status is extracted from historical data and used as a label, denoted as... The churn prediction loss is constructed using the standard binary cross-entropy loss function. :

[0103]

[0104] In particular, due to the end-to-end differentiability of the architecture of this invention, the gradient of the aforementioned loss function is not only used to optimize the parameters of the time series prediction model itself (e.g., TCN convolutional kernels, attention weights), but is also directly propagated to the time series graph neural network model in S2 through the backpropagation algorithm. Therefore, this training mechanism can force the underlying graph neural network to automatically learn the graph network topology patterns and interaction meta-path features that are most likely to cause future churn, thereby achieving synergy between data-driven graph structure optimization and churn classification tasks.

[0105] S5. By combining the future transaction value, network contribution value, and future churn risk of each customer node, the dynamic network adjustment value of each customer node is assessed.

[0106] It's important to note that traditional Customer Lifetime Value (LTV) or Customer CLV assessment models often focus solely on an individual customer's historical spending and expected future purchasing power, completely ignoring the social spillover effects within the customer's network structure. This can lead to businesses missing out on key figures or customers who, despite having lower personal spending, play a crucial role in providing advice within the community when allocating retention budgets. The loss of these key individuals can inevitably trigger a chain reaction of churn-like customer attrition. To address this, this invention designs a dynamic network-adjusted value model that integrates an individual's direct commercial value with the synergistic value of their network structure, dynamically discounting it based on churn risk.

[0107] Furthermore, the value of each customer node is defined as the sum of two parts.

[0108] Specifically, the first part is the expected value of an individual transaction calculated based on its future transaction and retention probabilities. The second part is the expected value of network contribution.

[0109] Furthermore, for the target customer node, its dynamic network adjustment value is denoted as... The calculation formula is as follows:

[0110]

[0111] in, This represents the expected value of individual transactions in the first part. This represents the expected value of the network contribution in the second part.

[0112] Furthermore, the expected value of individual transactions in the first part is calculated.

[0113] Furthermore, when assessing customer value, one should not rely solely on cross-sectional data from a single point in time, but rather consider their continued contribution over multiple forecast periods (e.g., the next 3 months).

[0114] Furthermore, based on the comprehensive temporal representation vector output by the temporal convolutional network with self-attention mechanism in S4, a parallel value regression branch (e.g., a multilayer perceptron containing ReLU activation function) is used to predict the customer's future... The value of individual transactions during a period (such as expected consumption amount) is denoted as :

[0115]

[0116] Secondly, based on the customer churn probability predicted in S4, the customer churn probability in the future is converted into the probability of churn in the next churn period. The retention probability during the period is denoted as In this embodiment, the retention probability can be derived from the single-period churn probability through a survival analysis function (e.g., a discrete-time survival model or a simple exponential decay probability), representing the customer's continuous survival up to the [number missing]th period. The probability of the period. Finally, the expected value of the individual transaction is obtained. The calculation formula is as follows:

[0117]

[0118] in, The total number of periods for future business assessment; The time discount rate for money or value, used to discount future income to its present value; As a risk adjustment multiplier, it means that if the risk of customer churn is extremely high (i.e., the risk adjustment multiplier approaches 0), then no matter how high their expected spending amount is, the expected value of their conversion into real revenue for the company will be significantly reduced.

[0119] Furthermore, the expected value of the network contribution in the second part is calculated.

[0120] Furthermore, the calculated network contribution value is multiplied by an expected persistence factor, which is determined by the probability of continuous retention over multiple future periods, based on predictions of the future churn risk of customer nodes.

[0121] Specifically, the network contribution value calculated using the proxy model and Monte Carlo sampling in S3 will be incorporated into the calculation. Since the network contribution value represents the incremental marginal revenue a customer currently brings to the global network (e.g., the synergistic revenue from attracting new users and activating surrounding nodes), calculating its total expected network value over its future lifecycle requires considering the duration of its presence in the network ecosystem. Therefore, an expected persistence factor is defined, denoted as... :

[0122]

[0123] The expected value of the network contribution of the target customer node. The calculation formula is:

[0124]

[0125] It's important to explain that the expected persistence factor is essentially a discounted cumulative value of a customer's expected future survival time. Multiplying the static network contribution value by this expected persistence factor has the following physical meaning: If a KOL (with extremely high network contribution value) has a very low risk of churn (the risk adjustment multiplier remains consistently high, resulting in a large expected persistence factor), then they can continue to exert a network radiation effect for a long time, thus reaping a large expected network contribution value. Conversely, if the model predicts that the KOL has a very high churn tendency (a sharp drop in the risk adjustment multiplier leads to a very small expected persistence factor), this indicates that their network value may collapse at any time. In this case, the enterprise operating system can accurately determine the maximum budget limit for marketing intervention and resource allocation for that customer by calculating the difference between the limit value under a risk-free state and the expected value discounted under current risk (i.e., the implicit churn cost), thereby avoiding the waste of marketing resources caused by blindly issuing subsidies.

[0126] Furthermore, combining the above processing, the final dynamic network adjustment value of the target customer node is the sum of the two parts mentioned above.

[0127] Furthermore, to ensure the accuracy of the individual transaction value prediction module, during the joint training phase of the entire artificial intelligence model, the future transaction amount recorded in the company's historical data is introduced as a label, denoted as... We construct a mean squared error loss function as the value regression loss, denoted as . :

[0128]

[0129] in, This is represented as the total forecast period.

[0130] Furthermore, by performing a multi-task weighted combination of this value regression loss and the churn prediction cross-entropy loss in S4, the total loss value can be obtained:

[0131]

[0132] in, and Hyperparameters for balancing the learning weights of the two tasks.

[0133] It should be noted that by minimizing the total loss value, the entire end-to-end deep learning framework can achieve backpropagation. This ensures that the generated dynamic node embedding vectors are not only sensitive to changes in the graph topology (improving the recall rate of churn prediction), but also highly aligned with the commercial benefits directly and indirectly created by customers (improving the accuracy of value regression).

[0134] Furthermore, to ensure the stable convergence of the aforementioned architecture including the proxy model, in one feasible implementation of this embodiment, it is divided into a two-stage alternating training phase and a business inference warning phase, specifically including:

[0135] Phase 1: Pre-training phase for representation and prediction tasks. Temporarily mask the network contribution value input in S3 (or initialize it to a zero vector). Use historical customer retention / churn status labels and historically recorded future transaction amount labels as supervision signals. Utilize the total loss value to train the temporal graph neural network model in S2, as well as the temporal convolutional networks and value regression prediction processes in S4 and S5, end-to-end until the network converges. At this point, lock and obtain stable temporal graph neural network model parameters.

[0136] Phase Two: Proxy Model Training and Shapley Value Game Solving. The parameters of the time-series graph neural network model trained in Phase One are frozen. Stable dynamic node embedding vectors for all customer nodes at various historical time points are extracted using this model. Then, S3 is executed to extract customer subgraphs from the historical database of the business system and extract the true feature function values ​​of the subgraphs within the observation window as labels to train the graph neural network proxy model. Once the proxy model converges, Monte Carlo sampling approximation calculations are performed to output stable network contribution values ​​for all nodes.

[0137] Phase Three: Fine-tuning and Online Business Inference. The stable network contribution value calculated in Phase Two is fed back into the fusion operation of S4, releasing the parameter freeze of the relevant prediction networks in S4 and S5, and performing a small number of iterative fine-tuning iterations. In actual online business monitoring scenarios, daily or weekly incremental dynamic heterogeneous customer interaction graph data can be input into the trained model pipeline, automatically outputting the current dynamic network adjustment value for each customer and pushing it to the decision engine of the enterprise customer relationship management system to generate differentiated customer retention and distribution strategies (e.g., triggering the highest level of dedicated customer service intervention for customers with high dynamic network adjustment value and high churn probability).

[0138] 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 it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A customer churn early warning and dynamic value assessment method based on graph neural networks, characterized in that, include: Construct a dynamic heterogeneous customer interaction graph, which includes customer nodes, non-customer nodes related to customers, and edges with temporal attributes that record the interaction types and interaction times between nodes; Based on the dynamic heterogeneous customer interaction graph, a time-series graph neural network model is used to learn and generate a dynamic node embedding vector for each customer node at different time points. The entire set of customer nodes in the dynamic heterogeneous customer interaction graph is defined as a set of participants in a collaborative game. The characteristic function of the collaborative game is approximated by a proxy model based on a graph neural network, and the network contribution value of each customer node is calculated based on the approximated characteristic function. By integrating the individual characteristics of each customer node, the dynamic node embedding vector, and the network contribution value, a fused feature is constructed, and the future churn risk of the customer node is predicted based on the fused feature. By combining the future transaction value, network contribution value, and future churn risk of each customer node, the dynamic network adjustment value of each customer node is assessed.

2. The customer churn early warning and dynamic value assessment method based on graph neural networks as described in claim 1, characterized in that, The construction of the dynamic heterogeneous customer interaction graph includes: Define customers, products, or services as different types of nodes; The social relationships between customers, the purchase or evaluation relationships between customers and products, and the participation relationships between customers and marketing activities are defined as different types of edges, and each edge is given a timestamp of the interaction and an attribute describing the intensity of the interaction.

3. The customer churn early warning and dynamic value assessment method based on graph neural networks as described in claim 1, characterized in that, For each of the client nodes, a dynamic node embedding vector is learned and generated at different time points, including: Multiple heterogeneous meta-paths are predefined to capture different semantic neighborhood relationships; For each meta-path, a weighted aggregation of the neighbor node information of the target customer node is performed through a graph attention mechanism that integrates time coding. By using a recurrent neural network unit, the neighborhood information aggregated at the current time step is fused with the node embedding vector at the previous time step to update the dynamic node embedding vector at the current time step.

4. The customer churn early warning and dynamic value assessment method based on graph neural networks as described in claim 1 or 3, characterized in that, The value of a network's contribution is calculated, including: Multiple customer subsets are randomly sampled from the dynamic heterogeneous customer interaction graph, and the corresponding real business value is calculated for each customer subset as training data. Train a graph neural network agent model such that the agent model can receive a subgraph consisting of a subset of customers as input and output a predicted value of the business value of the subgraph. By performing multiple random sorts on all customer nodes, in each sort, the trained graph neural network agent model is used to calculate the incremental value of the predicted business brought about by the target customer node being added to its sorting predecessor node set. The average of the predicted business value increments generated by the target customer node in all random sorts is used to determine its network contribution value.

5. The customer churn early warning and dynamic value assessment method based on graph neural networks as described in claim 4, characterized in that, The trained graph neural network agent model has a network structure including a graph isomorphic network layer for learning structured representations of subgraphs, and a set-to-set readout function for aggregating node-level representations into graph-level value predictions.

6. The customer churn early warning and dynamic value assessment method based on graph neural networks as described in claim 1, characterized in that, Predicting the future churn risk of the aforementioned customer nodes includes: The fused features of each customer node at multiple consecutive time points are used to form a time-series feature sequence; The time-series feature sequence is input into a time series prediction model, which outputs its probability of loss at future time points.

7. The customer churn early warning and dynamic value assessment method based on graph neural networks as described in claim 6, characterized in that, The time series prediction model is a temporal convolutional network with a self-attention mechanism. The temporal convolutional network is used to capture local and long-term dependency patterns in the temporal feature sequence, and the self-attention mechanism is used to identify key time points in the temporal feature sequence that play a decisive role in churn prediction.

8. The customer churn early warning and dynamic value assessment method based on graph neural networks as described in claim 1, characterized in that, Evaluating the value of the dynamic network adjustment includes: The value of each customer node is defined as the sum of two parts; The first part is the individual transaction expected value calculated based on its future transaction and retention probability; The second part is the expected value of network contribution.

9. The customer churn early warning and dynamic value assessment method based on graph neural networks as described in claim 8, characterized in that, The expected value of the network contribution includes: The calculated network contribution value is multiplied by an expected persistence factor, which is determined by the probability of continuous retention over multiple future periods, based on the prediction of the future churn risk of the customer node.

10. The customer churn early warning and dynamic value assessment method based on graph neural networks as described in claim 1, characterized in that, Before learning and generating a dynamic node embedding vector for each customer node at different time points, a step of processing the edges in the dynamic heterogeneous customer interaction graph is included, the step of which includes: An exponential decay function based on time half-life is applied to the weight of each edge, so that recent interactions have higher weights in the computation of the graph neural network model and the surrogate model than interactions that occur in the distant future.

Images