Customer life cycle value prediction method based on distributed adaptive alignment

Abstract

The invention provides a customer life cycle value prediction method based on distributed adaptive alignment. The method comprises the following steps: acquiring customer data; constructing a multi-distribution batch training sample based on the data; inputting the multi-distribution batch training samples into a simple network, and outputting downstream task sharing features; downstream task sharing features are input into a distribution approximation network DAM, projected to a high-dimensional space and normalized into probability distribution; the downstream task shared features are input into a multi-view feature extraction module MFVE, and feature expressions are obtained; optimizing the probability distribution and the feature expression by adopting a joint constraint attention alignment mechanism to obtain constrained feature expression; inputting the constrained feature expression into the MLP, and mapping and outputting a prediction result; according to the method, the technical problems that prediction errors are easy to amplify and multi-module output is difficult to effectively integrate when distribution selection is wrong in a traditional method are solved, the accuracy and the stability of customer value prediction are improved, and reliable technical support is provided for intelligent business decision making.

Description

Technical Field

[0001] This invention relates to the field of machine learning in artificial intelligence technology, and in particular to a method for predicting customer lifetime value based on distribution adaptive alignment. Background Technology

[0002] With the integration of full-domain tracking and real-time data warehousing, enterprises can now capture user behavior logs across the entire lifecycle, from initial exposure, registration, activity, and payment to inactivity or churn, at millisecond levels. CLTV has become the core North Star metric for measuring marketing ROI and refining operations. However, traditional estimation methods based on historical mean, RFM, or simple regression are limited by static features, sparse labels, and delayed updates. They cannot depict the dynamic value leaps of high-potential users at different lifecycle stages, nor can they provide reliable predictions in complex distributions with zero inflation, long tails, and multiple peaks, leading to budget mismatches, lost opportunities, and outdated strategies.

[0003] A deeper challenge stems from the dual complexity of data distribution and model architecture. First, customer consumption behavior is inherently ambiguous and transitional. Traditional "hard binning" forcibly categorizes users into a single distribution, generating irreversible misclassification errors in overlapping boundary regions, which are amplified step by step along the feature learning path. Second, the high-dimensional embeddings, probability distributions, and scalar predictions output by multi-module networks differ significantly in semantic space and numerical scale, making direct concatenation or weighting prone to information conflicts. Third, existing knowledge distillation is limited to one-way transmission from teacher to student, failing to leverage complementary signals between modules to achieve bidirectional constraints and synergistic enhancement. Fourth, independent loss functions pull against each other in multi-objective optimization, resulting in unstable training and difficulty converging to the Pareto front. Fifth, a large number of zero-value customers in zero-inflated log-normal data are simply merged into low-value intervals, ignoring their unique generation mechanisms, leading to prediction bias and decreased business interpretability.

[0004] Therefore, there is an urgent need for a CLTV prediction framework for large-scale unstructured behavioral logs and structured transaction data: characterizing boundary ambiguity through soft distribution allocation, aligning and fusing heterogeneous outputs from multiple modules using bidirectional distillation, constructing a unified loss function to collaboratively optimize prediction accuracy, distribution identification accuracy, and module consistency, while explicitly modeling the zero inflation mechanism, and ultimately accurately identifying high-value and high-risk customer groups under the constraints of limited computing power, limited labeling, and real-time inference, providing a real-time, reliable, and interventionist data foundation for differentiated incentives, budget reallocation, and growth strategies. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a customer lifetime value prediction method based on distribution adaptive alignment. This method includes: acquiring user consumption behavior data; constructing multi-distribution batch training samples based on the consumption behavior data; inputting the multi-distribution batch training samples into a simplified network to output downstream task shared features; inputting the downstream task shared features into a distributed approximation network (DAM), projecting them into a high-dimensional space and normalizing them into a probability distribution; inputting the downstream task shared features into a multi-view feature extraction module (MFVE) to obtain feature representations; optimizing the probability distributions and feature representations using a joint constrained attention alignment mechanism to obtain constrained feature representations; and inputting the constrained feature representations into an MLP to map and output prediction results.

[0006] The beneficial effects of this invention are:

[0007] This invention targets the entire payment transaction history. First, it uses lifetime value as an anchor to dynamically slice n distribution batches via a sliding window, both amplifying the training samples and naturally aligning the distribution gap between high and low value ranges. Then, it concatenates the three heterogeneous features—numerical, sequential, and categorical—into a unified embedding via a lightweight MLP. This results in extremely simple parameters, very low latency, yet provides high-quality common representations for subsequent modules. Next, it initiates dual-channel parallel extraction: the Distribution Approximation Network (DAM) projects the shared embedding into an n-dimensional probability vector, providing a soft label in real time indicating which value distribution the user belongs to; the Multi-View Feature Extraction (MVFE) uses a dual approach of MMoE and linear attention, weighting from an expert perspective and amplifying cross-distribution differences to output a more detailed customer profile. To avoid overfitting and excessive generalization, it introduces a joint constraint attention alignment—a two-way constraint of hard label distillation (DAM output) and soft label distillation (MVFE global weights), with adjustable temperature, achieving simultaneous optimization of "distribution awareness" and "feature awareness." During training, a three-in-one loss mechanism (Huber regression to preserve numerical accuracy, distributed cross-entropy to ensure classification alignment, and distillation regularization to ensure module collaboration) updates all parameters in a single backpropagation, resulting in faster and more stable convergence. The inference phase retains only three stages: DAM + MVFE + MLP, enabling lightweight and efficient online deployment. The entire process is recursive-free, using pure MLP and attention stacking, supporting distributed tensor parallelism and multi-GPU, multi-batch training, making it ideal for engineering deployment with zero barriers to entry. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a customer lifetime value prediction method based on distribution adaptive alignment according to an embodiment of the present invention;

[0009] Figure 2 This is a model structure diagram of a customer lifetime value prediction method based on distribution adaptive alignment according to an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] A customer lifetime value prediction method based on distribution adaptive alignment, such as Figure 1 As shown, user shopping behavior data within Amazon's open-source dataset and across Amazon's entire ecosystem are extracted. Multi-distribution batch training samples are constructed based on this data. These samples are then input into a simplified network to output downstream task shared features. These downstream task shared features are then input into a distributed approximation network (DAM), projected into a high-dimensional space, and normalized to a probability distribution. The downstream task shared features are then input into a multi-view feature extraction module (MFVE) to obtain feature representations. A joint constrained attention alignment mechanism is used to optimize the probability distribution and feature representations, resulting in constrained feature representations. Finally, the constrained feature representations are input into an MLP to map and output prediction results.

[0012] This invention proposes a customer lifetime value prediction method based on distribution adaptive alignment. This method includes: acquiring in-domain and out-of-domain payment behavior data of the user to be identified; inputting the data into a model to obtain the customer's lifetime value; and training the customer lifetime value prediction model, which includes the following steps:

[0013] S1: Obtain payment transaction data from user mini-programs, apps, etc., both within and outside the domain, and construct multi-distribution batch training samples;

[0014] Furthermore, the multi-distribution mentioned in step S1 includes:

[0015]

[0016] In the formula, This represents the lifetime value of the i-th user. Represents a set of users. This indicates which distribution the user belongs to. Indicates the total number of users. This indicates the number of distributions to be set. This indicates the user's position number after sorting.

[0017] S2: Output the shared feature embedding of downstream tasks by passing user data through a simple network;

[0018] Furthermore, such as Figure 2 The simplified shared layer shown includes:

[0019]

[0020] In the formula, This indicates a simple shared embedding network. This represents the dimension output by the simplified network. This represents a multilayer perceptron. These represent the input numerical values, sequences, and category features, respectively.

[0021] Furthermore, sequence features Input processing calculations include:

[0022]

[0023] In the formula, This represents the sequence input after conversion into the embedding representation. This represents a weighted summation of the sequence Embedding. This represents the activation function; here, a non-normalized function is used to represent the weights, emphasizing user behavior awareness. This indicates that the input mapping is set to a fixed latitude.

[0024] S3: Input the Embedding into the distributed approximation network DAM, project it onto a high-dimensional space and normalize it into a probability distribution;

[0025] Furthermore, such as Figure 2 The distributed approximation network DAM shown includes:

[0026]

[0027] In the formula, Indicates the output latitude is The vector, Indicates the quantity of the distribution. Embedding representing shared user characteristics Represents the weight matrix and bias term, respectively, and represents the random bias term, which adds a certain amount of weight to the model.

[0028] S4: Simultaneously, the feature embedding is input into the multi-view feature extraction module MFVE to obtain a richer feature representation;

[0029] Furthermore, the multi-view feature extraction module MVFE described in step S4 includes:

[0030] Step 1: Input the shared feature embedding into the hybrid expert network (MMoE), and use a gating mechanism to perform weighted summation for each expert's perspective to obtain the multi-view representation of each distribution.

[0031] Step 2: The multi-view representation is further amplified through a linear attention mechanism to amplify the differences between distributions and output the final embedding.

[0032] Step 3: Record the scores of the linear attention calculation and output the weights of each distribution for its mechanism.

[0033] Furthermore, such as Figure 2 The calculation rules for the hybrid expert network MMoE described above are as follows:

[0034]

[0035]

[0036] In the formula, A multi-view representation of each distribution. M represents the number of distributions, and M represents the number of experts. This represents an expert network, where the output is the expert's specific perspective. Let represent a gated network, where represents the expert view weights of the i-th distribution.

[0037] Furthermore, such as Figure 2 The calculation rules for the linear attention mechanism described above are as follows:

[0038]

[0039]

[0040] In the formula, This represents the activation function. and represent the trainable weight matrices, used here to calculate the differences between the representation distributions. This represents the weight of the current distribution relative to other distributions. express All weight matrices, Indicates the final embedded customer lifetime value

[0041] S5: Design a joint constraint attention alignment mechanism to collaboratively optimize modules DAM and MVFE.

[0042] Furthermore, such as Figure 2 The calculation of the joint constraint attention alignment mechanism in step S5 as shown includes:

[0043]

[0044]

[0045]

[0046] In the formula, This represents a special softmax function with a temperature parameter T. This indicates the temperature parameter used for distillation. This represents the n-dimensional vector output by the DAM module. These represent the hard-label distillation weights and soft-label distillation weights, respectively. This represents the average weight of each distribution in the MVFE output, used here to extract information from the DAM using soft-label distillation. This represents the smallest factor to prevent extreme values.

[0047] S6: Input the constrained feature representation into the MLP mapping to output the prediction result, and simultaneously combine multiple loss update models.

[0048] Furthermore, the calculation and prediction results in step S6 follow these rules:

[0049]

[0050] In the formula, This represents the user lifetime value predicted by the model. Indicates the final embedding, This represents the activation function, used here to ensure that the output is always greater than or equal to 0.

[0051] Furthermore, the combined losses in step S6 mainly include:

[0052]

[0053]

[0054]

[0055] In the formula, This represents the predicted value of a user's lifetime value. This indicates the critical point for switching losses. Indicates batch size, This indicates the distribution number to which the user actually belongs.

[0056] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A customer lifetime value prediction method based on distribution adaptive alignment, characterized in that, include: Obtain user consumption behavior data; Construct multi-distribution batch training samples based on consumer behavior data; Multiple batches of training samples are input into a simplified network, which outputs shared features for downstream tasks. The shared features of downstream tasks are input into the distributed approximation network (DAM), projected into a high-dimensional space, and normalized to a probability distribution; the shared features of downstream tasks are input into the multi-view feature extraction module (MFVE) to obtain the feature representation. A joint constraint attention alignment mechanism is used to optimize the probability distribution and feature representation to obtain the constrained feature representation; the constrained feature representation is then input into the MLP and the prediction result is mapped out.

2. The customer lifetime value prediction method based on distribution adaptive alignment according to claim 1, characterized in that, The multi-distribution batch training samples are: ； ； in, This represents the lifetime value of the i-th user. Represents a set of users. This indicates which distribution the user belongs to. This indicates the total number of users. This indicates the number of distributions to be set. This indicates the user's position number after sorting.

3. The customer lifetime value prediction method based on distribution adaptive alignment according to claim 1, characterized in that, Data processing in multi-distribution batch training samples of simplified networks includes: ； in, This indicates a simple shared embedding network. This represents the dimension output by the simplified network. This represents a multilayer perceptron. These represent the input numerical values, sequences, and category features, respectively.

4. The customer lifetime value prediction method based on distribution adaptive alignment according to claim 3, characterized in that, Sequence features for: ； in, This represents the sequence input after conversion into the embedding representation. This represents a weighted summation over the sequence Embedding. This represents the activation function. This indicates that the input mapping is set to a fixed latitude.

5. The customer lifetime value prediction method based on distribution adaptive alignment according to claim 1, characterized in that, The Distributed Approximation Network (DAM) processes shared features from downstream tasks, including: ； in, Indicates output latitude is The vector, Indicates the quantity of the distribution. Indicates shared user characteristics. These represent the weight matrix and the bias term, respectively.

6. The customer lifetime value prediction method based on distribution adaptive alignment according to claim 1, characterized in that, The Multi-View Feature Extraction (MVFE) module processes the shared features of downstream tasks by: inputting the shared features of downstream tasks into a hybrid expert network, using a gating mechanism to perform weighted summation for each expert's perspective to obtain the multi-view representations of each distribution; passing the multi-view representations through a linear attention mechanism to obtain the final embedding; recording the scores of the linear attention calculation, and outputting the weights of each distribution for its mechanism.

7. The customer lifetime value prediction method based on distribution adaptive alignment according to claim 6, characterized in that, The computational rules for hybrid expert networks are as follows: ； ； in, A multi-view representation of each distribution. M represents the number of distributions, and M represents the number of experts. Represents an expert network, This indicates a gated network.

8. The customer lifetime value prediction method based on distribution adaptive alignment according to claim 6, characterized in that, The computational rules for the linear attention mechanism are as follows: ； ； ； in, This represents the activation function. Let represent the trainable weight matrices, This represents the weight of the current distribution relative to other distributions. express All weight matrices, This indicates the final embedded customer lifetime value.

9. The customer lifetime value prediction method based on distribution adaptive alignment according to claim 1, characterized in that, The calculation of the joint constrained attention alignment mechanism includes: ； ； ； in, This represents a special softmax function with a temperature parameter T. This indicates the temperature parameter used for distillation. This represents the n-dimensional vector output by the DAM module. These represent the hard-label distillation weights and soft-label distillation weights, respectively. This represents the average weight of each distribution in the MVFE output. This represents the smallest factor.

10. A customer lifetime value prediction method based on distribution adaptive alignment according to claim 1, characterized in that, The rules for calculating prediction results include: ； in, This represents the user lifetime value predicted by the model. Indicates the final embedding, This represents the activation function.

Images