Abstract
As the most widely used payment method at this stage, mobile payment is more and more closely related to the blockchain economy. Traditional methods lack a certain degree of accuracy. This research proposes a feature-based and sequential-based Bilateral AM (BAM) and Convolutional Neural Network (CNN)-gated recurrent unit for the development and application of mobile payment and blockchain economy (Gated Recurrent Unit, GRU) hybrid model (BAM-CNN-GRU), select 5 feature parameters with high correlation with the blockchain for multivariate prediction. The introduction of BAM can automatically quantify the correlation between the input variables and the blockchain, and strengthen the expression of historical key information on the predicted output; the introduction of CNN can extract high-dimensional features that reflect the non-stationary dynamic changes of the blockchain. The proposed hybrid model achieves good results in both single-step and multi-step long-term series and multivariate input blockchain prediction. Compared with the other six methods, MAE is reduced by 75.45%, 64.74%, 62.84%, respectively. 59.41%, 45.54%, 44.16%. Compared with the BAM-GRU model, the CNN-GRU model, the GRU model, the LSTM model, the support vector machine SVM model and the BP model, the prediction accuracy of the hybrid model has been greatly improved, and it has a broader application prospect.
Introduction
Mobile payment systems have become the dominant transaction method in the modern digital landscape, increasingly intertwining with the blockchain economy. However, the underlying data for these systems is notoriously non-stationary and high-dimensional, making accurate prediction of system "health" or economic fluctuations difficult. Traditional machine learning models, such as Backpropagation (BP) neural networks and Support Vector Machines (SVM), often struggle with long-term dependencies and the "memory decay" inherent in complex time-series data. This paper introduces the BAM-CNN-GRU hybrid model, designed to bridge this gap by prioritizing relevant features and historical context.
Methodology: The BAM-CNN-GRU Architecture
The core innovation of this research is a "bilateral" architecture that applies attention mechanisms at both the beginning and the end of the data processing pipeline.
Bilateral Attention Mechanism (BAM)
The BAM component serves two primary functions:
- Feature-Based Attention (Input Side): Before the data enters the neural layers, the model evaluates five key input parameters: Sharing Economy (N1), Internet of Things (N2), Cloud Computing (Wf), Artificial Intelligence (po), and the Digital Economy (To). It assigns higher weights to the features most strongly correlated with the predicted output, effectively filtering out noise from irrelevant variables.
- Sequential-Based Attention (Output Side): After processing, a second attention layer looks back across the historical states of the recurrent unit. It identifies "key moments" in the past that are highly relevant to the current prediction, solving the lag and memory decay problems common in older models.
Feature Extraction via CNN
A 1D-Convolutional Neural Network (CNN) layer is integrated to extract high-dimensional features that reflect the non-stationary dynamic changes of the blockchain. This automated feature extraction captures complex patterns and "translation features" within the time-series data that linear analyses might overlook.
Sequential Processing with GRU
The Gated Recurrent Unit (GRU) layer handles the sequential nature of the data. Chosen for its computational efficiency compared to Long Short-Term Memory (LSTM) networks, the GRU remains highly effective at capturing the time-series dependencies necessary for economic forecasting.
Experimental Setup and Results
The study utilized a time-series dataset comprising the five performance parameters mentioned above, split into a 4:1 ratio for training and testing. The BAM-CNN-GRU model was evaluated against six baseline methods: BP, SVM, LSTM, standard GRU, CNN-GRU, and BAM-GRU.
The results demonstrate a significant leap in predictive accuracy:
- Mean Absolute Error (MAE): The proposed model achieved an MAE of 1.10. Compared to the other six methods, the MAE was reduced by 75.45% (vs BP), 64.74% (vs SVM), 62.84% (vs LSTM), 59.41% (vs GRU), 45.54% (vs CNN-GRU), and 44.16% (vs BAM-GRU).
- Stability: The Root Mean Square Error (RMSE) was reduced to 1.81, compared to 4.57 for the BP model, indicating that the hybrid model is much more consistent and less prone to large outliers in its predictions.
Discussion and Conclusion
The "bilateral" advantage of this model lies in its ability to determine both what features to prioritize and when historical data is most relevant. Beyond pure forecasting, this model offers significant utility in security; by accurately predicting blockchain fluctuations, operators can identify sudden anomalies that may signal security accidents or system failures.
While the model shows great promise, future research should address the computational overhead of such a complex architecture to ensure it can be deployed effectively in real-time mobile payment terminals. Additionally, testing on broader, publicly available financial datasets will further validate the model's generalizability across the digital economy.
Summary updated