Abstract
The new financial European regulations such as PSD2 are changing the retail banking services. Noticeably, the monitoring of the personal expenses is now opened to other institutions than retail banks. Nonetheless, the retail banks are looking to leverage the user-device authentication on the mobile banking applications to enhance the personal financial advertisement. To address the profiling of the authentication, we rely on tensor decomposition, a higher dimensional analogue of matrix decomposition. We use Paratuck2, which expresses a tensor as a multiplication of matrices and diagonal tensors, because of the imbalance between the number of users and devices. We highlight why Paratuck2 is more appropriate in this case than the popular CP tensor decomposition, which decomposes a tensor as a sum of rank-one tensors. However, the computation of Paratuck2 is computational intensive. We propose a new APproximate HEssian-based Newton resolution algorithm, APHEN, capable of solving Paratuck2 more accurately and faster than the other popular approaches based on alternating least square or gradient descent. The results of Paratuck2 are used for the predictions of users' authentication with neural networks. We apply our method for the concrete case of targeting clients for financial advertising campaigns based on the authentication events generated by mobile banking applications.
Introduction
Modern retail banks face a significant data modeling challenge in interpreting "user-device authentication" events. With the implementation of European PSD2 regulations, banks must find innovative ways to engage customers through Personal Finance Management (PFM). Authentication data is inherently multidimensional and characterized by a severe imbalance; for instance, a single user may authenticate multiple times daily across various devices (smartphones, tablets, laptops), creating a vast disparity between the number of users and connection events.
Standard matrix decomposition techniques fail to capture these complex higher-order interactions. While tensor decomposition provides a multidimensional alternative, the widely used CANDECOMP/PARAFAC (CP) decomposition often struggles with imbalanced data, leading to issues of underfitting or overfitting. This paper focuses on the Paratuck2 decomposition, which is better suited for modeling these asymmetries but is traditionally difficult and computationally expensive to solve.
Methodology
Paratuck2 Decomposition
Unlike CP decomposition, which treats all dimensions symmetrically, Paratuck2 expresses a tensor as a product of matrices and diagonal tensors. This structure is specifically chosen to model the "static asymmetry" between users and the devices they utilize. It allows for different latent factors for different dimensions, effectively addressing the imbalance problem where a few users might interact with many devices.
The APHEN Algorithm
The core contribution of this work is APHEN (Approximate HEssian-based Newton), a high-performance optimization algorithm. Traditional Newton methods offer quadratic convergence but are computationally prohibitive due to the need to calculate the full Hessian matrix (the second derivative). APHEN overcomes this by utilizing a finite difference approximation of the Hessian-vector product. This approach enables the algorithm to achieve superlinear convergence—outperforming standard gradient descent—without the excessive memory overhead associated with a full Newton method.
Experimental Evaluation
The proposed framework was evaluated using both synthetic benchmarks and real-world data:
- Synthetic Benchmarks: Tensors of varying sizes (from $5 \times 5 \times 5$ to $100 \times 100 \times 20$) were used to compare APHEN’s convergence speed and accuracy against six established algorithms, including Alternating Least Squares (ALS), Gradient Descent (GD), and Adam.
- Real-World Data: The authors utilized anonymized user-computer authentication logs from the Los Alamos National Laboratory (LANL), comprising over 60 million entries. A sample of 150 users and 300 computers over 50 time intervals was used to construct a $150 \times 300 \times 50$ tensor.
Results and Discussion
Optimization Performance
In comparative benchmarks, APHEN consistently demonstrated superior performance. It achieved the fastest convergence speeds and the lowest residual errors. For a $15 \times 15 \times 15$ tensor, APHEN reached nearly 100% accuracy (based on a log-residual metric), significantly outperforming BFGS and ALS.
Predictive Modeling
The latent factors derived from the Paratuck2 decomposition were fed into various neural network architectures—Multi-Layer Perceptron (MLP), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM)—to predict future authentication events. The LSTM model proved most effective, capturing the recurrent temporal patterns of user behavior. On the LANL dataset, the LSTM achieved a Mean Absolute Error (MAE) of 0.0379 and a Pearson correlation of 0.9753, substantially outperforming Decision Trees (MAE 0.0965).
Conclusion
This paper introduces a robust framework for profiling user behavior in mobile banking by combining Paratuck2 tensor decomposition with the efficient APHEN optimization algorithm. While APHEN shows great promise, future work will focus on improving GPU compatibility to handle industrial-scale tensors and incorporating deeper features such as in-app navigation patterns. The authors also intend to extend APHEN to other tensor decomposition models, further establishing its utility in numerical optimization.
Summary updated