User-Device Authentication in Mobile Banking Using APHEN for Paratuck2 Tensor Decomposition

Jérémy Charlier, Eric Falk, Radu State, Jean Hilger

TL;DR

This paper introduces APHEN, a high-performance optimization algorithm for Paratuck2 tensor decomposition, designed to handle the complex, imbalanced data found in mobile banking authentication. By combining this decomposition with LSTM neural networks, the authors provide a robust framework for profiling user behavior and predicting future banking activity to improve targeted financial advertising.

What Problem Does This Paper Address?

Modern retail banks face a data modeling challenge: how to make sense of "user-device authentication" events. With the rise of European PSD2 regulations, banks need better ways to engage customers through Personal Finance Management (PFM). However, authentication data is inherently multidimensional and "imbalanced." For example, a single user might authenticate dozens of times a day across multiple devices (phone, tablet, laptop), creating a massive disparity between the number of users and the number of connection events.

Standard matrix decompositions fail to capture these higher-order interactions. While Tensor Decomposition (the multidimensional version of matrix decomposition) is the right tool, the most popular method—CP decomposition—struggles with imbalanced data, often leading to underfitting or overfitting. The Paratuck2 decomposition is better suited for these asymmetries but is notoriously difficult and computationally expensive to solve using traditional methods like Alternating Least Squares (ALS) or Gradient Descent.

Methods

The authors propose a two-stage approach:

  1. APHEN (Approximate HEssian-based Newton): This is the paper's core algorithmic contribution. Traditional Newton methods provide incredibly fast (quadratic) convergence but require calculating the "Hessian matrix" (the second derivative), which is computationally ruinous for large tensors. APHEN bypasses this by using a finite difference approximation of the Hessian-vector product. This allows the algorithm to achieve "superlinear" convergence—faster than gradient descent but without the memory overhead of a full Newton method.
  2. Paratuck2 Decomposition: Unlike CP decomposition, which treats all dimensions equally, Paratuck2 expresses a tensor as a multiplication of matrices and diagonal tensors. This structure is specifically chosen to model the "static asymmetry" between users and the devices they use.
  3. Predictive Modeling: Once the tensor is decomposed into "latent factors" (hidden patterns of behavior), these factors are fed into various neural networks—Multi-Layer Perceptron (MLP), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM)—to predict future authentication events.

Datasets

  • Synthetic Benchmarks: Tensors ranging from $5 \times 5 \times 5$ to $100 \times 100 \times 20$ were used to test APHEN’s convergence speed and accuracy against six other algorithms (ALS, GD, Adam, etc.).
  • Real-World Data: The authors used the Los Alamos National Laboratory (LANL) anonymized user-computer authentication logs. This dataset contains over 60 million lines of logs. For the experiment, they sampled 150 users and 300 computers over 50 time intervals to build a $150 \times 300 \times 50$ tensor.

Key Insights and Ideas

  • The "Imbalance" Problem: The authors demonstrate that in imbalanced datasets (like 2 users connecting to 20 different computers), CP decomposition fails. If you set a low rank, you underfit the connections; if you set a high rank, you overfit the users. Paratuck2 solves this by allowing different latent factors for different dimensions.
  • Hessian-Free Advantage: APHEN proves that you don't need the entire Hessian matrix to get the benefits of second-order optimization. By approximating how the Hessian interacts with a descent direction, you can find the "optimal" solution with much lower numerical residual errors than standard gradient-based methods like Adam.
  • Temporal Recurrence: The study found that user authentication isn't random—it follows recurrent patterns based on daily activities. This makes the latent factors derived from tensor decomposition highly effective when paired with sequence-aware models like LSTMs.

Key Results

  • Optimization Performance: In benchmarks, APHEN consistently showed the fastest convergence speeds and the lowest residual errors. For a $15 \times 15 \times 15$ tensor, APHEN achieved nearly 100% accuracy (based on a log-residual metric), significantly outperforming BFGS and ALS.
  • Prediction Accuracy: When predicting the next two weeks of user activity on the LANL dataset, the LSTM model outperformed all others. It achieved a Mean Absolute Error (MAE) of 0.0379 and a Pearson correlation of 0.9753, indicating a very high degree of predictive power compared to Decision Trees (MAE 0.0965) or standard MLPs (MAE 0.0506).

Limitations and Future Work

  • Scalability: While APHEN is faster than its peers, the experiments were conducted on relatively small subsets of the LANL data (150 users). The authors note that improving GPU compatibility is a necessary next step to handle larger, "industrial-scale" tensors.
  • Feature Depth: The current model relies primarily on authentication timing and device labels. The authors suggest that future work should incorporate "navigation usage" (how a user moves through an app) and the specific time gaps between actions to further refine the financial awareness profiles.
  • Versatility: The authors plan to extend APHEN to other types of tensor decompositions beyond Paratuck2 to prove its general-purpose utility in the field of numerical optimization.

Summary updated

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 decompositions, a higher dimensional analogue of matrix decompositions. 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.
BibTeX
@inproceedings{charlier2018user,
  author = {Jérémy Charlier and Eric Falk and Radu State and Jean Hilger},
  title = {{User-Device Authentication in Mobile Banking Using APHEN for Paratuck2 Tensor Decomposition}},
  booktitle = {Workshops @ ICDM},
  year = {2018},
  url = {https://openreview.net/forum?id=YmKL4Yak8G},
}

Loading related content...

Figure