Abstract:Tensor decomposition is a popular technique for tensor completion, However most of the existing methods are based on linear or shallow model, when the data tensor becomes large and the observation data is very small, it is prone to over fitting and the performance decreases significantly. To address this problem, the completion method for a tensor based on a Biased Deep Tensor Factorization Network (BDTFN) is proposed. This method can not only overcome the shortcomings of traditional tensor factorization, but also deal with complex non-linear data. Firstly, the horizontal and lateral tensors corresponding to the observed values of the input tensors are used as inputs and projected to obtain their horizontal (lateral) potential feature tensors. Secondly, the horizontal (lateral) potential feature tensors are respectively constructed into a multilayer perceptron network. Finally, the horizontal and lateral output tensors are fused by constructing a bilinear pooling layer. Tensor forward-propagation is composed of those three step, and its parameters are updated by tensor back-propagation using the multivariable chain rule. In this paper, we consider the large-scale 5-minute traffic speed data set and use it to address the missing data imputation problem for large-scale spatiotemporal traffic data. In addition, we compare the numerical performance of the proposed algorithm with those for state-of-the-art approaches on video recovery and color image recovery. Numerical experimental results illustrate that our approach is not only much more accurate than those state-of-the-art methods, but it also has high speed.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is that when dealing with large - scale and high - dimensional spatio - temporal traffic data, existing tensor completion methods are prone to over - fitting and their performance drops significantly in the case of severe data missing. Specifically:
1. **Large - scale data and a small amount of observed data**: When the data tensor becomes very large while the observed data is very small, existing methods based on linear or shallow models are prone to over - fitting, resulting in a significant performance drop.
2. **Non - linear data processing**: Traditional tensor decomposition methods assume that the data is low - rank and can be generated by a linear combination of latent feature vectors, which is not ideal for dealing with complex and diverse missing data.
3. **Computational challenges**: Computational challenges faced when using existing methods to analyze large - scale and high - dimensional traffic data.
To solve these problems, the paper proposes a new method based on **Biased Deep Tensor Factorization Network (BDTFN)**, which not only overcomes the limitations of traditional tensor decomposition but also can handle complex non - linear data. By introducing the multi - layer perceptron network and the bilinear pooling layer, BDTFN can maintain high efficiency and accuracy when dealing with large - scale and high - dimensional data.
### Main innovation points:
- **Multi - layer perceptron network**: Construct a multi - layer perceptron network from the horizontal and lateral latent feature tensors of the input tensor.
- **Bilinear pooling layer**: Design a bilinear pooling layer, use the Hadamard product to fuse the horizontal and lateral network outputs, thereby obtaining the predicted model value.
- **Tensor back - propagation**: Calculate the gradient through the multivariate chain rule and update the parameters.
- **Combination of forward and back - propagation**: Propose a new tensor completion algorithm, extend the traditional tensor decomposition to a multi - layer neural network, and enhance the ability to handle non - linear data with complex structures.
### Application scenarios:
In the paper, a large - scale 5 - minute traffic speed data set is used to verify the effectiveness of this method in dealing with large - scale spatio - temporal traffic data, and comparative experiments are carried out with existing advanced methods. The experimental results show that BDTFN is superior to other methods in both accuracy and computational efficiency.
### Key formulas:
- **Tensor forward propagation**:
\[
\vec{U}_{i}^{(j + 1)}=\sigma\left(W_{j}^{u} * M \vec{U}_{i}^{(j)}+\vec{B}_{j}^{u}\right), \quad j = 1, \ldots, N - 1
\]
\[
\vec{V}_{i}^{(j + 1)}=\sigma\left(W_{j}^{v} * M \vec{V}_{i}^{(j)}+\vec{B}_{j}^{v}\right), \quad j = 1, \ldots, N - 1
\]
- **Bilinear pooling**:
\[
X(i, j,:)=\vec{H}^{T} * M \sigma\left(\vec{U}_{i}^{N} \odot \vec{V}_{j}^{N}\right)
\]
- **Loss function**:
\[
E(N)=\frac{1}{2n_{2}} \sum_{i = 1}^{n_{2}}\left\|H\left(\vec{Y}_{i}, c_{i}\right)\right\|_{F}^{2}+R(N)
\]
Through these improvements, BDTFN can provide a more accurate and efficient missing - data - filling solution when dealing with large - scale and high - dimensional spatio - temporal traffic data.