Tensor-Train-Based Tensor Union for Data Fusion With Its Multi-Modal Accurate Prediction

Influenced by the prevalence of mobile edge computing (MEC), massive consumption data is generated from customer applications. Tensor-based data representation and fusion methods are considered to effectively handle these heterogeneous consumption big data. However, the curse of dimensionality caused by high-order tensors and the limited processing power at the edge restrict the operation of data fusion on mobile edge devices. Therefore, this paper focuses on proposing a tensor-train (TT)-based lightweight data fusion method that suitable for applying in the MEC paradigm and diverse scenarios with its multi-modal accurate prediction. First, we develop a tensor union method for data fusion in various scenarios, including addition, averaging and Hadamard product. Then, a Union-Then-Decomposition in TT format (UTD-TT) algorithm is presented based on the proposed tensor union. Additionally, we propose three TT-based tensor union approaches appropriately to obtain fusion tensors in lower order space. Furthermore, a Union-By-Decomposition in TT format (UBD-TT) algorithm is presented to achieve TT-based lightweight data fusion. Extensive experimental results indicate that UTD-TT and UBD-TT have better prediction accuracy in predicting user behavior compared to the non-union approach. Besides, UBD-TT outperforms UTD-TT in computational efficiency while maintaining similar prediction accuracies.