This repository contains the implementation and experiments for our paper:
Robust Triple Tensor Decomposition for Corrupted and Incomplete Multiway Data
Tensors provide a principled language for structuring multiway signals across space, time, and modality, serving as a unifying scaffold for modern analysis.
They have enabled state-of-the-art advances in compression, completion, and representation learning across videos, hyperspectral imaging, and network telemetry ~[Kolda & Bader, 2009; Sedighin et al., 2024; Thanh et al., 2023; Wang et al., 2023; Vijayaraghavan et al., 2020].
Building on this foundation, our work introduces a robust TriTD framework that extends the Triple Tensor Decomposition model ~[Qi et al., 2021] to handle sparse corruptions, structured missingness, and foreground–background separation.
We propose a two-constraint ADMM optimization scheme with a convex
- Triple Tensor Decomposition (TriTD) factorization into three coupled 3-way cores
-
Robust data modeling via convex
$\ell_1$ sparse residual - Two-constraint ADMM solver with provable convergence (Lemma 1)
- Kronecker-free implementation using reshape–permute construction
-
Memory-efficient updates with small
$r^2 !\times! r^2$ Gram systems -
Benchmarked on:
- Tensor completion datasets:
sensor,taxi,network,chicago - Video background modeling:
Highway,Office,PETS2006,Sofa
- Tensor completion datasets:
We estimate low-rank cores ((A,B,C)) and sparse corruption (O) by
The augmented Lagrangian is optimized by ADMM with soft-thresholding for the sparse copy (E) and ridge-regularized least squares for ((A,B,C)).
Instead of building Kronecker products to form the mode-wise design matrices reshape/permute and a single GEMM per mode (e.g., $F=\text{RPAS}(B,C)$), reducing the cost from
- Tensor completion:
sensor,taxi,network,chicago - Video background modeling:
Highway,Office,PETS2006,Sofa(CDnet 2014; 300 consecutive frames per sequence)
Settings (completion): 10% missing; triple rank (r=5); missing treated as corrupted. Metrics: RRE and wall-clock Time (s).
| Method | sensor RRE | Time(s) | taxi RRE | Time(s) | network RRE | Time(s) | chicago RRE | Time(s) |
|---|---|---|---|---|---|---|---|---|
| Sofia | 0.341 | 15.95 | 0.584 | 598.24 | 0.963 | 12.01 | 0.352 | 194.36 |
| TRLRF | 0.316 | 25.58 | 0.280 | 1799.52 | 0.126 | 41.06 | 0.311 | 1318.22 |
| RC-FCTN | 0.337 | 2.46 | 0.380 | 128.44 | 1.083 | 5.08 | 0.247 | 29.30 |
| TTNN | 0.558 | 4.45 | 0.307 | 340.42 | 0.999 | 7.39 | 0.316 | 264.73 |
| TriTD-ADMM (Ours) | 0.279 | 2.53 | 0.338 | 53.90 | 0.143 | 1.72 | 0.321 | 20.69 |
TriTD-ADMM delivers competitive accuracy with state-of-the-art runtime across all four datasets.
Runtime (s) — lower is better
| Sequence | TTNN | Sofia | TRLRF | RC-FCTN | TriTD-ADMM (Ours) |
|---|---|---|---|---|---|
| Highway | 201.47 | 370.57 | 1031.97 | 50.64 | 33.68 |
| Sofa | 225.50 | 419.57 | 1147.48 | 56.92 | 37.05 |
| Office | 226.36 | 424.15 | 1148.17 | 56.64 | 43.98 |
| PETS2006 | 229.23 | 395.39 | 1215.11 | 92.62 | 35.93 |
Qualitatively, TriTD-ADMM cleanly separates foreground (O) from background (L) and avoids absorbing moving objects into (L) in dynamic scenes (e.g., Highway).
Per iteration, TriTD-ADMM costs
├── fast_robust_triple_tensor/ # TriTD-ADMM core (triple_decomp_ADMM.m, utilities)
├── other_methods/ # Other comparison methods
├── traffic_triple_comparison.m
└── video_triple_comparison.m
If you use this code, kindly please cite the following paper
[1] N.Q. Dang, D.M.Nhat, L.T. Thanh,N.L. Trung, K. Abed-Meraim. "Fast and Robust Triple Tensor Decomposition With Data Corruption". Proc. 51st IEEE ICASSP, 2026.