A Permutation-Invariant, Missing-View-Aware Multi-Perspective Recurrent Neural Network Cell for Time-Series Fusion

Recurrent models for multi-source data must combine synchronized streams while remaining usable when sources are reordered or unavailable. Multi-perspective long short-term memory (MP-LSTM) cells update a joint memory across perspectives, but a sequential update can make the result depend on source order. We implement a generalized chain comparator and replace its perspective-wise chain with shared projections and set attention over available views. All experiments use one fixed subject-independent split of the UCI Daily and Sports Activities (DSA) dataset, with subjects 7 and 8 as the two held-out test participants. On this split, the proposed cell uses 66.0% fewer parameters and has a 41.5% shorter recorded training time than the comparator. On a fixed 720-segment, seed-0 test subset, its inference-time accuracy remains 75.14% across eight random view permutations (standard deviation (SD) 0.00 percentage points), whereas the chain comparator falls from 76.39% to a mean of 24.67%. In the 10-class task, the mean full-view accuracy across three training seeds is 3.53 percentage points lower for the proposed cell. A cyclic missing-view schedule increases seed-0 accuracy with four of five views removed from 17.47% to 41.41% in the 19-class task, while three-seed mean full-view accuracy decreases by 3.40 percentage points. These results characterize order-invariant fusion and its robustness-accuracy trade-off in this controlled case study; they do not establish generalization beyond this dataset and held-out-subject split.