RuView/tools/bfi_learning/RESEARCH.md

NTU-Fi HumanID: comparison contract

The current short-window LSTM is an experimental baseline. Its results do not establish state of the art or identity recognition from the local BFI hardware. This note records sources checked on 2026-09-29 and the conditions needed for a fair comparison.

Resolve the split names before comparing results

MEASURED archive inventory: the public NTU-Fi-HumanID.zip contains 294 files in train_amp and 546 in test_amp, across 14 identities. Its SHA256 is 1fef4e66b088b94160b14f8b950538a5455f5a62514146e17e53d19f3732a094. Every file contains CSIamp, a real double array with shape (342, 2000). The private inventory, numeric validation, and per-file hashes reproduce these observations. No exact duplicate arrays were found, before or after decimation; that check does not establish independence of physical acquisition sessions.

The author code explicitly trains on test_amp and evaluates on train_amp for HumanID. This explains the README's 546 training / 294 test counts. Using the directory names literally produces a different experiment. See the HumanID branch in the author's loader configuration and the dataset description.

The initial local protocol used literal directory names. Preserve its split and receipts under an explicit archive-directory protocol label. A later author protocol needs a separate manifest and run. Do not relabel a completed run or swap its train and test assignments after inspecting results. Development on both assignments also prevents treating their overlapping records as a fresh, untouched holdout.

Published reference and differences

CLAIMED by SenseFi, supervised Table 3:

Model	HumanID accuracy	Parameters
CNN-5	97.14%	0.478 million
GRU	98.96%	0.079 million
LSTM	97.19%	0.105 million
BiLSTM	99.38%	0.210 million

These are paper results, averaged over three seeds. They are not measurements of this implementation. The paper's generic 80:20 split statement also differs from its HumanID count table, so reproduce an explicitly identified code/data version. SenseFi paper, Tables 2–3 and implementation details.

The author loader takes every fourth time sample and presents one full (3, 114, 500) recording to the model. Its LSTM has one layer with hidden width 64 and one prediction per recording. Our bounded LSTM uses shorter windows; averaging their probabilities is a different model and evaluation procedure. Author loader, author model.

There is a normalization discrepancy to resolve as well: the paper describes min-max normalization, while the released CSI loader uses fixed constants 42.3199 and 4.9802. Their fitting population is not established here. The local trainer instead fits feature normalization exclusively on training rows. Report that difference; do not adopt unexplained constants to chase a score. Paper, released preprocessing.

Next comparisons

1. Freeze source hashes, split assignments, preprocessing, model, seed list and selection rule. Use whole training files for validation and score a frozen checkpoint on test only after selection. Keep normalization within training.
2. Report recording accuracy, balanced accuracy, confusion matrix, majority baseline, recording coverage and windows per recording. Label window metrics separately. Correlated windows are not independent trials. Physical session separation remains unknown; sample indices are not measured seconds.
3. Compare full-recording GRU and CNN-5/LeNet, then BiLSTM, under the same frozen local protocol. These are practical author baselines, with no guaranteed improvement. Exact reproduction must also match the author's model readout and training schedule. SenseFi model implementations.
4. Keep future CSI-to-BFI transfer experiments separate. Require labeled BFI recordings, independent acquisition sessions, and target-device evaluation before claiming that a public CSI classifier works on the local link.

Retain dataset attribution and the private provenance receipt. The dataset release is CC BY 4.0; the author repository's MIT code license is separate.