1. Cold-Start Heterogeneous-Device Wireless Localization
Vincent Zheng
Advanced Digital Sciences Center, Singapore
April 29, 2016
Joint work with Hong Cao, Shenghua Gao, Aditi Adhikari,
Miao Lin and Kevin Chen-Chuan Chang

2. Signal-strength-based Localization
AP1
AP2
-30dBm
-50dBm
-70dBm
AP3
Build localization model (Offline Training)
Prediction location for a device (Online Testing)
A classification function
f: X  Y
as the localization model
X = (AP1, AP2, AP3)
Y = locations
(-30, -50, -70) dBm
Location y1
(-40, -45, -75) dBm
Location y2
……. …
X = (AP1, AP2, AP3) Y
(-33, -47, -73) dBm ?
……. …

3. Challenges Heterogeneous device Cold start
Traditional methods assume test data has the same distribution with the training data
Use same (or, at least very similar) device in training-testing
However, devices are heterogeneous, which hurts performance
Cold start
No calibration data on the new devices
Same y, different x

4. Problem Formulation
Transfer learning in an extreme setting with no training data in the target domain
Training
Surveyor Device (S)
Target Device (T)
X = (AP1, AP2, AP3)
Y = locations
(-30, -50, -70) dBm
Location y1
(-40, -45, -75) dBm
Location y2
……. …
X = (AP1, AP2, AP3)
Y = locations
Testing
Robust feature learning
Target Device (T)
1) Same y, same x across devices
X = (AP1, AP2, AP3)
Y = locations
(-36, -58, -79) dBm ?
(-34, -43, -65) dBm
……. …
2) Build localization model

5. Previous Work Methods Source domain labeled training data
Target domain labeled training data
Target domain unlabeled training data
LFT (Haeberlen et al. 2004) ✔ ✗
KDEFT (Park et al. 2011)
LatentMTL (Zheng et al. 2008)
ULFT (Tsui et al. 2009)
KLIEP (Sugiyama et al. 2008)
KMM (Huang et al. 2007)
SKM (Zhang et al. 2013)
HLF (Kjaergaard & Munk 2008)
ECL (Yedavalli et al. 2005)
Ours

6. Our Insights Robust features
Pairwise RSS comparison is robust (e.g., AP1 - AP2 > 0)
Phenomenon can be explained by radio propagation theory
But not discriminative (e.g., y1 vs. y2)
High-order Pairwise (HOP) features are robust and discriminative (e.g., AP3 - AP1 > AP1 - AP2)

7. Formulating HOP Features
Formulation
Pairwise comparison to encode relative AP-location closeness
Multiple pairwise comparison in a linear combination
Allow randomness in signal value variation
An HOP feature corresponds to a pair of ({ck1,k2}, b)
Eq.(4) can be rewritten as
Numeric input x, binary output h, linear mapping from x to h

8. Learning HOP Features
Robust Feature Constrained Restricted Boltzmann Machine
Gaussian-Bernoulli RBM as the feature learning framework
We extend it to incorporate the HOP constraint
Finally, we combine feature learning and localization model training
P(x(k)) is the likelihood of observation x(k) generated by a set of latent variables hj’s
L1 is to find the robust HOP features, L2 is to ensure the HOP features are discriminative

9. Experimental Setup Public data sets Baselines Evaluation metric
2 environments
6 heterogeneous devices
4 pairs of S-T devices
Baselines
Ignoring heterogeneity
SVM
Transfer learning to handle heterogeneity
HLF: heuristic pairwise AP value ratio (limited to 2nd-order)
ECL: pairwise AP value ranking (still limited to 2nd-order)
Evaluation metric
Accuracy w.r.t. error distance
A prediction is “correct” if it is within K meters of the ground truth location

10. Results
Our model achieves 23.1%– 91.3% relative accuracy improvement than the best state-of-the-art baseline.