Please enable JavaScript.
Coggle requires JavaScript to display documents.
Tanir Ozcelebi - Behavioral context recognition (Challenges/limitations…
Tanir Ozcelebi - Behavioral context recognition
Societal motivation
breakout area --> sensors --> controlling products (like lights)
Scientific motivation
Missing 'clean' data
Imbalanced and noisy data with missing values
Data set
60 users with phones/watches
300,000+ measurements (multimodal)
Utilize 6 modalities, 166 features, 51 labels (binary classification)
Research question
Our goal is robust multi-label context classification
Challenge 1: unbalanced dataset (missing data)
Synthesis of data
Benefit: running model on synthetic data --> you don't need user data so no privacy issues
Challenge 2: Modality (sensor) loss
Achieve clean data that gives the best accuracy of data
Reconstructing modality by looking how it interactswith other modalities in the past
Challenge 3: Creating a tolerant machine learning model (with missing values)
A machine model should not stop working when it misses data
Contribution to research domain
Challenges/limitations
Challenge 1: Pretty accurate model with training on synthetic and testing on real data
Challenge 2: Reconstruction of single missing modality
Using hidden units layers
Multi label classification network
Root mean square error for each modality averaged over 5 folds
Dropping 3 modalities out of 6
Challenge 3: Modality specific network's configuration
No feature enginering
Task specific networks
Extended architecture
Results
Train the model with 4 modalities
Their model performs a little bit worse then the other (left) model, but it is a robust model
Methods
Autoencoder
Unsupervised representation learning technique (deep neural network)
Trained to reconstruct its own input x
Discriminator network
Trained to recognize wheter its input is a Gaussian instance
Implementation
Implementation and testing in Tensorflow
Use only clean data to train and evaluate synthesis
Impute noisy data
Alternative approaches
Potential applications
References
