m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p...
Transcript of m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p...
![Page 1: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/1.jpg)
1
m-TSNE: A Framework for Visualizing High-Dimensional Multivariate Time Series
Minh Nguyen, Sanjay Purushotham, Hien To, and Cyrus Shahabi
{minhnngu, spurusho, hto, shahabi}@usc.eduNov. 12, 2016, Chicago, IL, USA
VAHC - Workshop on Visual Analytics in Healthcare
AMIA 2016 Annual Symposium
![Page 2: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/2.jpg)
2
Outline
• Motivation
• Related Work
– Univariate Time Series
– Multidimensional Data
• Our Framework: m-TSNE (Multivariate Time Series t-Distributed Stochastic Neighbor Embedding)
• Experiment & Results
– Human Performance ATOM-HP dataset
– Electroencephalography EEG dataset
• Summary
![Page 3: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/3.jpg)
3
Motivation
• Sensors development, e-Health platforms (EHR, mobile health,…)
• Multivariate Time Series (MTS) in Healthcare:– Human vital signs
– Remote patient monitoring data
– Medical sensors data
– Lab results
– …
![Page 4: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/4.jpg)
4
Related Work• Univariate Time Series Data:
– Jointly visualize multiple lines / stacked graphs of features
0
100
200
300
400
500
600
700
800
900
1000
0 6 1218 0 6 1218 0 6 1218 0 6 1218 0 6 1218 0 6 1218 0 6 1218
steps
0
100
200
300
400
500
600
700
800
900
1000
0 6 1218 0 6 1218 0 6 1218 0 6 1218 0 6 1218 0 6 1218 0 6 1218
steps
Multiple lines [Playfair 1786] Stacked graphs [Byron & Wattenberg ‘08]
![Page 5: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/5.jpg)
5
Related Work• Univariate Time Series Data:
– Jointly visualize multiple lines / stacked graphs of features
0
100
200
300
400
500
600
700
800
900
1000
0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18
steps total_calories average_heart_rate
peak_heart_rate lowest_heart_rate
0
200
400
600
800
1000
1200
1400
0 6 1218 0 6 1218 0 6 1218 0 6 1218 0 6 1218 0 6 1218 0 6 1218
steps total_calories average_heart_rate
peak_heart_rate lowest_heart_rate
Multiple lines [Playfair 1786] Stacked graphs [Byron & Wattenberg ‘08]
![Page 6: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/6.jpg)
6
Related Work• Multidimensional Data
– Use Radar chart [Chambers ‘83], Parallel coordinates [Inselberg ‘85] approach
0.00E+00
5.00E-01
1.00E+00
1.50E+00
2.00E+00
2.50E+00
3.00E+00steps
total_calories
average_heart_ratepeek_heart_rate
lowest_heart_rate
![Page 7: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/7.jpg)
7
Related Work• Multidimensional Data
– Use Radar chart [Chambers ‘83], Parallel coordinates [Inselberg ‘85] approach
0.00E+00
5.00E-01
1.00E+00
1.50E+00
2.00E+00
2.50E+00
3.00E+00
3.50E+00
4.00E+004.50E+00
steps
total_calories
average_heart_ratepeek_heart_rate
lowest_heart_rate
![Page 8: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/8.jpg)
8
Our Framework: m-TSNE
Challenges:
• Long time series
• Multiple dimensions
• Comparing multiple Multivariate Time Series (MTS)
Previous techniques focus on Refine & Represent data visually.
Using Machine Learning techniques in visualization to help / support in Data Insight Observation.
![Page 9: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/9.jpg)
9
Our Framework: m-TSNE
• Consider each MTS as a data point
• Build map where points distances describe MTS similarities
• Embedding: Minimize the discrepancy between high-dimensional space MTS data points and low-dimensional space data points
High Dimensional Low DimensionalMapping
![Page 10: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/10.jpg)
10
Our Framework: m-TSNE
steps total_calories average_heart_rate peak_heart_rate lowest_heart_rate246 83 75 94 67188 86 66 75 60
24 79 63 73 560 82 65 74 59
50 81 64 77 6042 80 65 86 56
0 82 62 67 60
Raw data MTS X = {X1, X2,…, Xn}Each Xi is a feature (a univariate time series).
![Page 11: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/11.jpg)
11
Similarity of 2 MTS
• Each window/segmentation is an MTS X’ = X = {X’1, X’2,…, X’n}. X’i is a univariate time series within a window length (e.g. a day, a month)
• Different similarity metrics:EROS: Extended Frobenius norm [Yang & Shahabi ‘04]
![Page 12: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/12.jpg)
12
Low-dimensional Projection (t-SNE)• pij : high-dimensional similarity of 2 MTS data points xi and xj
• qij : low-dimensional similarity of 2 MTS data points yi and yj
• Move points using gradient descent
Low Dimensional
• Minimize Kullback-Leibler divergence: [Maaten & Hinton ‘08]
![Page 13: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/13.jpg)
13
Low-dimensional Projection (t-SNE)• pij : high-dimensional similarity of 2 MTS data points xi and xj
• qij : low-dimensional similarity of 2 MTS data points yi and yj
• Move points using gradient descent
Low Dimensional
• Minimize Kullback-Leibler divergence: [Maaten & Hinton ‘08]
![Page 14: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/14.jpg)
14
Low-dimensional Projection (t-SNE)• pij : high-dimensional similarity of 2 MTS data points xi and xj
• qij : low-dimensional similarity of 2 MTS data points yi and yj
• Move points using gradient descent
Low Dimensional
• Minimize Kullback-Leibler divergence: [Maaten & Hinton ‘08]
![Page 15: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/15.jpg)
15
Experimental Setup
2 MTS Datasets:
Analytical Technologies to Objectively Measure Human Performance (ATOM-HP) Dataset
Control vs. Alcoholic Electroencephalography (EEG) Dataset
Home monitoring data of anonymized cancer patients Control vs. alcoholic subject performing trials
5 features: Steps, Total Calories, Heartrate (average, lowest, peak)
64 features: 64 electrodes placed on the subject’s scalps
2 chemotherapy cycles: 60 days Each trial’s duration is 1s.
Data sample rate: per hour Data sample rate: 3.9-msec (256Hz)
There are 8 patients (more patients are being enrolled in this on-going study)
There are 20 subjects (10 controlled, 10 alcoholic). Each subject performs 30 trials.
![Page 16: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/16.jpg)
16
Results
• ATOM-HP Dataset
• Monitoring data of one patient• Each point is a daily MTS
• 3 Distinct Clusters of Points:High Performance (Active) daysLow Performance (Inactive) daysNoisy Sensor Data
• Any further relationship between points and chemotherapy treatment?
![Page 17: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/17.jpg)
17
Results
• ATOM-HP Dataset
![Page 18: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/18.jpg)
18
Results
• EEG Dataset
• Each point is a trial performed by a control / alcoholic subject
• Show a manifold:Inside: Control subjectOutside + Outliers: Alcoholic subject
![Page 19: m-TSNE: A Framework for Visualizing High-Dimensional ... · Low-dimensional Projection (t-SNE) •p ij: high-dimensional similarity of 2 MTS data points x i and x j •q ij: low-dimensional](https://reader035.fdocuments.net/reader035/viewer/2022081404/5f04c17a7e708231d40f8bb3/html5/thumbnails/19.jpg)
19
Summary
• Conclusion
– m-TSNE: a framework to visualize high-dimensional MTS data
– Empirical evaluation on two healthcare datasets: ATOM-HP dataset, and EEG dataset
• Future Work– More subjects / data in on-going study ATOM-HP
– Dynamically visualize high-dimensional MTS data
– Adding HCI for visualization results
– Applying using different features / variables in MTS