Icon

JKISeason3-18_​berti093

Creating the neural network

Accuracy on test: 99,438%

Should combine the uniformity columns

Feature that highly contribute to the prediction:

- Uniformity (Average "Uniformity of Cell Size" and "Uniformity of Cell Shape" to avoid multicollinearity)

Features that moderately contribute to the prediction:

- Clump Thickness
- Bare Nuclei
- Bland Chromatin

Features that contributing little to the prediction:

- Marginal Adhesion
- Mitoses
- Normal Nucleoli
- Single Epithelial Cell Size
Scoring ofthe model
Scorer
training a local GLM for each input instance
Compute LIME
Filter out the constant column
Column Filter
Transposefor thevisualization
Table Transposer
Collectexplanations
Loop End
RowID
Input layer8 nodes
Keras Input Layer
Hidden layer64 nodes, RelU
Keras Dense Layer
Oversamplethe "4" classas it is inminority
SMOTE
Filter outsample codenumber
Column Filter
Hidden layer16 nodes, RelU
Keras Dense Layer
Output layer2 nodes, Softmax
Keras Dense Layer
Start of the lime loop
LIME Loop Start
For visualization
Post-Processing
Hidden layer32 nodes, RelU
Keras Dense Layer
Reading tumor.csv
CSV Reader
Test ofthe neuralnetwork
Keras Network Executor
Batch the limeobservations
GroupBy
Check the linearcorrelationof the features
Linear Correlation
Createdummiesfrom Class
One to Many
Visualizing LIME
Bar Chart (JavaScript)
Test ofthe neuralnetwork
Keras Network Executor
Trainingof the neuralnetword
Keras Network Learner
Coding of the classregarding the probabilities
Rule Engine
Create "Uniformity"column by averaginguniformity of size and shape
Column Aggregator
80% Training20% Test
Partitioning

Nodes

Extensions

Links