FeatureFu[l] contains a collection of library/tools for advanced feature engineering, such as using extended s-expression based feature transformation, to derive features on top of other features, or convert a light weighted model (logistical regression or decision tree) into a feature, in an intuitive way without touching any code.
Sample use cases:
-
Feature normalization
"(min 1 (max (+ (* slope x) intercept) 0))" : scale feature x with slope and intercept, and normalize to [0,1]
-
Feature combination
"(‐ (log2 (+ 5 impressions)) (log2 (+ 1 clicks)))" : combine #impression and #clicks into a smoothed CTR style feature
-
Nonlinear featurization
"(if (> query_doc_matches 0) 0 1)" : negation of a query/document matching feature
-
Cascading modeling
"(sigmoid (+ (+ (..) w1) w0))" : convert a logistic regression model into a feature
-
Model combination (e.g. combine decision tree and linear regression)
"(+ (* model1_score w1) (* model2_score w2))" : combine two model scores into one final score
Expr: A super fast and simple evaluator for mathematical s-expressions written in Java.
Using it is as simple as:
VariableRegistry variableRegistry=new VariableRegistry();
Expr expression = Expression.parse("(sigmoid (+ (* a x) b))",variableRegistry);
Variable x = variableRegistry.findVariable("x");
Variable a = variableRegistry.findVariable("a");
Variable b = variableRegistry.findVariable("b");
expression.evaluate();
Map<String,Double> varMap = new HashMap<String,Double>();
varMap.put("x",0.2);
varMap.put("a",0.6);
varMap.put("b",0.8);
variableRegistry.refresh(varMap);
expression.evaluate();
gradle clean build
##Test
cd build/expr/lib
$java -cp expr-1.0.jar Expression "(+ 0.5 (* (/ 15 1000) (ln (- 55 12))))"
=(0.5+((15.0/1000.0)*ln((55.0-12.0))))
=0.5564180017354035
tree
└── +
├── 0.5
└── *
├── /
| ├── 15.0
| └── 1000.0
└── ln
└── -
├── 55.0
└── 12.0
expr is available under:
<dependency>
<groupId>com.linkedin.featurefu</groupId>
<artifactId>expr</artifactId>
<version>0.0.3</version>
</dependency>
dependencies { compile "com.linkedin.featurefu:expr:0.0.3" }