associative array perceptron, and custom feature extractor

.gitignore

@@ -0,0 +1,2 @@
+/node_modules
+/.project

associative.js

@@ -0,0 +1,126 @@
+/**
+ * Utilities for associative arrays (= hashes = Javascript objects).
+ * @author Erel Segal-Halevi
+ */
+
+var api = {};
+
+/**
+ * add one associative array to another.
+ * @param target [input and output]
+ * @param source [input]: will be added to target.
+ */
+api.add  = function(target, source) {
+	for (var feature in source) {
+		if (!(feature in target))
+			target[feature]=0;
+		target[feature] += source[feature];
+	}
+}
+
+/**
+ * multiply one associative array by another.
+ * @param target [input and output]
+ * @param source [input]: target will be multiplied by it.
+ */
+api.multiply  = function(target, source) {
+	for (var feature in source) {
+		if (!(feature in target))
+			target[feature]=1;
+		target[feature] *= source[feature];
+	}
+}
+
+/**
+ * multiply an associative array by a scalar.
+ * @param target [input and output]
+ * @param source [input]: target will be multiplied by it.
+ */
+api.multiply_scalar  = function(target, source) {
+	for (var feature in target) {
+		target[feature] *= source;
+	}
+}
+
+/**
+ * calculate the scalar product of the given two arrays.
+ * @param features [input]
+ * @param weights [input]
+ * @note Usually, there are much less features than weights.
+ */
+api.inner_product = function(features, weights) {
+	var result = 0;
+	for (var feature in features) {
+			if (feature in weights) {
+					result += features[feature] * weights[feature]
+			} else {
+					/* the sample contains a feature that was never seen in training - ignore it for now */ 
+			}
+	}
+	return result;
+}
+
+api.sum_of_values = function(weights) {
+	var result = 0;
+	for (var feature in weights)
+		result += weights[feature];
+	return result;
+}
+
+api.sum_of_absolute_values = function(weights) {
+	var result = 0;
+	for (var feature in weights)
+		result += Math.abs(weights[feature]);
+	return result;
+}
+
+api.sum_of_square_values = function(weights) {
+	var result = 0;
+	for (var feature in weights)
+		result += Math.pow(weights[feature],2);
+	return result;
+}
+
+/**
+ * Normalize the given associative array, such that the sum of values is 1.
+ * Unless, of course, the current sum is 0, in which case, nothing is done. 
+ */
+api.normalize_sum_of_values_to_1 = function(features) {
+	var sum = api.sum_of_absolute_values(features);
+	if (sum!=0)
+		api.multiply_scalar(features, 1/sum);
+}
+
+/**
+ * Normalize the given associative array, such that the sum of squares of the values is 1.
+ * Unless, of course, the current sum is 0, in which case, nothing is done. 
+ */
+api.normalize_sum_of_squares_to_1 = function(features) {
+	var sum = api.sum_of_square_values(features);
+	if (sum!=0)
+		api.multiply_scalar(features, 1/Math.sqrt(sum));
+}
+
+
+/**
+ * @param array [input]
+ * @return a string of the given associative array, sorted by keys.
+ */
+api.stringify_sorted = function(weights, separator) {
+	var result = "{" + separator;
+	var keys = Object.keys(weights);
+	keys.sort();
+	var last = keys.length-1;
+	for (i = 0; i <= last; i++) {
+		var key = keys[i];
+		var weight = weights[key]; 
+		result += '"'+key+'": '+weight;
+		if (i<last) result+=",";
+		result += separator;
+	}
+	result += "}";
+	return result;	
+}
+
+module.exports = api;
+

demos/.gitignore

@@ -0,0 +1 @@
+/rte3model

demos/and.js

@@ -0,0 +1,36 @@
+/**
+ * A perceptron that calculates the boolean "and" function
+ */
+
+var perceptron = require('../perceptron')
+
+var and = perceptron()
+
+and.train([1, 1], 1)
+and.train([0, 1], 0)
+and.train([1, 0], 0)
+and.train([0, 0], 0)
+
+// practice makes perfect (we hope...)
+while(!and.retrain()) {}
+
+console.log("2 FEATURES: ");
+console.log(and.perceive([1, 1])) // 1
+console.log(and.perceive([0, 1])) // 0
+console.log(and.perceive([1, 0])) // 0
+console.log(and.perceive([0, 0])) // 0
+console.log(and.perceive([2, 0])) // ?
+console.log(and.perceive([0, 2])) // ?
+
+console.log("3 FEATURES: ");
+console.log(and.perceive([0, 0, 0])) // 0
+console.log(and.perceive([0, 0, 1])) // 0
+console.log(and.perceive([0, 1, 0])) // 0
+console.log(and.perceive([1, 0, 0])) // 0
+console.log(and.perceive([1, 1, 0])) // 0
+console.log(and.perceive([1, 0, 1])) // 0
+console.log(and.perceive([0, 1, 1])) // 0
+console.log(and.perceive([1, 1, 1])) // 0
+
+console.dir(and);
+

demos/rtedemo.js

@@ -0,0 +1,62 @@
+/**
+ * A perceptron for recognizing textual entailment, based on BIUTEE's proofs. 
+ */
+
+var perceptron = require('../perceptron_associative')
+  , associative = require('../associative')
+  , fs = require('fs');
+
+var opts = {
+		feature_extractor: function(pair, features) {
+			if (pair.task)
+				features[pair.task]=1; 
+			if (pair.hypothesis_length)
+				features.InverseHypothesisLength = 1 / pair.hypothesis_length;
+			for (var iStep=0; iStep<pair.proof.length; ++iStep) {
+				var step = pair.proof[iStep];
+				associative.add(features, step);
+			}
+		},
+		learningrate: 1,
+		do_averaging: false,
+		do_normalization: true,
+		debug: 0
+};
+var rte = perceptron(opts);
+
+var devpairs = JSON.parse(fs.readFileSync('rte3devproofs.json', 'utf8'));
+var testpairs = JSON.parse(fs.readFileSync('rte3testproofs.json', 'utf8'));
+
+for (var iPair=0; iPair<devpairs.length; ++iPair)
+	rte.train(devpairs[iPair], devpairs[iPair].decision);
+
+// test save and load:
+rte.save("rte3model");
+var rte2 = perceptron(opts);
+rte2.load("rte3model");
+//rte = rte2;
+
+console.log("\n\nafter first train: ");
+console.log("test on train data: ");
+console.dir(rte.test_on_train());
+
+console.log("test on test data: ");
+rte.test_start();
+for (var iPair=0; iPair<testpairs.length; ++iPair)
+	rte.test(testpairs[iPair], testpairs[iPair].decision);
+console.dir(rte.test_results());
+for (var i=1; i<=10; ++i) {
+	rte.retrain();
+	if (i==1 || i==10 || i==20) {
+		console.log("\n\nafter retrain "+i+":");
+		console.log("test on train data: ");
+		console.dir(rte.test_on_train());
+		console.log("test on test data: ");
+		rte.test_start();
+		for (var iPair=0; iPair<testpairs.length; ++iPair) 
+			rte.test(testpairs[iPair], testpairs[iPair].decision);
+		console.dir(rte.test_results());
+	}
+}
+
+

package.json

@@ -16,9 +16,12 @@
   "directories": {
     "test": "tests"
   },
-  "dependencies": {},
+  "dependencies": {
+    "mkpath": "~0.1.0",
+    "should": "~1.2.2"
+  },
   "devDependencies": {
-    "specify": "~1.1.2"
+    "specify": "*"
   },
   "keywords": [
     "machine learning",