Initial commit

readme.md

@@ -0,0 +1,8 @@
+## TrustGraph
+
+An implementation of the basic EigenTrust algorithm (http://nlp.stanford.edu/pubs/eigentrust.pdf).
+
+The algorithm is meant to find the trustworthiness of peers in a distributed system. A (potentially sparse) matrix is populated with values representing how much peers trust each other. A map is also populated with how much trust is extended by default to a sub-set of peers. From that starting point, the algorithm converges on the global trustworthiness of each peer.
+
+### To-Do
+This is a first pass. It does not yet implement distributed EigenTrust. There is also some room to improve error handling.

trustgraph.go

@@ -0,0 +1,142 @@
+// Package trustGraph is based on EigenTrust
+// http://nlp.stanford.edu/pubs/eigentrust.pdf
+package trustGraph
+
+import (
+	"errors"
+)
+
+// Group represents a group of peers. Peers need to be given unique, int IDs.
+// Certainty represents the threshold of RMS change at which the algorithm will
+// escape. Max is the maximum number of loos the algorithm will perform before
+// escaping (regardless of certainty). These default to 0.001 and 200
+// respectivly and generally don't need to be changed.
+type Group struct {
+	trustGrid    map[int]map[int]float32
+	initialTrust map[int]float32
+	Certainty    float32
+	Max          int
+	Alpha        float32
+}
+
+// NewGroup is the constructor for Group.
+func NewGroup() Group {
+	return Group{
+		trustGrid:    map[int]map[int]float32{},
+		initialTrust: map[int]float32{},
+		Certainty:    0.001,
+		Max:          200,
+		Alpha:        0.95,
+	}
+}
+
+// Add will add or override a trust relationship. The first arg is the peer who
+// is extending trust, the second arg is the peer being trusted (by the peer
+// in the first arg). The 3rd arg is the amount of trust, which must be
+func (g Group) Add(truster, trusted int, amount float32) (err error) {
+	err = float32InRange(amount)
+	if err == nil {
+		a, ok := g.trustGrid[truster]
+		if !ok {
+			a = map[int]float32{}
+			g.trustGrid[truster] = a
+		}
+		a[trusted] = amount
+	}
+	return
+}
+
+// InitialTrust sets the vaulues used to seed the calculation as well as the
+// corrective factor used by Alpha.
+func (g Group) InitialTrust(trusted int, amount float32) (err error) {
+	err = float32InRange(amount)
+	if err == nil {
+		g.initialTrust[trusted] = amount
+	}
+	return
+}
+
+// float32InRange is a helper to check that a value is 0.0 <= x <= 1.0
+func float32InRange(x float32) error {
+	if x < 0 {
+		return errors.New("Trust amount cannot be less than 0")
+	}
+	if x > 1 {
+		return errors.New("Trust amount cannot be greater than 1")
+	}
+	return nil
+}
+
+// Compute will approximate the trustworthyness of each peer from the
+// information known of how much peers trust eachother.
+// It wil loop, upto g.Max times or until the average difference between
+// iterations is less than g.Certainty.
+func (g Group) Compute() map[int]float32 {
+	if len(g.initialTrust) == 0 {
+		return map[int]float32{}
+	}
+	t0 := g.initialTrust //trust map for previous iteration
+
+	for i := 0; i < g.Max; i++ {
+		t1 := *g.computeIteration(&t0) // trust map for current iteration
+		d := avgD(&t0, &t1)
+		t0 = t1
+		if d < g.Certainty {
+			break
+		}
+	}
+
+	return t0
+}
+
+// computeIteration is broken out of Compute to aid comprehension. It is the
+// inner loop of Compute. It loops over every value in t (the current trust map)
+// and looks up how much trust that peer extends to every other peer. The
+// product of the direct trust and indirect trust
+func (g Group) computeIteration(t0 *map[int]float32) *map[int]float32 {
+
+	t1 := map[int]float32{}
+	for truster, directTrust := range *t0 {
+		for trusted, indirectTrust := range g.trustGrid[truster] {
+			if trusted != truster {
+				t1[trusted] += directTrust * indirectTrust
+			}
+		}
+	}
+
+	// normalize the trust values
+	// in the EigenTrust paper, this was not done every step, but I prefer to
+	// Not doing it means the diff (d) needs to be normalized in
+	// proportion to the values (because they increase with every iteration)
+	highestTrust := float32(0)
+	for _, v := range t1 {
+		if v > highestTrust {
+			highestTrust = v
+		}
+	}
+	//Todo handle highestTrust == 0
+	for i, v := range t1 {
+		t1[i] = (v/highestTrust)*g.Alpha + (1-g.Alpha)*g.initialTrust[i]
+	}
+
+	return &t1
+}
+
+// abs is helper to take abs of float32
+func abs(x float32) float32 {
+	if x < 0 {
+		return -x
+	}
+	return x
+}
+
+// avgD is helper to compare 2 maps of float32s and return the average
+// difference between them
+func avgD(t0, t1 *map[int]float32) float32 {
+	d := float32(0)
+	for i, v := range *t1 {
+		d += abs(v - (*t0)[i])
+	}
+	d = d / float32(len(*t0))
+	return d
+}

trustgraph_test.go

@@ -0,0 +1,113 @@
+package trustGraph
+
+import (
+	"math"
+	"math/rand"
+	"testing"
+	"time"
+)
+
+func TestBasic(t *testing.T) {
+	g := NewGroup()
+	g.Add(1, 2, 1)
+	g.Add(1, 3, .5)
+	g.Add(2, 1, 1)
+	g.Add(2, 3, .5)
+	g.Add(3, 1, 1)
+	g.Add(3, 2, 1)
+
+	g.InitialTrust(1, 1)
+
+	out := g.Compute()
+
+	if out[1] < 0.975 {
+		t.Error("Trust in node 1 should be closer to 1.00")
+	}
+
+	if out[2] < 0.93 {
+		t.Error("Trust in node 2 should be closer to 1.00")
+	}
+	if out[3] < 0.4 || out[3] > 0.6 {
+		t.Error("Trust in node 3 should be closer to 0.50")
+	}
+}
+
+func TestRand(t *testing.T) {
+	peers := 200
+	rand.Seed(time.Now().UTC().UnixNano())
+	g := NewGroup()
+
+	//randomly set actual trust values for peers
+	actualTrust := make([]float32, peers)
+	for i := 0; i < peers; i++ {
+		actualTrust[i] = rand.Float32()
+	}
+
+	// peer0 is set to and granted 100% trust
+	actualTrust[0] = 1
+	g.InitialTrust(0, 1)
+
+	// set 30% of trust values to +/- 10% of actual trust
+	for i := 0; i < peers; i++ {
+		for j := 0; j < peers; j++ {
+			if rand.Float32() > .7 {
+				g.Add(i, j, randNorm(actualTrust[j]))
+			}
+		}
+	}
+
+	// compute trust
+	out := g.Compute()
+
+	// find RMS error
+	e := float32(0)
+	for i := 0; i < peers; i++ {
+		x := actualTrust[i] - out[i]
+		e += x * x
+	}
+	e = float32(math.Sqrt(float64(e / float32(peers))))
+
+	if e > .2 {
+		t.Error("RMS Error should be less than 20% for a 30% full trust grid of 200 nodes")
+	}
+}
+
+// randNorm takes a float and returns a value within +/- 10%,
+// without going over 1
+func randNorm(x float32) float32 {
+	r := rand.Float32()*.2 + .9
+	x *= r
+	if x > 1 {
+		return 1
+	}
+	return x
+}
+
+func TestRangeError(t *testing.T) {
+	g := NewGroup()
+
+	err := g.Add(1, 2, 1.1)
+	if err.Error() != "Trust amount cannot be greater than 1" {
+		t.Error("Expected error")
+	}
+
+	err = g.Add(1, 2, -1)
+	if err.Error() != "Trust amount cannot be less than 0" {
+		t.Error("Expected error less than 0 error")
+	}
+
+	err = g.Add(1, 2, 1)
+	if err != nil {
+		t.Error("Did not expected error")
+	}
+
+	err = g.Add(1, 2, 0)
+	if err != nil {
+		t.Error("Did not expected error")
+	}
+
+	err = g.Add(1, 2, 0.5)
+	if err != nil {
+		t.Error("Did not expected error")
+	}
+}