use best first search, not greedy routing

README.md

@@ -4,4 +4,8 @@ Hierarchical Navigable Small World - demonstration of concept implementation in
 
 Implementation mainly referenced the paper [Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs](https://arxiv.org/abs/1603.09320) however I made simplifications.
 
-I don't care about performance. That said, we can still compare the relative running time?
+I don't care about performance. That said, we can still compare the relative running time?
+
+## Update:
+
+Fixed code to avoid using greedy routing. Previous results are in result-greedy.csv. We have not yet seen a difference in results, but the logic is different.

hnsw.py

@@ -77,10 +77,10 @@ class HNSW:
         self.level_count[l] += 1
         elem = HNSWTower(coord, l)
         L = self.ep.level
-        for level in range(L, l, -1):
-            ep = self.route_layer(coord, ep, level)
         # enter points (multiple, sorted by distance from near to far)
         eps = [(distance(coord, ep.coord), ep)]
+        for level in range(L, l, -1):
+            eps = self.search_layer(coord, eps, 1, level)        
         for level in range(min(l,L), -1, -1):
             W = self.search_layer(coord, eps, self.ef_construction, level)
             neighbors = [entry[1] for entry in sorted(W)[:self.M]]
@@ -94,7 +94,7 @@ class HNSW:
         if l > L:
             self.ep = elem
     def route_layer(self, coord, ep, level):
-        # returns node that is closes to coord at level, starting from "ep"
+        # greedy routing starting at "ep". This method shouldn't be used.
         while True:
             best = None
             min_d = distance(coord, ep.coord)
@@ -142,8 +142,8 @@ class HNSW:
     def k_nn_search(self, coord, K, ef):
         ep = self.ep
         L = ep.level
-        for level in range(L, 0, -1):
-            ep = self.route_layer(coord, ep, level)
         eps = [(distance(coord, ep.coord), ep)]
+        for level in range(L, 0, -1):
+            eps = self.search_layer(coord, eps, 1, level)
         W = self.search_layer(coord, eps, ef, 0)
         return sorted(W)[:K]

result-greedy.csv

@@ -0,0 +1,21 @@
+N_corpus,Dim,M,ef_construction,build_time,K,ef,recall,query_ms,N_query,seed
+1000,10,10,50,1.2735,5,10,0.8280,0.4636,100,43
+1000,20,10,50,1.3208,5,10,0.6200,0.5730,100,43
+1000,30,10,50,1.2967,5,10,0.6540,0.5083,100,43
+1000,40,10,50,1.3569,5,10,0.6140,0.4980,100,43
+1000,50,10,50,1.3414,5,10,0.5340,0.5513,100,43
+1000,70,10,50,1.3196,5,10,0.5560,0.5107,100,43
+1000,100,10,50,1.3419,5,10,0.5160,0.5179,100,43
+1000,100,12,50,1.6164,5,10,0.5760,0.5652,100,43
+1000,100,14,50,1.8366,5,10,0.6000,0.6434,100,43
+1000,100,16,50,2.1186,5,10,0.6420,0.6895,100,43
+1000,100,18,50,2.4118,5,10,0.6880,0.7652,100,43
+1000,100,20,50,2.7659,5,10,0.6940,0.8060,100,43
+2000,100,20,50,6.2879,5,10,0.5740,0.9183,100,43
+3000,100,20,50,10.1791,5,10,0.5280,1.0457,100,43
+4000,100,20,50,14.1797,5,10,0.4720,1.1173,100,43
+5000,100,20,50,18.9001,5,10,0.4240,1.1408,100,43
+5000,100,10,50,9.2855,5,10,0.3220,0.6739,100,43
+5000,100,10,100,13.6119,5,10,0.3160,0.6867,100,43
+5000,100,10,20,6.0684,5,10,0.3080,0.6659,100,43
+5000,100,10,10,4.9333,5,10,0.2520,0.6278,100,43

result.csv

@@ -1,21 +1,2 @@
 N_corpus,Dim,M,ef_construction,build_time,K,ef,recall,query_ms,N_query,seed
-1000,10,10,50,1.2735,5,10,0.8280,0.4636,100,43
-1000,20,10,50,1.3208,5,10,0.6200,0.5730,100,43
-1000,30,10,50,1.2967,5,10,0.6540,0.5083,100,43
-1000,40,10,50,1.3569,5,10,0.6140,0.4980,100,43
-1000,50,10,50,1.3414,5,10,0.5340,0.5513,100,43
-1000,70,10,50,1.3196,5,10,0.5560,0.5107,100,43
-1000,100,10,50,1.3419,5,10,0.5160,0.5179,100,43
-1000,100,12,50,1.6164,5,10,0.5760,0.5652,100,43
-1000,100,14,50,1.8366,5,10,0.6000,0.6434,100,43
-1000,100,16,50,2.1186,5,10,0.6420,0.6895,100,43
-1000,100,18,50,2.4118,5,10,0.6880,0.7652,100,43
-1000,100,20,50,2.7659,5,10,0.6940,0.8060,100,43
-2000,100,20,50,6.2879,5,10,0.5740,0.9183,100,43
-3000,100,20,50,10.1791,5,10,0.5280,1.0457,100,43
-4000,100,20,50,14.1797,5,10,0.4720,1.1173,100,43
-5000,100,20,50,18.9001,5,10,0.4240,1.1408,100,43
-5000,100,10,50,9.2855,5,10,0.3220,0.6739,100,43
-5000,100,10,100,13.6119,5,10,0.3160,0.6867,100,43
-5000,100,10,20,6.0684,5,10,0.3080,0.6659,100,43
-5000,100,10,10,4.9333,5,10,0.2520,0.6278,100,43
+5000,100,10,10,5.0158,5,10,0.2520,0.6253,100,43