Merge pull request #45 from antvis/feat/add-cosine-similarity

Yanyan-Wang · web-flow · commit 646a2bc40fa2 · 2021-12-16T15:03:28.000+08:00
feat: add nodesCosineSimilarity算法-基于节点属性计算余弦相似度
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,4 +1,7 @@
 # ChangeLog
+#### 0.1.17
+
+- feat: add consine-similarity algorithm and nodes-consine-similarity algorithm;
 
 #### 0.1.16
 
diff --git a/packages/graph/package.json b/packages/graph/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@antv/algorithm",
-  "version": "0.1.16",
+  "version": "0.1.17",
   "description": "graph algorithm",
   "keywords": [
     "graph",
diff --git a/packages/graph/src/cosine-similarity.ts b/packages/graph/src/cosine-similarity.ts
@@ -0,0 +1,27 @@
+import Vector from './utils/vector';
+/**
+ * cosine-similarity算法 计算余弦相似度
+ * @param item 元素
+ * @param targetItem 目标元素
+ */
+const cosineSimilarity = (
+  item: number[],
+  targetItem: number[],
+): number => {
+  // 目标元素向量
+  const targetItemVector = new Vector(targetItem);
+  // 目标元素向量的模长
+  const targetNodeNorm2 = targetItemVector.norm2();
+  // 元素向量
+  const itemVector = new Vector(item);
+  // 元素向量的模长
+  const itemNorm2 = itemVector.norm2();
+  // 计算元素向量和目标元素向量的点积
+  const dot = targetItemVector.dot(itemVector);
+  const norm2Product = targetNodeNorm2 * itemNorm2;
+  // 计算元素向量和目标元素向量的余弦相似度
+  const cosineSimilarity = norm2Product ? dot / norm2Product : 0;
+  return cosineSimilarity;
+}
+
+export default cosineSimilarity;
diff --git a/packages/graph/src/index.ts b/packages/graph/src/index.ts
@@ -12,6 +12,8 @@ import labelPropagation from './label-propagation';
 import louvain from './louvain';
 import iLouvain from './i-louvain';
 import kCore from './k-core';
+import cosineSimilarity from './cosine-similarity';
+import nodesCosineSimilarity from './nodes-cosine-similarity';
 import minimumSpanningTree from './mts';
 import pageRank from './pageRank';
 import GADDI from './gaddi';
@@ -42,6 +44,8 @@ export {
   louvain,
   iLouvain,
   kCore,
+  cosineSimilarity,
+  nodesCosineSimilarity,
   minimumSpanningTree,
   pageRank,
   getNeighbors,
@@ -71,6 +75,8 @@ export default {
   louvain,
   iLouvain,
   kCore,
+  cosineSimilarity,
+  nodesCosineSimilarity,
   minimumSpanningTree,
   pageRank,
   getNeighbors,
diff --git a/packages/graph/src/louvain.ts b/packages/graph/src/louvain.ts
@@ -2,7 +2,7 @@ import { clone } from '@antv/util';
 import getAdjMatrix from './adjacent-matrix';
 import { NodeConfig, ClusterData, GraphData, ClusterMap } from './types';
 import Vector from './utils/vector';
-import { secondReg, dateReg } from './constants/time';
+import { getPropertyWeight } from './utils/node-properties';
 
 const getModularity = (
   nodes: NodeConfig[],
@@ -28,59 +28,6 @@ const getModularity = (
   return modularity;
 }
 
-// 获取所有属性并排序
-const getAllSortProperties = (nodes: NodeConfig[] = []) => {
-  const propertyKeyInfo = {};
-  nodes.forEach(node => {
-    Object.keys(node.properties).forEach(propertyKey => {
-      // 目前过滤只保留可以转成数值型的或日期型的, todo: 统一转成one-hot特征向量
-      if (!`${node.properties[propertyKey]}`.match(secondReg) && 
-        !`${node.properties[propertyKey]}`.match(dateReg) && 
-        isNaN(Number(node.properties[propertyKey])) || propertyKey === 'id') {
-        if (propertyKeyInfo.hasOwnProperty(propertyKey)) {
-          delete propertyKeyInfo[propertyKey];
-        }
-        return;
-      }
-      if (propertyKeyInfo.hasOwnProperty(propertyKey)) {
-        propertyKeyInfo[propertyKey] += 1;
-      } else {
-        propertyKeyInfo[propertyKey] = 1;
-      }
-    })
-  })
-
-  // 取top50的属性
-  const sortKeys = Object.keys(propertyKeyInfo).sort((a,b)=>{
-    return propertyKeyInfo[b] - propertyKeyInfo[a];
-  });
-  return sortKeys.length < 100 ? sortKeys : sortKeys.slice(0, 100);
-}
-
-const processProperty = (properties, propertyKeys) => propertyKeys.map(key => {
-  if (properties.hasOwnProperty(key)) {
-    // 可以转成数值的直接转成数值
-    if (!isNaN(Number(properties[key]))) {
-      return Number(properties[key]);
-    }
-    // 时间型的转成时间戳
-    if (properties[key].match(secondReg) || properties[key].match(dateReg)) {
-      // @ts-ignore
-      return Number(Date.parse(new Date(properties[key]))) / 1000;
-    }
-  }
-  return 0;
-})
-
-// 获取属性特征权重
-const getPropertyWeight = (propertyKeys, nodes) => {
-  let allPropertiesWeight = [];
-  for (let i = 0; i < nodes.length; i++) {
-    allPropertiesWeight[i] = processProperty(nodes[i].properties, propertyKeys);
-  }
-  return allPropertiesWeight;
-}
-
 // 模块惯性度，衡量属性相似度
 const getInertialModularity = (
   nodes: NodeConfig[] = [],
@@ -171,9 +118,8 @@ const louvain = (
         node.properties.nodeType = nodeTypeInfo.findIndex(nodeType => nodeType === node.nodeType);
       })
     }
-    const propertyKeys = getAllSortProperties(nodes);
     // 所有节点属性特征向量集合
-    allPropertiesWeight = getPropertyWeight(propertyKeys, nodes);
+    allPropertiesWeight = getPropertyWeight(nodes);
   }
  
   let uniqueId = 1;
diff --git a/packages/graph/src/nodes-cosine-similarity.ts b/packages/graph/src/nodes-cosine-similarity.ts
@@ -0,0 +1,41 @@
+import { clone } from '@antv/util';
+import { NodeConfig } from './types';
+import { getPropertyWeight } from './utils/node-properties';
+import cosineSimilarity from './cosine-similarity';
+/**
+ *  nodes-cosine-similarity算法 基于节点属性计算余弦相似度(基于种子节点寻找相似节点)
+ * @param nodes 图节点数据
+ * @param seedNode 种子节点
+ */
+const nodesCosineSimilarity = (
+  nodes: NodeConfig[] = [],
+  seedNode: NodeConfig,
+): {
+  allCosineSimilarity: number[],
+  similarNodes: NodeConfig[],
+} => {
+  const similarNodes = clone(nodes.filter(node => node.id !== seedNode.id));
+  const seedNodeIndex = nodes.findIndex(node => node.id === seedNode.id);
+  // 所有节点属性特征向量集合
+  const allPropertiesWeight = getPropertyWeight(nodes);
+  // 种子节点属性
+  const seedNodeProperties = allPropertiesWeight[seedNodeIndex];
+
+  const allCosineSimilarity: number[] = [];
+  similarNodes.forEach((node, index) => {
+    if (node.id !== seedNode.id) {
+      // 节点属性
+      const nodeProperties = allPropertiesWeight[index];
+      // 计算节点向量和种子节点向量的余弦相似度
+      const cosineSimilarityValue = cosineSimilarity(nodeProperties, seedNodeProperties);
+      allCosineSimilarity.push(cosineSimilarityValue);
+      node.cosineSimilarity = cosineSimilarityValue;
+    }
+  });
+
+  // 将返回的节点按照余弦相似度大小排序
+  similarNodes.sort((a, b) => b.cosineSimilarity - a.cosineSimilarity);
+  return { allCosineSimilarity, similarNodes };
+}
+
+export default nodesCosineSimilarity;
diff --git a/packages/graph/src/utils/node-properties.ts b/packages/graph/src/utils/node-properties.ts
@@ -0,0 +1,62 @@
+import { NodeConfig } from '../types';
+import { secondReg, dateReg } from '../constants/time';
+
+// 获取所有属性并排序
+export const getAllSortProperties = (nodes: NodeConfig[] = [], n: number = 100) => {
+  const propertyKeyInfo = {};
+  nodes.forEach(node => {
+    if (!node.properties) {
+      return;
+    }
+    Object.keys(node.properties).forEach(propertyKey => {
+      // 目前过滤只保留可以转成数值型的或日期型的, todo: 统一转成one-hot特征向量或者embedding
+      if (propertyKey === 'id' || !`${node.properties[propertyKey]}`.match(secondReg) && 
+        !`${node.properties[propertyKey]}`.match(dateReg) && 
+        isNaN(Number(node.properties[propertyKey]))) {
+        if (propertyKeyInfo.hasOwnProperty(propertyKey)) {
+            delete propertyKeyInfo[propertyKey];
+        }
+        return;
+      }
+      if (propertyKeyInfo.hasOwnProperty(propertyKey)) {
+        propertyKeyInfo[propertyKey] += 1;
+      } else {
+        propertyKeyInfo[propertyKey] = 1;
+      }
+    })
+  })
+
+  // 取top50的属性
+  const sortKeys = Object.keys(propertyKeyInfo).sort((a,b) => propertyKeyInfo[b] - propertyKeyInfo[a]);
+  return sortKeys.length < n ? sortKeys : sortKeys.slice(0, n);
+}
+  
+const processProperty = (properties, propertyKeys) => propertyKeys.map(key => {
+  if (properties.hasOwnProperty(key)) {
+    // 可以转成数值的直接转成数值
+    if (!isNaN(Number(properties[key]))) {
+      return Number(properties[key]);
+    }
+    // 时间型的转成时间戳
+    if (properties[key].match(secondReg) || properties[key].match(dateReg)) {
+      // @ts-ignore
+      return Number(Date.parse(new Date(properties[key]))) / 1000;
+    }
+  }
+  return 0;
+})
+  
+// 获取属性特征权重
+export const getPropertyWeight = (nodes: NodeConfig[]) => {
+  const propertyKeys = getAllSortProperties(nodes);
+  let allPropertiesWeight = [];
+  for (let i = 0; i < nodes.length; i++) {
+    allPropertiesWeight[i] = processProperty(nodes[i].properties, propertyKeys);
+  }
+  return allPropertiesWeight;
+}
+
+export default {
+  getAllSortProperties,
+  getPropertyWeight,
+}
diff --git a/packages/graph/src/utils/vector.ts b/packages/graph/src/utils/vector.ts
@@ -21,7 +21,7 @@ class Vector {
     }
     if (this.arr.length === otherArr.length) {
       let res = [];
-      for(let key in this.arr) {
+      for (let key in this.arr) {
         res[key] = this.arr[key] + otherArr[key];
       }
       return new Vector(res);
@@ -38,7 +38,7 @@ class Vector {
     }
     if (this.arr.length === otherArr.length) {
       let res = [];
-      for(let key in this.arr) {
+      for (let key in this.arr) {
         res[key] = this.arr[key] - otherArr[key];
       }
       return new Vector(res);
@@ -47,15 +47,15 @@ class Vector {
 
   avg(length) {
     let res = [];
-    for(let key in this.arr) {
+    for (let key in this.arr) {
       res[key] = this.arr[key] / length;
     }
     return new Vector(res);
   }
 
   negate() {
     let res = [];
-    for(let key in this.arr) {
+    for (let key in this.arr) {
       res[key] = - this.arr[key];
     }
     return new Vector(res);
@@ -69,7 +69,7 @@ class Vector {
     }
     if (this.arr.length === otherArr.length) {
       let res = 0;
-      for(let key in this.arr) {
+      for (let key in this.arr) {
         res += Math.pow(this.arr[key] - otherVector.arr[key], 2);
       }
       return res;
@@ -83,11 +83,38 @@ class Vector {
     cloneArr.sort((a, b) => a - b);
     const max = cloneArr[cloneArr.length - 1];
     const min = cloneArr[0];
-    for(let key in this.arr) {
+    for (let key in this.arr) {
       res[key] = (this.arr[key] - min) / (max - min);
     }
     return new Vector(res);
   }
+
+  // 2范数 or 模长
+  norm2() {
+    if (!this.arr?.length) {
+      return 0;
+    }
+    let res = 0;
+      for (let key in this.arr) {
+        res += Math.pow(this.arr[key], 2);
+      }
+    return Math.sqrt(res);
+  }
+
+  // 两个向量的点积
+  dot(otherVector) {
+    const otherArr = otherVector.arr;
+      if (!this.arr?.length || !otherArr?.length) {
+        return 0;
+      }
+      if (this.arr.length === otherArr.length) {
+        let res = 0;
+        for (let key in this.arr) {
+          res += this.arr[key] * otherVector.arr[key];
+        }
+        return res;
+      }
+    }
 }
 
 export default Vector;
diff --git a/packages/graph/tests/unit/cosineSimilarity-spec.ts b/packages/graph/tests/unit/cosineSimilarity-spec.ts
@@ -0,0 +1,41 @@
+import { cosineSimilarity } from '../../src';
+
+describe('cosineSimilarity abnormal demo: ', () => {
+  it('item contains only zeros: ', () => {
+    const item = [0, 0, 0];
+    const targetTtem = [3, 1, 1];
+    const cosineSimilarityValue = cosineSimilarity(item, targetTtem);
+    expect(cosineSimilarityValue).toBe(0);
+  });
+  it('targetTtem contains only zeros: ', () => {
+    const item = [3, 5, 2];
+    const targetTtem = [0, 0, 0];
+    const cosineSimilarityValue = cosineSimilarity(item, targetTtem);
+    expect(cosineSimilarityValue).toBe(0);
+  });
+  it('item and targetTtem both contains only zeros: ', () => {
+    const item = [0, 0, 0];
+    const targetTtem = [0, 0, 0];
+    const cosineSimilarityValue = cosineSimilarity(item, targetTtem);
+    expect(cosineSimilarityValue).toBe(0);
+  });
+});
+
+describe('cosineSimilarity normal demo: ', () => {
+  it('demo similar: ', () => {
+    const item = [30, 0, 100];
+    const targetTtem = [32, 1, 120];
+    const cosineSimilarityValue = cosineSimilarity(item, targetTtem);
+    expect(cosineSimilarityValue).toBeGreaterThanOrEqual(0);
+    expect(cosineSimilarityValue).toBeLessThan(1);
+    expect(Number(cosineSimilarityValue.toFixed(3))).toBe(0.999);
+  });
+  it('demo dissimilar: ', () => {
+    const item = [10, 300, 2];
+    const targetTtem = [1, 2, 30];
+    const cosineSimilarityValue = cosineSimilarity(item, targetTtem);
+    expect(cosineSimilarityValue).toBeGreaterThanOrEqual(0);
+    expect(cosineSimilarityValue).toBeLessThan(1);
+    expect(Number(cosineSimilarityValue.toFixed(3))).toBe(0.074);
+  });
+});
diff --git a/packages/graph/tests/unit/nodesCosineSimilarity-spec.ts b/packages/graph/tests/unit/nodesCosineSimilarity-spec.ts

Original file line number	Diff line number	Diff line change
`@@ -1,6 +1,6 @@`
`1`	`1`	`{`
`2`	`2`	`"name": "@antv/algorithm",`
`3`		`- "version": "0.1.16",`
	`3`	`+ "version": "0.1.17",`
`4`	`4`	`"description": "graph algorithm",`
`5`	`5`	`"keywords": [`
`6`	`6`	`"graph",`