基于Graph Embedding的行业因子向量化

核心观点

传统方法在处理行业因子信息时通常采用one-hot将行业转变为哑变量进行处理（对于某些特殊模型，可以采用category变量）。这两者均无法表示行业之间的内在关联关系。且哑变量为稀疏矩阵，信息密度较低，增加了计算开销。本文利用行业相关性矩阵构建图结构，采用Graph

Embedding的算法Node2vec将行业因子进行向量化处理，得到的行业因子能够有效的包括行业内在相关性。提纲03讨论02行业Embedding01Graph

Embedding介绍

Embedding 嵌入（Embedding）是指一个数学结构经映射包含到另一个结构中。某个物件X称为嵌入到另一个物件Y中，是指有一个保持结构的单射f:X→Y，这个映射f就给出了一个嵌入万物皆可Embedding（word

embdding、item

embedding、graph

embedding）经典的Word

Embedding算法包括：

Word2vec、

GloVe、FastText等大语言Embedding模型

：

BERT、EloMo、GPT等

One-hot 独热编码（one-hot）：用N位状态寄存器来对N个状态进行编码在NLP中，将单词转换为独热向量缺点：维数灾难语义鸿沟I [1,0,0,0]am [0,1,0,0]a [0,0,1,0]quant [0,0,0,1]

Word2vec 资料来源：中信建投Word2vec: 将单词转换为连续向量基于连续向量，计算单词相似度：model.most_similar(‘贵州茅台’)：('洋河股份',

0.9215401411056519)('酒鬼酒',

0.9041466116905212)('600519',

0.899872362613678)('海天味业',

0.8987340331077576)('口子窖',

0.8912389874458313)('顺鑫农业',

0.8851066827774048)('五粮液',

0.8838584423065186)('老白干酒',

0.8835573196411133)('山西汾酒',

0.8828251361846924)('古井贡酒',

0.8824747204780579)Iamaquant[[0.312334,

-0.128755,0.560548,

0.30743…][0.22087,

0.02268,0.288854,

0.488231…][0.335961,

-0.178007,0.054437,

0.895583…][-0.006239,

0.098319,0.46317,

0.31525…]

Graph

Embedding 图数据结构用来表达物体之间的关联关系图的定义：𝐺=（𝑉,

𝐸）V代表顶点集合，E代表边邻接矩阵：常见的图嵌入算法包括DeepWalk,

LINE,

Node2Vec,

SDNE,

Struct2Vec,

GraphSAGE等图算法能够有效的计算节点之间的关系图嵌入（Graph

Embedding）和图神经网络（Graph

Neural

Network,

GNN）两者互相关联，embedding生成的向量可以作为GNN的输入，同时某些embedding需要用到GNN的算法。

DeepWalk 在图中采用随机游走生成结点序列，利用Word2Vec（SkipGram）算法进行无监督训练，生成embedding结果资料来源：DeepWalk:

online

learning

of

social

representations,

中信建投 资料来源：DeepWalk:

online

learning

of

social

representations,

中信建投优点：算法简单高效缺点：没有考虑边的属性，无法跳出有向图中出度为0的节点Node2VecNode2vec是在DeepWalk的基础上，引入参数p和q，将随机游走序列调节为宽度优先搜索和深度优先搜索的序列生成资料来源：

node2vec:

Scalable

Feature

Learning

forNetworks,中信建投优点：可以学习同质图和同构图结构缺点：超参p，q需要调节𝜋𝑣𝑥=𝛼𝑝𝑞(𝑡,𝑥)·

𝑤𝑣𝑥资料来源：

node2vec:

Scalable

Feature

Learning

for

Networks,中信建投

行业因子

传统方法在处理行业因子信息时通常采用one-hot将行业转变为哑变量进行处理（对于某些特殊模型，可以采用category变量）两者均无法表示行业之间的内在关联关系且哑变量为稀疏矩阵，信息密度较低，增加了计算开销资料来源：中信建投

行业因子Word Embedding 直接采用Word

Embedding提取行业名称的向量能够反映行业之间的关系，但与二级市场的行业关系存在一定的区别，此关系主要来文本资料，映射到二级市场会发生一定的改变单次训练，无法回测Node2Vec资料来源：WIND,中信建投构造有权完全无向图，指数收益率相关性矩阵作为邻接矩阵节点为行业，边为行业之间的相关性使用Node2vec

对行业因子进行embeddingNode2vecwalk_length:10num_walks:1e6window_size:

5embed_size:

30每个行业均可用30维向量表示，利用TSNE将向量压缩到2维可视化展示

讨论

相比于行业哑变量，行业graph

embedding向量信息密度更高，且向量之间内在的包含了行业之间的相关性关系。Embedding向量可以作为各类机器学习和深度学习模型的输入，更好的提高预测效果。资料来源：中信建投

风险提示

本报告中所有数据结果是基于历史统计结果的展示，未来有可能发生风格切换导致因子失效的风险。模型运行存在一定的随机性，初始化随机数种子会对结果产生影响，单次运行结果可能会有一定偏差。历史数据的区间选择会对结果产生一定的影响。模型参数的不同会影响最终结果。模型对计算资源要求较高，运算量不足会导致结果存在一定的欠拟合风险。本文所有模型结果均来自历史数据，模型存在统计误差，不保证模型未来的有效性，对投资不构成任何建议。行业因子Graph

Embedding00 -0.0771 0.0532 -0.0053 -0.1924 0.0565 -0.0076 0.0307 -0.0288 -0.2259 0.11810 -0.03411 -0.05512 -0.00813 -0.00214 -0.04415 0.21016 -0.09817 0.05218 0.10719 -0.08420 0.07621 0.12522 -0.09023 -0.10424 -0.07825 -0.04026 -0.12327 -0.21228 -0.07429 -0.11010.1680.0190.0390.101-0.036-0.0760.028-0.020-0.066-0.0310.198-0.0430.124-0.061-0.116-0.052-0.1830.038-0.040-0.1870.0860.1110.076-0.103-0.1280.001-0.007-0.1140.067-0.1122-0.014-0.082-0.172-0.110-0.0590.043-0.153-0.006-0.1700.0320.1670.035-0.026-0.0910.1610.1310.003-0.0980.0690.0180.1200.1130.1010.143-0.0280.1130.0740.0740.0280.0513-0.227-0.153-0.0770.0380.060-0.023-0.057-0.0470.000-0.1360.101-0.1990.012-0.076-0.0540.0050.0460.1620.059-0.021-0.0140.056-0.033-0.0060.072-0.0220.0060.020-0.0270.2634-0.008-0.068-0.1600.1570.018-0.1590.064-0.062-0.1230.009-0.246-0.0890.0590.0080.0960.0640.0930.0010.001-0.1190.1520.1730.045-0.038-0.063-0.0200.0810.2120.0250.05350.0190.0070.110-0.154-0.0270.2200.0120.0090.0290.019-0.030-0.053-0.034-0.0970.091-0.149-0.094-0.016-0.003-0.2150.1320.214-0.0160.0020.002-0.131-0.004-0.0040.0430.0626-0.111-0.180-0.089-0.041-0.276-0.105-0.093-0.1460.1390.1490.1430.0310.1830.0980.2520.072-0.0340.0970.180-0.003-0.1250.028-0.1780.0160.1840.195-0.0050.0160.0290.1277-0.042-0.069-0.1410.099-0.0110.005-0.054-0.0420.1570.161-0.0210.158-0.070-0.014-0.043-0.104-0.0290.080-0.0420.167-0.0400.093-0.0680.045-0.099-0.0220.254-0.0620.0690.10180.019-0.084-0.0970.015-0.101-0.1580.133-0.081-0.021-0.047-0.208-0.190-0.2280.0910.031-0.0190.055-0.0620.073-0.0110.1380.0050.1340.174-0.341-0.082-0.196-0.119-0.0730.09090.0120.0430.088-0.178-0.061-0.005-0.173-0.1890.131-0.081-0.1130.0290.1010.021-0.1500.0840.030-0.028-0.059-0.0060.0860.0550.0860.064-0.0600.0720.0560.0940.0230.08510-0.094-0.052-0.0200.022-0.1090.094-0.023-0.015-0.116-0.0350.012-0.0200.0910.0290.099-0.003-0.0150.1430.1400.152-0.025-0.0750.079-0.009-0.002-0.0320.087-0.2010.025-0.05411-0.098-0.0090.0700.0770.1650.010-0.1300.0550.0450.0520.0070.0090.012-0.0220.035-0.053-0.0420.021-0.076-0.098-0.083-0.073-0.0630.197-0.0930.049-0.213-0.082-0.012-0.136120.1560.0990.1580.0620.1740.171-0.015-0.204-0.0200.2500.1970.0550.059-0.016-0.061-0.1650.042-0.026-0.0800.044-0.159-0.111-0.164-0.0830.0150.1080.1020.091-0.1940.014130.0360.126-0.1250.0400.0500.0340.074-0.056-0.009-0.171-0.038-0.0110.091-0.0260.1170.105-0.1770.0400.174-0.0440.1720.0460.0610.208-0.044-0.148-0.050-0.058-0.262-0.127140.108-0.0190.1500.088-0.0300.140-0.133-0.021-0.0700.1360.0940.0390.2130.1290.0450.0650.1330.007-0.003-0.026-0.094-0.078-0.0590.1630.104-0.1890.030-0.1030.0580.026150.106-0.058-0.008-0.177-0.069-0.089-0.054-0.030-0.204-0.105-0.0660.044-0.011-0.0150.114-0.1460.0340.183-0.0270.105-0.072-0.038-0.0880.041-0.0350.025-0.0380.054-0.0910.06916-0.0470.1360.1910.1760.0100.085-0.028-0.0700.0280.0250.1040.080-0.088-0.0540.1900.0710.0690.1880.085-0.007-0.049-0.1200.058-0.001-0.070-0.0030.0350.1090.1580.054170.0880.1530.0790.0320.1070.000-0.062-0.0030.040-0.123-0.142-0.304-0.0180.0820.1040.029-0.043-0.0610.0780.1000.1180.0350.0400.000-0.0330.1100.082-0.0330.028-0.046180.0580.117-0.1090.084-0.099-0.0260.048-0.120-0.007-0.027-0.031-0.212-0.043-0.220-0.125-0.0530.159-0.089-0.0630.0930.1070.0510.0380.056-0.060-0.099-0.222-0.090-0.031-0.18519-0.080-0.226-0.132-0.051-0.1460.035-0.091-0.0890.0630.0500.013-0.065-0.1940.2070.0760.129-0.0640.2270.1460.0400.0050.0410.0860.0210.138-0.0390.0330.182-0.047-0.125200.158-0.031-0.112-0.0440.1230.133-0.022-0.0340.172-0.013-0.0030.0210.0740.0070.1620.1330.0740.0790.045-0.0310.042-0.0340.122-0.177-0.101-0.115-0.174-0.1130.0300.02621-0.173-0.1270.037-0.2120.004-0.208-0.1440.0450.0490.0840.014-0.0180.071-0.1140.1680.0220.0360.0910.1840.0260.058-0.0040.057-0.092-0.003-0.172-0.028-0.059-0.129-0.24822-0.092-0.249-0.071-0.011-0.2260.134-0.0310.1120.0160.008-0.0520.0220.027-0.1600.0220.150-0.0110.0660.066-0.073-0.008-0.0360.012-0.0820.0560.1880.108-0.009-0.249-0.060230.095-0.055-0.117-0.022-0.0440.152-0.1150.0250.0070.0240.0070.0130.0660.010-0.031-0.0550.1220.1990.140-0.001-0.074-0.104-0.1590.0230.0500.1190.0280.0520.177-0.067240.0640.045-0.082-0.1090.052-0.0490.1750.0710.0920.150-0.019-0.104-0.004-0.0610.0080.0020.0590.054-0.086-0.181-0.033-0.1320.1300.1410.0770.0820.1780.0160.042-0.007250.0530.056-0.021-0.019-0.068-0.002-0.0230.1960.112-0.0670.073-0.128-0.0420.2040.0620.0750.2040.0640.081-0.0060.0380.000-0.052-0.011-0.199-0.0930.044