![\"\"](\"https:\/\/tensor.agenthub.uk\/wp-content\/uploads\/2024\/05\/image-10.png\")
\u76f8\u4fe1\u8fd9\u4e2a\u56fe\u4f60\u4e5f\u770b\u5410\u4e86\uff0c\u4e8e\u662f\u6211\u4eec\u8fd8\u662f\u91cd\u65b0\u753b\u4e00\u4e0b\u5427..<\/p>\n\n\n
![\"\"](\"https:\/\/tensor.agenthub.uk\/wp-content\/uploads\/2024\/05\/image-11.png\")
<\/figure>\n<\/div>\n\n\n\u8fd9\u6837\u770b\u8d77\u6765\u6709\u6ca1\u6709\u5f88\u6e05\u6670\uff5e\uff5e\u7b80\u76f4\u6e05\u6670\u5230\u5bb6\u4e86\uff5e\uff5e\uff5e\u975e\u5e38\u7a00\u758f\u7684\u9ad8\u7eac\u7279\u5f81$x$\u8f93\u5165\u5230\u6a21\u578b\uff0cEmbedding\u5c42\u4ed6\u4eec\u5904\u7406\u6210\u7a20\u5bc6\u7684\u4f4e\u7eac\u7279\u5f81$D$\uff0c\u4e4b\u540e\u9001\u5165FM\u90e8\u5206\u548cDeep\u90e8\u5206\uff0c\u8fd9\u6837FM\u90e8\u5206\u8f93\u51fa\u7684\u7ed3\u679c\uff1a\u00a0<\/p>\n\n\n\n
$$out_{FM} = FM(D)$$<\/p>\n\n\n\n
Deep\u90e8\u5206\u7684\u8f93\u51fa\uff1a<\/p>\n\n\n\n
$$out_{Deep} = Deep(D)$$<\/p>\n\n\n\n
\u6a21\u578b\u6700\u7ec8\u7684\u8f93\u51fa<\/p>\n\n\n\n
$$out = out_{FM} + out_{Deep}$$<\/p>\n\n\n\n
\u6574\u4e2a\u6a21\u578b\u7684\u53c2\u6570\uff0c\u5305\u62ecEmbedding\u5c42\u7684weights\u662f\u4e00\u8d77\u66f4\u65b0\uff0c\u5927\u6982\u56e0\u4e3aEmbedding\u5c42\u4e0d\u5149\u5f97\u8003\u8651\u5e95\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u8fd8\u5f97\u8003\u8651\u9ad8\u9636\u7684\uff0c\u6240\u4ee5\u4f1a\u6709\u66f4\u597d\u7684\u7279\u5f81\u8868\u8fbe\u3002<\/p>\n\n\n\n
2.\u8f93\u5165\u5c42+Embedding\u5c42<\/h2>\n\n\n\n![\"\"](\"https:\/\/tensor.agenthub.uk\/wp-content\/uploads\/2024\/05\/image-12-1024x353.png\")
<\/figure>\n<\/div>\n\n\n\u8f93\u5165\u5c42\u662fonehot\u7684\u7279\u5f81\uff0c\u6bcf\u4e2a\u989c\u8272\u8868\u793a\u4e00\u4e2afield\uff0c\u8fd9\u4e2afield\u4e2d\u7684\u6bcf\u7ef4\u6570\u636e\u53ea\u8ddf\u4e0a\u9762Embedding\u5c42\u7684\u5bf9\u5e94field\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff0c\u5168\u94fe\u63a5\u64cd\u4f5c\u3002\u8fd9\u6837\uff0c\u6bcf\u4e2afield\u90fd\u4f1aembedding\u5230\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\uff0c\u8fd9\u4e2a\u5411\u91cf\u7684\u957f\u5ea6\u5728FM\u4e2d\u6807\u8bb0\u4e3a$k$\u3002<\/p>\n\n\n\n
\u7eff\u8272\u7684\u7279\u5f81\u53ea\u8ddf\u7eff\u8272\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff5e\u7eff\u8272\u795e\u7ecf\u5143\u8f93\u51fa\u76844\u4e2a\u503c\u5c31\u662f\u7eff\u8272\u8fd9\u4e2afield\u7684Embedding\u5411\u91cf\u3002<\/p>\n\n\n\n
\u6211\u4eec\u770b\u4e0b\u8fd9\u4e2a\u795e\u7ecf\u5143\u7684\u64cd\u4f5c\uff1a<\/p>\n\n\n
\n![\"\"](\"https:\/\/tensor.agenthub.uk\/wp-content\/uploads\/2024\/05\/image-13-1024x114.png\")
<\/figure>\n<\/div>\n\n\n\u8fd9\u7eff\u8272\uff0c\u8fd8\u771f\u662f\u751f\u673a\u76ce\u7136\u7684\u989c\u8272~ \u7eff\u6cb9\u6cb9\u7684\uff0c\u6df1\u7eff\u8272\u90a3\u4e2a\u662f1\uff0c\u5176\u4ed6\u6d45\u7eff\u8272\u90fd\u662f0\uff0c\u6240\u4ee5\u8fd9\u4e2a\u795e\u7ecf\u5143\u53ea\u6709$w_6$\u7559\u4e0b\u4e86\uff0c\u4e4b\u540e\u5176\u4ed6\u4e09\u4e2a\u7eff\u8272\u795e\u7ecf\u5143\u4e5f\u53ea\u6709\u4ed6\u4eec\u7684$w_6$\u7559\u4e0b\u3002<\/p>\n\n\n\n
\u6309\u7167\u8fd9\u4e2a\u601d\u8def\u5b9e\u9645\u4e0a\uff0c\u6bcf\u4e2afield\u4e2d\u53ea\u6709\u4e00\u4e2a\u4e3a1\u7684\u53c2\u4e0e\u5b9e\u9645\u8fd0\u7b97\uff0c\u5176\u4ed6\u90fd\u662f\u6253\u9171\u6cb9\u7684\uff0c\u4e8e\u662f\u6211\u4eec\u5f88\u591amodel\u7684Input\u5230Embedding\u4ee3\u7801\u5b9e\u73b0\uff0c\u662f\u76f4\u63a5\u4f7f\u7528\u6846\u67b6\u7684Lookup Table\uff0c\u6bd4\u5982[pytorch\u7684Embedding\u5c42]\u3002<\/p>\n\n\n\n
\u5982\u679c\u4f7f\u7528torch.nn.Embedding\u6765\u5b9e\u73b0\uff0c\u8fd8\u8282\u7701\u4e86\u5c06\u7c7b\u522b\u7279\u5f81onehot\u7684\u64cd\u4f5c\uff0c\u6846\u67b6\u76f4\u63a5\u968f\u673a\u751f\u6210$\\text{field_length} \\times k$\u7684\u77e9\u9635\uff0c\u8fd9\u4e2a\u77e9\u9635\u53ef\u4ee5\u53ef\u4ee5\u770b\u4f5c\u662fembedding\u5c42$k$\u4e2a\u7eff\u8272\u7684\u795e\u7ecf\u5143\u6240\u6709weights\u7ec4\u6210\u7684\u77e9\u9635\u3002\u6253\u4e2a\u6bd4\u65b9\uff0c\u57ce\u5e02\u6709600\u7eac\uff0c\u6211\u4eec\u5e0c\u671bEmbedding\u52305\u7eac\u7684\u7a20\u5bc6\u5411\u91cf\uff0c\u76f4\u63a5\u4f1a\u751f\u6210$600 \\times 5$\u7684weights\u77e9\u9635\uff0c\u5982\u679c\u5317\u4eac\u7684\u7f16\u53f7\u662f1\u90a3\u5c31\u76f4\u63a5\u8fd4\u56deweights\u7684\u7b2c\u4e00\u5217\u5411\u91cf\uff0c\u8fd9\u4e2a\u5411\u91cf\u5c31\u662f\u5317\u4eac\u7684\u5411\u91cf\u8868\u8fbe\u3002\u5728\u4e4b\u540e\u7684FM\u5c42\u548cDeep\u5c42\uff0c\u5317\u4eac\u8fd9\u4e2a\u57ce\u5e02\u7279\u5f81\u5c31\u4ee55\u7ef4\u7684\u5411\u91cf\u53c2\u4e0e\u8fd0\u7b97\u3002<\/p>\n\n\n\n
3.FM\u5c42<\/h2>\n\n\n\n![\"\"](\"https:\/\/tensor.agenthub.uk\/wp-content\/uploads\/2024\/05\/image-14.png\")
<\/figure>\n<\/div>\n\n\nFM\u5c42\u7684\u5b9e\u73b0\u4e0a\u7bc7\u6587\u7ae0\u804a\u8fc7\u4e86\uff0c\u7279\u5f81\u90fd\u88abonehot\u8fc7\u4e86\uff0c\u6bcf\u4e2afield\u90fd\u53d8\u6210\u4e86\u4e00\u4e2a$k$\u7eac\u7684\u5411\u91cf\uff0c\u4e0a\u56fe\u4e2d\u5c31\u662f\u6bcf\u4e2afield\u5bf9\u5e94\u7684$k$\u4e2a\u795e\u7ecf\u5143\uff0c\u8fd9$k$\u4e2a\u795e\u7ecf\u5143\u8f93\u51fa\u7684\u5c31\u662f\u5f53\u524d\u7684$V_i$\uff0c\u56e0\u4e3a\u7279\u5f81\u503c\u672c\u8eab\u5c31\u662f1\uff0c\u6240\u4ee5$\\left \\langle V_i, V_j \\right \\rangle x_ix_j$\uff0c\u5c31\u53ea\u6709\u5411\u91cf\u5185\u79ef\u90e8\u5206\u4e86\u3002\u6240\u4ee5\u7279\u5f81\u5230FM\u7684\u8fc7\u7a0b\u53ef\u4ee5\u7b80\u8ff0\u5982\u4e0b\uff1a<\/p>\n\n\n\n
\u4ece\u6bcf\u4e2afield\u5bf9\u5e94\u7684lookup table\u91cc\uff0c\u627e\u5230\u6837\u672c\u4e2d\u6bcf\u4e2afield\u5bf9\u5e94\u7684\u90a3\u4e00\u5217\uff0c\u7ec4\u6210\u4e00\u4e2a\u77e9\u9635\uff0c\u8fd9\u4e2a\u77e9\u9635\u5c31\u662fFM\u4e2d\u7684$V \\in R^{\\text{fl}\\times k}$,\u5176\u4e2d$\\text{fl}$\u662f\u591a\u5c11\u4e2afield\uff0c\u4e4b\u540e\u6267\u884cFM\u7684\u4e8c\u9636\u7ec4\u5408\u90e8\u5206\u7684\u64cd\u4f5c\u3002\u7b80\u5316\u4e4b\u540e\u7684\u64cd\u4f5c\u6d41\u7a0b\u662f\u5bf9\u77e9\u9635$V$\u6309\u884c\u52a0\u548c\u5e73\u65b9\uff0c\u51cf\u53bb\u5bf9\u77e9\u9635$V$\u5e73\u65b9\u52a0\u548c\uff0cFM\u5c42\u7684\u6570\u5b66\u8868\u8fbe:<\/p>\n\n\n\n
$$y_{FM} = \\left \\langle w,x \\right \\rangle + \\sum_{j_1=1}^{d}\\sum_{j_2=j_1+1}^{d} \\left \\langle V_i, V_j \\right \\rangle x_{j1}x_{j2}$$<\/p>\n\n\n\n
4.Deep\u5c42<\/h2>\n\n\n\n![\"\"](\"https:\/\/tensor.agenthub.uk\/wp-content\/uploads\/2024\/05\/image-15.png\")
<\/figure>\n<\/div>\n\n\nDeep\u5c42\u662f\u4e00\u4e2a\u6bd4\u8f83\u7b80\u5355\u7684\u6f5c\u4e8f\u795e\u7ecf\u7f51\u7edc\uff0c\u7528\u6765\u5b66\u4e60\u9ad8\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u4eceembedding\u51fa\u6765\u7684$\\text{field_length} \\times k$\u4e2a\u795e\u7ecf\u5143\u76f4\u63a5\u94fe\u63a5\u5230\u540e\u7eed\u7684Deep\u5c42\u3002\u4f46\u662f\u8fd9$\\text{field_length} \\times k$\u4e2a\u795e\u7ecf\u5143\u5176\u5b9e\u662f\u96b6\u5c5e\u4e8e\u4e0d\u540c\u7684fields\uff0c\u8fdb\u5165\u540e\u7eedDeep Network\u9636\u6bb5\u540e\uff0c\u96b6\u5c5e\u4e8e\u540c\u4e00\u4e2afield\u7684\u591a\u4e2a\u795e\u7ecf\u5143\u5f7c\u6b64\u95f4\u4e5f\u4f1a\u53c2\u4e0e\u7279\u5f81\u7ec4\u5408\uff0c\u76f4\u89c9\u4e0a\u770b\u8fd9\u662f\u5f88\u4e0d\u5408\u7406\u7684\uff0c18\u5e74\u5fae\u8f6f\u63d0\u51fa\u7684xDeepFM\u5bf9\u8fd9\u4e00\u90e8\u5206\u8fdb\u884c\u4e86\u4fee\u6539\u3002<\/p>\n\n\n\n
\u539f\u6587\u8bf4\u8fd9\u91cc\u6709\u4e24\u4e2a\u6bd4\u8f83\u597d\u73a9\u7684\u5730\u65b9\uff1a<\/p>\n\n\n\n
\n- while the lengths of different input field vectors can be different, their embeddings are of the same size (k);<\/li>\n\n\n\n
- the latent feature vectors ($V$) in FM now server as network weights which are learned and used to compress the input field vectors to the embedding vectors.<\/li>\n<\/ol>\n\n\n\n
\u4e0d\u540cfield\u90fd\u88abembedding\u5230\u76f8\u540c\u957f\u5ea6\u7684\u5411\u91cf\u4e2d\uff0c\u8fd9\u6837\u505a\u7684\u597d\u5904\u662f\u6a21\u578b\u7ed3\u6784\u6bd4\u8f83\u7edf\u4e00\uff0c\u5404\u5c42\u94fe\u63a5\u6ca1\u6709\u590d\u6742\u7684\u673a\u6784\uff0c\u4f46\u662f\u4e5f\u9650\u5236\u4e86field\u7684\u8868\u8fbe\u3002FM\u7684\u9690\u7279\u5f81\u5411\u91cf$V$\u662f\u4ece\u7f51\u7edc\u91cc\u76f4\u63a5\u5b66\u5230\u7684\uff0c\u5b83\u4e5f\u8d1f\u8d23\u5c06\u8f93\u5165\u7684\u8d85\u7ea7\u7a00\u758f\u7684\u9ad8\u7eac\u7279\u5f81\u53d8\u6210\u7a20\u5bc6\u7684\u5411\u91cf(embedding)\u3002<\/p>\n\n\n\n
5.\u603b\u7ed3\uff08\u719f\u8bfb\u5e76\u80cc\u8bf5\uff09<\/h2>\n\n\n\n\nDeepFM trains a deep component and an FM component jointly. It gains performance improvement from these advantages:\u00a01) it does not need any pre-training;2) it learns both high- and low-order feature interactions;\u00a03) it introduces a sharing strategy of feature embedding to avoid feature engineering.\u00a0<\/p>\n<\/blockquote>\n\n\n\n
\u6309\u7167\u539f\u8bba\u6587\u7684\u8bf4\u6cd5\uff0cDeepFM\u7684\u4f18\u52bf\u662f\u53ef\u4ee5\u5b9e\u73b0end-to-end\u7684\u8bad\u7ec3\uff0c\u4e0d\u7528\u5bf9\u7279\u5f81\u8fdb\u884c\u989d\u5916\u5904\u7406\u3002\u800c\u4e14\uff0cFM\u90e8\u5206\u548cDeep\u90e8\u5206\u5171\u4eab\u8f93\u5165\uff0c\u53ef\u4ee5\u66f4\u6709\u6548\u7684\u8bad\u7ec3\uff0c\u4f46\u662f\u4e3a\u4ec0\u4e48\u5c31\u6709\u6548\u4e86\u5462\uff1f\u8fd9\u5927\u6982\u8fd8\u662f\u56e0\u4e3aEmbedding\u5c42\u5b66\u4e60\u7684\u597d\uff0c\u5bf9\u7279\u5f81\u7684\u8868\u8fbe\u66f4\u9760\u8c31\u3002<\/p>\n\n\n\n
\u5c31\u56e0\u4e3a\u540c\u65f6\u4ece\u539f\u59cb\u7279\u5f81\u91cc\u9762\u83b7\u53d6\u9ad8\u9636\u4f4e\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u5c31\u66f4\u6709\u6548\u4e86\uff1f\u8fd8\u771f\u662f\uff0c\u5728\u539f\u6587\u5b9e\u9a8c\u90e8\u5206\u7ed9\u51fa\u4e86\u4e0d\u5171\u4eab\u8f93\u5165\u7684\u60c5\u51b5\uff0c\u628awide & deep\u7684LR\u6362\u6210\u4e86FM\uff0c\u4ece\u7ed9\u5b9a\u7684\u6548\u679c\u770b\uff0c\u5171\u4eab\u8f93\u5165\u786e\u5b9e\u8981\u66f4\u597d\uff0c\u76f4\u63a5\u7528FM\u66ff\u6362LR\u7684wide&deep\u6a21\u578b\u6548\u679c\u8fd8\u4e0d\u5982\u4e0d\u66ff\u6362\u7684\u597d\uff0c\u4e0b\u56fe\u4e2dFM&DNN\u5c31\u662f\u521a\u624d\u8bf4\u4e0d\u5171\u4eab\u8f93\u5165\u7684\u60c5\u51b5\u3002<\/p>\n\n\n
\n![\"\"](\"https:\/\/tensor.agenthub.uk\/wp-content\/uploads\/2024\/05\/image-16.png\")
<\/figure>\n<\/div>\n\n\n\u53c2\u8003\u6587\u732e\uff1aHuifeng Guo, Ruiming Tang, et al. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. In Proceedings of the 26th International Joint Conference on Artificial Intelligence. 1725\u20131731.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"\u4e4b\u524d\u8bb0\u5f55\u4e86FM\uff0c\u5927\u6982\u5728\u5ea7\u7684\u5404\u4f4d\u5df2\u7ecf\u60f3\u5230\u540e\u9762\u4f1a\u804aDeepFM\uff0c\u7eb5\u7136FM\u5230DeepFM\u4e2d\u95f4\u9694\u4e86\u4f17\u591a\u7ecf\u5178\u7b97\u6cd5\uff0c\u4e3a\u4e86\u7167\u987e\u76f4\u89c9\u4e0a\u7684\u611f\u53d7\uff0c\u8fd8\u662f\u8bf4\u4e0bDeepFM\u3002 DeepFM\u4e5f\u662f\u4e00\u79cdwide&deep\u5f62\u5f0f\u7684\u6a21\u578b\uff0c\u6a21\u578b\u540c\u65f6\u8003\u8651\u4f4e\u9636\u548c\u9ad8\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0cFM\u90e8\u5206\u8d1f\u8d23\u4f4e\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0cdeep\u90e8\u5206\u8d1f\u8d23\u9ad8\u9636\u7279\u5f81\u7ec4\u5408\u3002 1.\u603b\u7ed3\u6784 \u76f8\u4fe1\u8fd9\u4e2a\u56fe\u4f60\u4e5f\u770b\u5410\u4e86\uff0c\u4e8e\u662f\u6211\u4eec\u8fd8\u662f\u91cd\u65b0\u753b\u4e00\u4e0b\u5427.. \u8fd9\u6837\u770b\u8d77\u6765\u6709\u6ca1\u6709\u5f88\u6e05\u6670\uff5e\uff5e\u7b80\u76f4\u6e05\u6670\u5230\u5bb6\u4e86\uff5e\uff5e\uff5e\u975e\u5e38\u7a00\u758f\u7684\u9ad8\u7eac\u7279\u5f81$x$\u8f93\u5165\u5230\u6a21\u578b\uff0cEmbedding\u5c42\u4ed6\u4eec\u5904\u7406\u6210\u7a20\u5bc6\u7684\u4f4e\u7eac\u7279\u5f81$D$\uff0c\u4e4b\u540e\u9001\u5165FM\u90e8\u5206\u548cDeep\u90e8\u5206\uff0c\u8fd9\u6837FM\u90e8\u5206\u8f93\u51fa\u7684\u7ed3\u679c\uff1a\u00a0 $$out_{FM} = FM(D)$$ Deep\u90e8\u5206\u7684\u8f93\u51fa\uff1a $$out_{Deep} = Deep(D)$$ \u6a21\u578b\u6700\u7ec8\u7684\u8f93\u51fa $$out = out_{FM} + out_{Deep}$$ \u6574\u4e2a\u6a21\u578b\u7684\u53c2\u6570\uff0c\u5305\u62ecEmbedding\u5c42\u7684weights\u662f\u4e00\u8d77\u66f4\u65b0\uff0c\u5927\u6982\u56e0\u4e3aEmbedding\u5c42\u4e0d\u5149\u5f97\u8003\u8651\u5e95\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u8fd8\u5f97\u8003\u8651\u9ad8\u9636\u7684\uff0c\u6240\u4ee5\u4f1a\u6709\u66f4\u597d\u7684\u7279\u5f81\u8868\u8fbe\u3002 2.\u8f93\u5165\u5c42+Embedding\u5c42 \u8f93\u5165\u5c42\u662fonehot\u7684\u7279\u5f81\uff0c\u6bcf\u4e2a\u989c\u8272\u8868\u793a\u4e00\u4e2afield\uff0c\u8fd9\u4e2afield\u4e2d\u7684\u6bcf\u7ef4\u6570\u636e\u53ea\u8ddf\u4e0a\u9762Embedding\u5c42\u7684\u5bf9\u5e94field\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff0c\u5168\u94fe\u63a5\u64cd\u4f5c\u3002\u8fd9\u6837\uff0c\u6bcf\u4e2afield\u90fd\u4f1aembedding\u5230\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\uff0c\u8fd9\u4e2a\u5411\u91cf\u7684\u957f\u5ea6\u5728FM\u4e2d\u6807\u8bb0\u4e3a$k$\u3002 \u7eff\u8272\u7684\u7279\u5f81\u53ea\u8ddf\u7eff\u8272\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff5e\u7eff\u8272\u795e\u7ecf\u5143\u8f93\u51fa\u76844\u4e2a\u503c\u5c31\u662f\u7eff\u8272\u8fd9\u4e2afield\u7684Embedding\u5411\u91cf\u3002 \u6211\u4eec\u770b\u4e0b\u8fd9\u4e2a\u795e\u7ecf\u5143\u7684\u64cd\u4f5c\uff1a \u8fd9\u7eff\u8272\uff0c\u8fd8\u771f\u662f\u751f\u673a\u76ce\u7136\u7684\u989c\u8272~ \u7eff\u6cb9\u6cb9\u7684\uff0c\u6df1\u7eff\u8272\u90a3\u4e2a\u662f1\uff0c\u5176\u4ed6\u6d45\u7eff\u8272\u90fd\u662f0\uff0c\u6240\u4ee5\u8fd9\u4e2a\u795e\u7ecf\u5143\u53ea\u6709$w_6$\u7559\u4e0b\u4e86\uff0c\u4e4b\u540e\u5176\u4ed6\u4e09\u4e2a\u7eff\u8272\u795e\u7ecf\u5143\u4e5f\u53ea\u6709\u4ed6\u4eec\u7684$w_6$\u7559\u4e0b\u3002 \u6309\u7167\u8fd9\u4e2a\u601d\u8def\u5b9e\u9645\u4e0a\uff0c\u6bcf\u4e2afield\u4e2d\u53ea\u6709\u4e00\u4e2a\u4e3a1\u7684\u53c2\u4e0e\u5b9e\u9645\u8fd0\u7b97\uff0c\u5176\u4ed6\u90fd\u662f\u6253\u9171\u6cb9\u7684\uff0c\u4e8e\u662f\u6211\u4eec\u5f88\u591amodel\u7684Input\u5230Embedding\u4ee3\u7801\u5b9e\u73b0\uff0c\u662f\u76f4\u63a5\u4f7f\u7528\u6846\u67b6\u7684Lookup Table\uff0c\u6bd4\u5982[pytorch\u7684Embedding\u5c42]\u3002 \u5982\u679c\u4f7f\u7528torch.nn.Embedding\u6765\u5b9e\u73b0\uff0c\u8fd8\u8282\u7701\u4e86\u5c06\u7c7b\u522b\u7279\u5f81onehot\u7684\u64cd\u4f5c\uff0c\u6846\u67b6\u76f4\u63a5\u968f\u673a\u751f\u6210$\\text{field_length} […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[29],"class_list":["post-629","post","type-post","status-publish","format-standard","hentry","category-machine-learning","tag-29"],"_links":{"self":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=629"}],"version-history":[{"count":6,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629\/revisions"}],"predecessor-version":[{"id":642,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629\/revisions\/642"}],"wp:attachment":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=629"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=629"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=629"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
<\/figure>\n<\/div>\n\n\n\u8f93\u5165\u5c42\u662fonehot\u7684\u7279\u5f81\uff0c\u6bcf\u4e2a\u989c\u8272\u8868\u793a\u4e00\u4e2afield\uff0c\u8fd9\u4e2afield\u4e2d\u7684\u6bcf\u7ef4\u6570\u636e\u53ea\u8ddf\u4e0a\u9762Embedding\u5c42\u7684\u5bf9\u5e94field\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff0c\u5168\u94fe\u63a5\u64cd\u4f5c\u3002\u8fd9\u6837\uff0c\u6bcf\u4e2afield\u90fd\u4f1aembedding\u5230\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\uff0c\u8fd9\u4e2a\u5411\u91cf\u7684\u957f\u5ea6\u5728FM\u4e2d\u6807\u8bb0\u4e3a$k$\u3002<\/p>\n\n\n\n
\u7eff\u8272\u7684\u7279\u5f81\u53ea\u8ddf\u7eff\u8272\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff5e\u7eff\u8272\u795e\u7ecf\u5143\u8f93\u51fa\u76844\u4e2a\u503c\u5c31\u662f\u7eff\u8272\u8fd9\u4e2afield\u7684Embedding\u5411\u91cf\u3002<\/p>\n\n\n\n
\u6211\u4eec\u770b\u4e0b\u8fd9\u4e2a\u795e\u7ecf\u5143\u7684\u64cd\u4f5c\uff1a<\/p>\n\n\n
<\/figure>\n<\/div>\n\n\n\u8fd9\u7eff\u8272\uff0c\u8fd8\u771f\u662f\u751f\u673a\u76ce\u7136\u7684\u989c\u8272~ \u7eff\u6cb9\u6cb9\u7684\uff0c\u6df1\u7eff\u8272\u90a3\u4e2a\u662f1\uff0c\u5176\u4ed6\u6d45\u7eff\u8272\u90fd\u662f0\uff0c\u6240\u4ee5\u8fd9\u4e2a\u795e\u7ecf\u5143\u53ea\u6709$w_6$\u7559\u4e0b\u4e86\uff0c\u4e4b\u540e\u5176\u4ed6\u4e09\u4e2a\u7eff\u8272\u795e\u7ecf\u5143\u4e5f\u53ea\u6709\u4ed6\u4eec\u7684$w_6$\u7559\u4e0b\u3002<\/p>\n\n\n\n
\u6309\u7167\u8fd9\u4e2a\u601d\u8def\u5b9e\u9645\u4e0a\uff0c\u6bcf\u4e2afield\u4e2d\u53ea\u6709\u4e00\u4e2a\u4e3a1\u7684\u53c2\u4e0e\u5b9e\u9645\u8fd0\u7b97\uff0c\u5176\u4ed6\u90fd\u662f\u6253\u9171\u6cb9\u7684\uff0c\u4e8e\u662f\u6211\u4eec\u5f88\u591amodel\u7684Input\u5230Embedding\u4ee3\u7801\u5b9e\u73b0\uff0c\u662f\u76f4\u63a5\u4f7f\u7528\u6846\u67b6\u7684Lookup Table\uff0c\u6bd4\u5982[pytorch\u7684Embedding\u5c42]\u3002<\/p>\n\n\n\n
\u5982\u679c\u4f7f\u7528torch.nn.Embedding\u6765\u5b9e\u73b0\uff0c\u8fd8\u8282\u7701\u4e86\u5c06\u7c7b\u522b\u7279\u5f81onehot\u7684\u64cd\u4f5c\uff0c\u6846\u67b6\u76f4\u63a5\u968f\u673a\u751f\u6210$\\text{field_length} \\times k$\u7684\u77e9\u9635\uff0c\u8fd9\u4e2a\u77e9\u9635\u53ef\u4ee5\u53ef\u4ee5\u770b\u4f5c\u662fembedding\u5c42$k$\u4e2a\u7eff\u8272\u7684\u795e\u7ecf\u5143\u6240\u6709weights\u7ec4\u6210\u7684\u77e9\u9635\u3002\u6253\u4e2a\u6bd4\u65b9\uff0c\u57ce\u5e02\u6709600\u7eac\uff0c\u6211\u4eec\u5e0c\u671bEmbedding\u52305\u7eac\u7684\u7a20\u5bc6\u5411\u91cf\uff0c\u76f4\u63a5\u4f1a\u751f\u6210$600 \\times 5$\u7684weights\u77e9\u9635\uff0c\u5982\u679c\u5317\u4eac\u7684\u7f16\u53f7\u662f1\u90a3\u5c31\u76f4\u63a5\u8fd4\u56deweights\u7684\u7b2c\u4e00\u5217\u5411\u91cf\uff0c\u8fd9\u4e2a\u5411\u91cf\u5c31\u662f\u5317\u4eac\u7684\u5411\u91cf\u8868\u8fbe\u3002\u5728\u4e4b\u540e\u7684FM\u5c42\u548cDeep\u5c42\uff0c\u5317\u4eac\u8fd9\u4e2a\u57ce\u5e02\u7279\u5f81\u5c31\u4ee55\u7ef4\u7684\u5411\u91cf\u53c2\u4e0e\u8fd0\u7b97\u3002<\/p>\n\n\n\n
3.FM\u5c42<\/h2>\n\n\n\n<\/figure>\n<\/div>\n\n\nFM\u5c42\u7684\u5b9e\u73b0\u4e0a\u7bc7\u6587\u7ae0\u804a\u8fc7\u4e86\uff0c\u7279\u5f81\u90fd\u88abonehot\u8fc7\u4e86\uff0c\u6bcf\u4e2afield\u90fd\u53d8\u6210\u4e86\u4e00\u4e2a$k$\u7eac\u7684\u5411\u91cf\uff0c\u4e0a\u56fe\u4e2d\u5c31\u662f\u6bcf\u4e2afield\u5bf9\u5e94\u7684$k$\u4e2a\u795e\u7ecf\u5143\uff0c\u8fd9$k$\u4e2a\u795e\u7ecf\u5143\u8f93\u51fa\u7684\u5c31\u662f\u5f53\u524d\u7684$V_i$\uff0c\u56e0\u4e3a\u7279\u5f81\u503c\u672c\u8eab\u5c31\u662f1\uff0c\u6240\u4ee5$\\left \\langle V_i, V_j \\right \\rangle x_ix_j$\uff0c\u5c31\u53ea\u6709\u5411\u91cf\u5185\u79ef\u90e8\u5206\u4e86\u3002\u6240\u4ee5\u7279\u5f81\u5230FM\u7684\u8fc7\u7a0b\u53ef\u4ee5\u7b80\u8ff0\u5982\u4e0b\uff1a<\/p>\n\n\n\n
\u4ece\u6bcf\u4e2afield\u5bf9\u5e94\u7684lookup table\u91cc\uff0c\u627e\u5230\u6837\u672c\u4e2d\u6bcf\u4e2afield\u5bf9\u5e94\u7684\u90a3\u4e00\u5217\uff0c\u7ec4\u6210\u4e00\u4e2a\u77e9\u9635\uff0c\u8fd9\u4e2a\u77e9\u9635\u5c31\u662fFM\u4e2d\u7684$V \\in R^{\\text{fl}\\times k}$,\u5176\u4e2d$\\text{fl}$\u662f\u591a\u5c11\u4e2afield\uff0c\u4e4b\u540e\u6267\u884cFM\u7684\u4e8c\u9636\u7ec4\u5408\u90e8\u5206\u7684\u64cd\u4f5c\u3002\u7b80\u5316\u4e4b\u540e\u7684\u64cd\u4f5c\u6d41\u7a0b\u662f\u5bf9\u77e9\u9635$V$\u6309\u884c\u52a0\u548c\u5e73\u65b9\uff0c\u51cf\u53bb\u5bf9\u77e9\u9635$V$\u5e73\u65b9\u52a0\u548c\uff0cFM\u5c42\u7684\u6570\u5b66\u8868\u8fbe:<\/p>\n\n\n\n
$$y_{FM} = \\left \\langle w,x \\right \\rangle + \\sum_{j_1=1}^{d}\\sum_{j_2=j_1+1}^{d} \\left \\langle V_i, V_j \\right \\rangle x_{j1}x_{j2}$$<\/p>\n\n\n\n
4.Deep\u5c42<\/h2>\n\n\n\n<\/figure>\n<\/div>\n\n\nDeep\u5c42\u662f\u4e00\u4e2a\u6bd4\u8f83\u7b80\u5355\u7684\u6f5c\u4e8f\u795e\u7ecf\u7f51\u7edc\uff0c\u7528\u6765\u5b66\u4e60\u9ad8\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u4eceembedding\u51fa\u6765\u7684$\\text{field_length} \\times k$\u4e2a\u795e\u7ecf\u5143\u76f4\u63a5\u94fe\u63a5\u5230\u540e\u7eed\u7684Deep\u5c42\u3002\u4f46\u662f\u8fd9$\\text{field_length} \\times k$\u4e2a\u795e\u7ecf\u5143\u5176\u5b9e\u662f\u96b6\u5c5e\u4e8e\u4e0d\u540c\u7684fields\uff0c\u8fdb\u5165\u540e\u7eedDeep Network\u9636\u6bb5\u540e\uff0c\u96b6\u5c5e\u4e8e\u540c\u4e00\u4e2afield\u7684\u591a\u4e2a\u795e\u7ecf\u5143\u5f7c\u6b64\u95f4\u4e5f\u4f1a\u53c2\u4e0e\u7279\u5f81\u7ec4\u5408\uff0c\u76f4\u89c9\u4e0a\u770b\u8fd9\u662f\u5f88\u4e0d\u5408\u7406\u7684\uff0c18\u5e74\u5fae\u8f6f\u63d0\u51fa\u7684xDeepFM\u5bf9\u8fd9\u4e00\u90e8\u5206\u8fdb\u884c\u4e86\u4fee\u6539\u3002<\/p>\n\n\n\n
\u539f\u6587\u8bf4\u8fd9\u91cc\u6709\u4e24\u4e2a\u6bd4\u8f83\u597d\u73a9\u7684\u5730\u65b9\uff1a<\/p>\n\n\n\n
\n- while the lengths of different input field vectors can be different, their embeddings are of the same size (k);<\/li>\n\n\n\n
- the latent feature vectors ($V$) in FM now server as network weights which are learned and used to compress the input field vectors to the embedding vectors.<\/li>\n<\/ol>\n\n\n\n
\u4e0d\u540cfield\u90fd\u88abembedding\u5230\u76f8\u540c\u957f\u5ea6\u7684\u5411\u91cf\u4e2d\uff0c\u8fd9\u6837\u505a\u7684\u597d\u5904\u662f\u6a21\u578b\u7ed3\u6784\u6bd4\u8f83\u7edf\u4e00\uff0c\u5404\u5c42\u94fe\u63a5\u6ca1\u6709\u590d\u6742\u7684\u673a\u6784\uff0c\u4f46\u662f\u4e5f\u9650\u5236\u4e86field\u7684\u8868\u8fbe\u3002FM\u7684\u9690\u7279\u5f81\u5411\u91cf$V$\u662f\u4ece\u7f51\u7edc\u91cc\u76f4\u63a5\u5b66\u5230\u7684\uff0c\u5b83\u4e5f\u8d1f\u8d23\u5c06\u8f93\u5165\u7684\u8d85\u7ea7\u7a00\u758f\u7684\u9ad8\u7eac\u7279\u5f81\u53d8\u6210\u7a20\u5bc6\u7684\u5411\u91cf(embedding)\u3002<\/p>\n\n\n\n
5.\u603b\u7ed3\uff08\u719f\u8bfb\u5e76\u80cc\u8bf5\uff09<\/h2>\n\n\n\n\nDeepFM trains a deep component and an FM component jointly. It gains performance improvement from these advantages:\u00a01) it does not need any pre-training;2) it learns both high- and low-order feature interactions;\u00a03) it introduces a sharing strategy of feature embedding to avoid feature engineering.\u00a0<\/p>\n<\/blockquote>\n\n\n\n
\u6309\u7167\u539f\u8bba\u6587\u7684\u8bf4\u6cd5\uff0cDeepFM\u7684\u4f18\u52bf\u662f\u53ef\u4ee5\u5b9e\u73b0end-to-end\u7684\u8bad\u7ec3\uff0c\u4e0d\u7528\u5bf9\u7279\u5f81\u8fdb\u884c\u989d\u5916\u5904\u7406\u3002\u800c\u4e14\uff0cFM\u90e8\u5206\u548cDeep\u90e8\u5206\u5171\u4eab\u8f93\u5165\uff0c\u53ef\u4ee5\u66f4\u6709\u6548\u7684\u8bad\u7ec3\uff0c\u4f46\u662f\u4e3a\u4ec0\u4e48\u5c31\u6709\u6548\u4e86\u5462\uff1f\u8fd9\u5927\u6982\u8fd8\u662f\u56e0\u4e3aEmbedding\u5c42\u5b66\u4e60\u7684\u597d\uff0c\u5bf9\u7279\u5f81\u7684\u8868\u8fbe\u66f4\u9760\u8c31\u3002<\/p>\n\n\n\n
\u5c31\u56e0\u4e3a\u540c\u65f6\u4ece\u539f\u59cb\u7279\u5f81\u91cc\u9762\u83b7\u53d6\u9ad8\u9636\u4f4e\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u5c31\u66f4\u6709\u6548\u4e86\uff1f\u8fd8\u771f\u662f\uff0c\u5728\u539f\u6587\u5b9e\u9a8c\u90e8\u5206\u7ed9\u51fa\u4e86\u4e0d\u5171\u4eab\u8f93\u5165\u7684\u60c5\u51b5\uff0c\u628awide & deep\u7684LR\u6362\u6210\u4e86FM\uff0c\u4ece\u7ed9\u5b9a\u7684\u6548\u679c\u770b\uff0c\u5171\u4eab\u8f93\u5165\u786e\u5b9e\u8981\u66f4\u597d\uff0c\u76f4\u63a5\u7528FM\u66ff\u6362LR\u7684wide&deep\u6a21\u578b\u6548\u679c\u8fd8\u4e0d\u5982\u4e0d\u66ff\u6362\u7684\u597d\uff0c\u4e0b\u56fe\u4e2dFM&DNN\u5c31\u662f\u521a\u624d\u8bf4\u4e0d\u5171\u4eab\u8f93\u5165\u7684\u60c5\u51b5\u3002<\/p>\n\n\n
\n<\/figure>\n<\/div>\n\n\n\u53c2\u8003\u6587\u732e\uff1aHuifeng Guo, Ruiming Tang, et al. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. In Proceedings of the 26th International Joint Conference on Artificial Intelligence. 1725\u20131731.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"\u4e4b\u524d\u8bb0\u5f55\u4e86FM\uff0c\u5927\u6982\u5728\u5ea7\u7684\u5404\u4f4d\u5df2\u7ecf\u60f3\u5230\u540e\u9762\u4f1a\u804aDeepFM\uff0c\u7eb5\u7136FM\u5230DeepFM\u4e2d\u95f4\u9694\u4e86\u4f17\u591a\u7ecf\u5178\u7b97\u6cd5\uff0c\u4e3a\u4e86\u7167\u987e\u76f4\u89c9\u4e0a\u7684\u611f\u53d7\uff0c\u8fd8\u662f\u8bf4\u4e0bDeepFM\u3002 DeepFM\u4e5f\u662f\u4e00\u79cdwide&deep\u5f62\u5f0f\u7684\u6a21\u578b\uff0c\u6a21\u578b\u540c\u65f6\u8003\u8651\u4f4e\u9636\u548c\u9ad8\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0cFM\u90e8\u5206\u8d1f\u8d23\u4f4e\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0cdeep\u90e8\u5206\u8d1f\u8d23\u9ad8\u9636\u7279\u5f81\u7ec4\u5408\u3002 1.\u603b\u7ed3\u6784 \u76f8\u4fe1\u8fd9\u4e2a\u56fe\u4f60\u4e5f\u770b\u5410\u4e86\uff0c\u4e8e\u662f\u6211\u4eec\u8fd8\u662f\u91cd\u65b0\u753b\u4e00\u4e0b\u5427.. \u8fd9\u6837\u770b\u8d77\u6765\u6709\u6ca1\u6709\u5f88\u6e05\u6670\uff5e\uff5e\u7b80\u76f4\u6e05\u6670\u5230\u5bb6\u4e86\uff5e\uff5e\uff5e\u975e\u5e38\u7a00\u758f\u7684\u9ad8\u7eac\u7279\u5f81$x$\u8f93\u5165\u5230\u6a21\u578b\uff0cEmbedding\u5c42\u4ed6\u4eec\u5904\u7406\u6210\u7a20\u5bc6\u7684\u4f4e\u7eac\u7279\u5f81$D$\uff0c\u4e4b\u540e\u9001\u5165FM\u90e8\u5206\u548cDeep\u90e8\u5206\uff0c\u8fd9\u6837FM\u90e8\u5206\u8f93\u51fa\u7684\u7ed3\u679c\uff1a\u00a0 $$out_{FM} = FM(D)$$ Deep\u90e8\u5206\u7684\u8f93\u51fa\uff1a $$out_{Deep} = Deep(D)$$ \u6a21\u578b\u6700\u7ec8\u7684\u8f93\u51fa $$out = out_{FM} + out_{Deep}$$ \u6574\u4e2a\u6a21\u578b\u7684\u53c2\u6570\uff0c\u5305\u62ecEmbedding\u5c42\u7684weights\u662f\u4e00\u8d77\u66f4\u65b0\uff0c\u5927\u6982\u56e0\u4e3aEmbedding\u5c42\u4e0d\u5149\u5f97\u8003\u8651\u5e95\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u8fd8\u5f97\u8003\u8651\u9ad8\u9636\u7684\uff0c\u6240\u4ee5\u4f1a\u6709\u66f4\u597d\u7684\u7279\u5f81\u8868\u8fbe\u3002 2.\u8f93\u5165\u5c42+Embedding\u5c42 \u8f93\u5165\u5c42\u662fonehot\u7684\u7279\u5f81\uff0c\u6bcf\u4e2a\u989c\u8272\u8868\u793a\u4e00\u4e2afield\uff0c\u8fd9\u4e2afield\u4e2d\u7684\u6bcf\u7ef4\u6570\u636e\u53ea\u8ddf\u4e0a\u9762Embedding\u5c42\u7684\u5bf9\u5e94field\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff0c\u5168\u94fe\u63a5\u64cd\u4f5c\u3002\u8fd9\u6837\uff0c\u6bcf\u4e2afield\u90fd\u4f1aembedding\u5230\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\uff0c\u8fd9\u4e2a\u5411\u91cf\u7684\u957f\u5ea6\u5728FM\u4e2d\u6807\u8bb0\u4e3a$k$\u3002 \u7eff\u8272\u7684\u7279\u5f81\u53ea\u8ddf\u7eff\u8272\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff5e\u7eff\u8272\u795e\u7ecf\u5143\u8f93\u51fa\u76844\u4e2a\u503c\u5c31\u662f\u7eff\u8272\u8fd9\u4e2afield\u7684Embedding\u5411\u91cf\u3002 \u6211\u4eec\u770b\u4e0b\u8fd9\u4e2a\u795e\u7ecf\u5143\u7684\u64cd\u4f5c\uff1a \u8fd9\u7eff\u8272\uff0c\u8fd8\u771f\u662f\u751f\u673a\u76ce\u7136\u7684\u989c\u8272~ \u7eff\u6cb9\u6cb9\u7684\uff0c\u6df1\u7eff\u8272\u90a3\u4e2a\u662f1\uff0c\u5176\u4ed6\u6d45\u7eff\u8272\u90fd\u662f0\uff0c\u6240\u4ee5\u8fd9\u4e2a\u795e\u7ecf\u5143\u53ea\u6709$w_6$\u7559\u4e0b\u4e86\uff0c\u4e4b\u540e\u5176\u4ed6\u4e09\u4e2a\u7eff\u8272\u795e\u7ecf\u5143\u4e5f\u53ea\u6709\u4ed6\u4eec\u7684$w_6$\u7559\u4e0b\u3002 \u6309\u7167\u8fd9\u4e2a\u601d\u8def\u5b9e\u9645\u4e0a\uff0c\u6bcf\u4e2afield\u4e2d\u53ea\u6709\u4e00\u4e2a\u4e3a1\u7684\u53c2\u4e0e\u5b9e\u9645\u8fd0\u7b97\uff0c\u5176\u4ed6\u90fd\u662f\u6253\u9171\u6cb9\u7684\uff0c\u4e8e\u662f\u6211\u4eec\u5f88\u591amodel\u7684Input\u5230Embedding\u4ee3\u7801\u5b9e\u73b0\uff0c\u662f\u76f4\u63a5\u4f7f\u7528\u6846\u67b6\u7684Lookup Table\uff0c\u6bd4\u5982[pytorch\u7684Embedding\u5c42]\u3002 \u5982\u679c\u4f7f\u7528torch.nn.Embedding\u6765\u5b9e\u73b0\uff0c\u8fd8\u8282\u7701\u4e86\u5c06\u7c7b\u522b\u7279\u5f81onehot\u7684\u64cd\u4f5c\uff0c\u6846\u67b6\u76f4\u63a5\u968f\u673a\u751f\u6210$\\text{field_length} […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[29],"class_list":["post-629","post","type-post","status-publish","format-standard","hentry","category-machine-learning","tag-29"],"_links":{"self":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=629"}],"version-history":[{"count":6,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629\/revisions"}],"predecessor-version":[{"id":642,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629\/revisions\/642"}],"wp:attachment":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=629"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=629"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=629"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
<\/figure>\n<\/div>\n\n\nFM\u5c42\u7684\u5b9e\u73b0\u4e0a\u7bc7\u6587\u7ae0\u804a\u8fc7\u4e86\uff0c\u7279\u5f81\u90fd\u88abonehot\u8fc7\u4e86\uff0c\u6bcf\u4e2afield\u90fd\u53d8\u6210\u4e86\u4e00\u4e2a$k$\u7eac\u7684\u5411\u91cf\uff0c\u4e0a\u56fe\u4e2d\u5c31\u662f\u6bcf\u4e2afield\u5bf9\u5e94\u7684$k$\u4e2a\u795e\u7ecf\u5143\uff0c\u8fd9$k$\u4e2a\u795e\u7ecf\u5143\u8f93\u51fa\u7684\u5c31\u662f\u5f53\u524d\u7684$V_i$\uff0c\u56e0\u4e3a\u7279\u5f81\u503c\u672c\u8eab\u5c31\u662f1\uff0c\u6240\u4ee5$\\left \\langle V_i, V_j \\right \\rangle x_ix_j$\uff0c\u5c31\u53ea\u6709\u5411\u91cf\u5185\u79ef\u90e8\u5206\u4e86\u3002\u6240\u4ee5\u7279\u5f81\u5230FM\u7684\u8fc7\u7a0b\u53ef\u4ee5\u7b80\u8ff0\u5982\u4e0b\uff1a<\/p>\n\n\n\n
\u4ece\u6bcf\u4e2afield\u5bf9\u5e94\u7684lookup table\u91cc\uff0c\u627e\u5230\u6837\u672c\u4e2d\u6bcf\u4e2afield\u5bf9\u5e94\u7684\u90a3\u4e00\u5217\uff0c\u7ec4\u6210\u4e00\u4e2a\u77e9\u9635\uff0c\u8fd9\u4e2a\u77e9\u9635\u5c31\u662fFM\u4e2d\u7684$V \\in R^{\\text{fl}\\times k}$,\u5176\u4e2d$\\text{fl}$\u662f\u591a\u5c11\u4e2afield\uff0c\u4e4b\u540e\u6267\u884cFM\u7684\u4e8c\u9636\u7ec4\u5408\u90e8\u5206\u7684\u64cd\u4f5c\u3002\u7b80\u5316\u4e4b\u540e\u7684\u64cd\u4f5c\u6d41\u7a0b\u662f\u5bf9\u77e9\u9635$V$\u6309\u884c\u52a0\u548c\u5e73\u65b9\uff0c\u51cf\u53bb\u5bf9\u77e9\u9635$V$\u5e73\u65b9\u52a0\u548c\uff0cFM\u5c42\u7684\u6570\u5b66\u8868\u8fbe:<\/p>\n\n\n\n
$$y_{FM} = \\left \\langle w,x \\right \\rangle + \\sum_{j_1=1}^{d}\\sum_{j_2=j_1+1}^{d} \\left \\langle V_i, V_j \\right \\rangle x_{j1}x_{j2}$$<\/p>\n\n\n\n
4.Deep\u5c42<\/h2>\n\n\n\n<\/figure>\n<\/div>\n\n\nDeep\u5c42\u662f\u4e00\u4e2a\u6bd4\u8f83\u7b80\u5355\u7684\u6f5c\u4e8f\u795e\u7ecf\u7f51\u7edc\uff0c\u7528\u6765\u5b66\u4e60\u9ad8\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u4eceembedding\u51fa\u6765\u7684$\\text{field_length} \\times k$\u4e2a\u795e\u7ecf\u5143\u76f4\u63a5\u94fe\u63a5\u5230\u540e\u7eed\u7684Deep\u5c42\u3002\u4f46\u662f\u8fd9$\\text{field_length} \\times k$\u4e2a\u795e\u7ecf\u5143\u5176\u5b9e\u662f\u96b6\u5c5e\u4e8e\u4e0d\u540c\u7684fields\uff0c\u8fdb\u5165\u540e\u7eedDeep Network\u9636\u6bb5\u540e\uff0c\u96b6\u5c5e\u4e8e\u540c\u4e00\u4e2afield\u7684\u591a\u4e2a\u795e\u7ecf\u5143\u5f7c\u6b64\u95f4\u4e5f\u4f1a\u53c2\u4e0e\u7279\u5f81\u7ec4\u5408\uff0c\u76f4\u89c9\u4e0a\u770b\u8fd9\u662f\u5f88\u4e0d\u5408\u7406\u7684\uff0c18\u5e74\u5fae\u8f6f\u63d0\u51fa\u7684xDeepFM\u5bf9\u8fd9\u4e00\u90e8\u5206\u8fdb\u884c\u4e86\u4fee\u6539\u3002<\/p>\n\n\n\n
\u539f\u6587\u8bf4\u8fd9\u91cc\u6709\u4e24\u4e2a\u6bd4\u8f83\u597d\u73a9\u7684\u5730\u65b9\uff1a<\/p>\n\n\n\n
\n- while the lengths of different input field vectors can be different, their embeddings are of the same size (k);<\/li>\n\n\n\n
- the latent feature vectors ($V$) in FM now server as network weights which are learned and used to compress the input field vectors to the embedding vectors.<\/li>\n<\/ol>\n\n\n\n
\u4e0d\u540cfield\u90fd\u88abembedding\u5230\u76f8\u540c\u957f\u5ea6\u7684\u5411\u91cf\u4e2d\uff0c\u8fd9\u6837\u505a\u7684\u597d\u5904\u662f\u6a21\u578b\u7ed3\u6784\u6bd4\u8f83\u7edf\u4e00\uff0c\u5404\u5c42\u94fe\u63a5\u6ca1\u6709\u590d\u6742\u7684\u673a\u6784\uff0c\u4f46\u662f\u4e5f\u9650\u5236\u4e86field\u7684\u8868\u8fbe\u3002FM\u7684\u9690\u7279\u5f81\u5411\u91cf$V$\u662f\u4ece\u7f51\u7edc\u91cc\u76f4\u63a5\u5b66\u5230\u7684\uff0c\u5b83\u4e5f\u8d1f\u8d23\u5c06\u8f93\u5165\u7684\u8d85\u7ea7\u7a00\u758f\u7684\u9ad8\u7eac\u7279\u5f81\u53d8\u6210\u7a20\u5bc6\u7684\u5411\u91cf(embedding)\u3002<\/p>\n\n\n\n
5.\u603b\u7ed3\uff08\u719f\u8bfb\u5e76\u80cc\u8bf5\uff09<\/h2>\n\n\n\n\nDeepFM trains a deep component and an FM component jointly. It gains performance improvement from these advantages:\u00a01) it does not need any pre-training;2) it learns both high- and low-order feature interactions;\u00a03) it introduces a sharing strategy of feature embedding to avoid feature engineering.\u00a0<\/p>\n<\/blockquote>\n\n\n\n
\u6309\u7167\u539f\u8bba\u6587\u7684\u8bf4\u6cd5\uff0cDeepFM\u7684\u4f18\u52bf\u662f\u53ef\u4ee5\u5b9e\u73b0end-to-end\u7684\u8bad\u7ec3\uff0c\u4e0d\u7528\u5bf9\u7279\u5f81\u8fdb\u884c\u989d\u5916\u5904\u7406\u3002\u800c\u4e14\uff0cFM\u90e8\u5206\u548cDeep\u90e8\u5206\u5171\u4eab\u8f93\u5165\uff0c\u53ef\u4ee5\u66f4\u6709\u6548\u7684\u8bad\u7ec3\uff0c\u4f46\u662f\u4e3a\u4ec0\u4e48\u5c31\u6709\u6548\u4e86\u5462\uff1f\u8fd9\u5927\u6982\u8fd8\u662f\u56e0\u4e3aEmbedding\u5c42\u5b66\u4e60\u7684\u597d\uff0c\u5bf9\u7279\u5f81\u7684\u8868\u8fbe\u66f4\u9760\u8c31\u3002<\/p>\n\n\n\n
\u5c31\u56e0\u4e3a\u540c\u65f6\u4ece\u539f\u59cb\u7279\u5f81\u91cc\u9762\u83b7\u53d6\u9ad8\u9636\u4f4e\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u5c31\u66f4\u6709\u6548\u4e86\uff1f\u8fd8\u771f\u662f\uff0c\u5728\u539f\u6587\u5b9e\u9a8c\u90e8\u5206\u7ed9\u51fa\u4e86\u4e0d\u5171\u4eab\u8f93\u5165\u7684\u60c5\u51b5\uff0c\u628awide & deep\u7684LR\u6362\u6210\u4e86FM\uff0c\u4ece\u7ed9\u5b9a\u7684\u6548\u679c\u770b\uff0c\u5171\u4eab\u8f93\u5165\u786e\u5b9e\u8981\u66f4\u597d\uff0c\u76f4\u63a5\u7528FM\u66ff\u6362LR\u7684wide&deep\u6a21\u578b\u6548\u679c\u8fd8\u4e0d\u5982\u4e0d\u66ff\u6362\u7684\u597d\uff0c\u4e0b\u56fe\u4e2dFM&DNN\u5c31\u662f\u521a\u624d\u8bf4\u4e0d\u5171\u4eab\u8f93\u5165\u7684\u60c5\u51b5\u3002<\/p>\n\n\n
\n<\/figure>\n<\/div>\n\n\n\u53c2\u8003\u6587\u732e\uff1aHuifeng Guo, Ruiming Tang, et al. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. In Proceedings of the 26th International Joint Conference on Artificial Intelligence. 1725\u20131731.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"\u4e4b\u524d\u8bb0\u5f55\u4e86FM\uff0c\u5927\u6982\u5728\u5ea7\u7684\u5404\u4f4d\u5df2\u7ecf\u60f3\u5230\u540e\u9762\u4f1a\u804aDeepFM\uff0c\u7eb5\u7136FM\u5230DeepFM\u4e2d\u95f4\u9694\u4e86\u4f17\u591a\u7ecf\u5178\u7b97\u6cd5\uff0c\u4e3a\u4e86\u7167\u987e\u76f4\u89c9\u4e0a\u7684\u611f\u53d7\uff0c\u8fd8\u662f\u8bf4\u4e0bDeepFM\u3002 DeepFM\u4e5f\u662f\u4e00\u79cdwide&deep\u5f62\u5f0f\u7684\u6a21\u578b\uff0c\u6a21\u578b\u540c\u65f6\u8003\u8651\u4f4e\u9636\u548c\u9ad8\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0cFM\u90e8\u5206\u8d1f\u8d23\u4f4e\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0cdeep\u90e8\u5206\u8d1f\u8d23\u9ad8\u9636\u7279\u5f81\u7ec4\u5408\u3002 1.\u603b\u7ed3\u6784 \u76f8\u4fe1\u8fd9\u4e2a\u56fe\u4f60\u4e5f\u770b\u5410\u4e86\uff0c\u4e8e\u662f\u6211\u4eec\u8fd8\u662f\u91cd\u65b0\u753b\u4e00\u4e0b\u5427.. \u8fd9\u6837\u770b\u8d77\u6765\u6709\u6ca1\u6709\u5f88\u6e05\u6670\uff5e\uff5e\u7b80\u76f4\u6e05\u6670\u5230\u5bb6\u4e86\uff5e\uff5e\uff5e\u975e\u5e38\u7a00\u758f\u7684\u9ad8\u7eac\u7279\u5f81$x$\u8f93\u5165\u5230\u6a21\u578b\uff0cEmbedding\u5c42\u4ed6\u4eec\u5904\u7406\u6210\u7a20\u5bc6\u7684\u4f4e\u7eac\u7279\u5f81$D$\uff0c\u4e4b\u540e\u9001\u5165FM\u90e8\u5206\u548cDeep\u90e8\u5206\uff0c\u8fd9\u6837FM\u90e8\u5206\u8f93\u51fa\u7684\u7ed3\u679c\uff1a\u00a0 $$out_{FM} = FM(D)$$ Deep\u90e8\u5206\u7684\u8f93\u51fa\uff1a $$out_{Deep} = Deep(D)$$ \u6a21\u578b\u6700\u7ec8\u7684\u8f93\u51fa $$out = out_{FM} + out_{Deep}$$ \u6574\u4e2a\u6a21\u578b\u7684\u53c2\u6570\uff0c\u5305\u62ecEmbedding\u5c42\u7684weights\u662f\u4e00\u8d77\u66f4\u65b0\uff0c\u5927\u6982\u56e0\u4e3aEmbedding\u5c42\u4e0d\u5149\u5f97\u8003\u8651\u5e95\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u8fd8\u5f97\u8003\u8651\u9ad8\u9636\u7684\uff0c\u6240\u4ee5\u4f1a\u6709\u66f4\u597d\u7684\u7279\u5f81\u8868\u8fbe\u3002 2.\u8f93\u5165\u5c42+Embedding\u5c42 \u8f93\u5165\u5c42\u662fonehot\u7684\u7279\u5f81\uff0c\u6bcf\u4e2a\u989c\u8272\u8868\u793a\u4e00\u4e2afield\uff0c\u8fd9\u4e2afield\u4e2d\u7684\u6bcf\u7ef4\u6570\u636e\u53ea\u8ddf\u4e0a\u9762Embedding\u5c42\u7684\u5bf9\u5e94field\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff0c\u5168\u94fe\u63a5\u64cd\u4f5c\u3002\u8fd9\u6837\uff0c\u6bcf\u4e2afield\u90fd\u4f1aembedding\u5230\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\uff0c\u8fd9\u4e2a\u5411\u91cf\u7684\u957f\u5ea6\u5728FM\u4e2d\u6807\u8bb0\u4e3a$k$\u3002 \u7eff\u8272\u7684\u7279\u5f81\u53ea\u8ddf\u7eff\u8272\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff5e\u7eff\u8272\u795e\u7ecf\u5143\u8f93\u51fa\u76844\u4e2a\u503c\u5c31\u662f\u7eff\u8272\u8fd9\u4e2afield\u7684Embedding\u5411\u91cf\u3002 \u6211\u4eec\u770b\u4e0b\u8fd9\u4e2a\u795e\u7ecf\u5143\u7684\u64cd\u4f5c\uff1a \u8fd9\u7eff\u8272\uff0c\u8fd8\u771f\u662f\u751f\u673a\u76ce\u7136\u7684\u989c\u8272~ \u7eff\u6cb9\u6cb9\u7684\uff0c\u6df1\u7eff\u8272\u90a3\u4e2a\u662f1\uff0c\u5176\u4ed6\u6d45\u7eff\u8272\u90fd\u662f0\uff0c\u6240\u4ee5\u8fd9\u4e2a\u795e\u7ecf\u5143\u53ea\u6709$w_6$\u7559\u4e0b\u4e86\uff0c\u4e4b\u540e\u5176\u4ed6\u4e09\u4e2a\u7eff\u8272\u795e\u7ecf\u5143\u4e5f\u53ea\u6709\u4ed6\u4eec\u7684$w_6$\u7559\u4e0b\u3002 \u6309\u7167\u8fd9\u4e2a\u601d\u8def\u5b9e\u9645\u4e0a\uff0c\u6bcf\u4e2afield\u4e2d\u53ea\u6709\u4e00\u4e2a\u4e3a1\u7684\u53c2\u4e0e\u5b9e\u9645\u8fd0\u7b97\uff0c\u5176\u4ed6\u90fd\u662f\u6253\u9171\u6cb9\u7684\uff0c\u4e8e\u662f\u6211\u4eec\u5f88\u591amodel\u7684Input\u5230Embedding\u4ee3\u7801\u5b9e\u73b0\uff0c\u662f\u76f4\u63a5\u4f7f\u7528\u6846\u67b6\u7684Lookup Table\uff0c\u6bd4\u5982[pytorch\u7684Embedding\u5c42]\u3002 \u5982\u679c\u4f7f\u7528torch.nn.Embedding\u6765\u5b9e\u73b0\uff0c\u8fd8\u8282\u7701\u4e86\u5c06\u7c7b\u522b\u7279\u5f81onehot\u7684\u64cd\u4f5c\uff0c\u6846\u67b6\u76f4\u63a5\u968f\u673a\u751f\u6210$\\text{field_length} […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[29],"class_list":["post-629","post","type-post","status-publish","format-standard","hentry","category-machine-learning","tag-29"],"_links":{"self":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=629"}],"version-history":[{"count":6,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629\/revisions"}],"predecessor-version":[{"id":642,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629\/revisions\/642"}],"wp:attachment":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=629"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=629"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=629"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
<\/figure>\n<\/div>\n\n\nDeep\u5c42\u662f\u4e00\u4e2a\u6bd4\u8f83\u7b80\u5355\u7684\u6f5c\u4e8f\u795e\u7ecf\u7f51\u7edc\uff0c\u7528\u6765\u5b66\u4e60\u9ad8\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u4eceembedding\u51fa\u6765\u7684$\\text{field_length} \\times k$\u4e2a\u795e\u7ecf\u5143\u76f4\u63a5\u94fe\u63a5\u5230\u540e\u7eed\u7684Deep\u5c42\u3002\u4f46\u662f\u8fd9$\\text{field_length} \\times k$\u4e2a\u795e\u7ecf\u5143\u5176\u5b9e\u662f\u96b6\u5c5e\u4e8e\u4e0d\u540c\u7684fields\uff0c\u8fdb\u5165\u540e\u7eedDeep Network\u9636\u6bb5\u540e\uff0c\u96b6\u5c5e\u4e8e\u540c\u4e00\u4e2afield\u7684\u591a\u4e2a\u795e\u7ecf\u5143\u5f7c\u6b64\u95f4\u4e5f\u4f1a\u53c2\u4e0e\u7279\u5f81\u7ec4\u5408\uff0c\u76f4\u89c9\u4e0a\u770b\u8fd9\u662f\u5f88\u4e0d\u5408\u7406\u7684\uff0c18\u5e74\u5fae\u8f6f\u63d0\u51fa\u7684xDeepFM\u5bf9\u8fd9\u4e00\u90e8\u5206\u8fdb\u884c\u4e86\u4fee\u6539\u3002<\/p>\n\n\n\n
\u539f\u6587\u8bf4\u8fd9\u91cc\u6709\u4e24\u4e2a\u6bd4\u8f83\u597d\u73a9\u7684\u5730\u65b9\uff1a<\/p>\n\n\n\n
- \n
- while the lengths of different input field vectors can be different, their embeddings are of the same size (k);<\/li>\n\n\n\n
- the latent feature vectors ($V$) in FM now server as network weights which are learned and used to compress the input field vectors to the embedding vectors.<\/li>\n<\/ol>\n\n\n\n
\u4e0d\u540cfield\u90fd\u88abembedding\u5230\u76f8\u540c\u957f\u5ea6\u7684\u5411\u91cf\u4e2d\uff0c\u8fd9\u6837\u505a\u7684\u597d\u5904\u662f\u6a21\u578b\u7ed3\u6784\u6bd4\u8f83\u7edf\u4e00\uff0c\u5404\u5c42\u94fe\u63a5\u6ca1\u6709\u590d\u6742\u7684\u673a\u6784\uff0c\u4f46\u662f\u4e5f\u9650\u5236\u4e86field\u7684\u8868\u8fbe\u3002FM\u7684\u9690\u7279\u5f81\u5411\u91cf$V$\u662f\u4ece\u7f51\u7edc\u91cc\u76f4\u63a5\u5b66\u5230\u7684\uff0c\u5b83\u4e5f\u8d1f\u8d23\u5c06\u8f93\u5165\u7684\u8d85\u7ea7\u7a00\u758f\u7684\u9ad8\u7eac\u7279\u5f81\u53d8\u6210\u7a20\u5bc6\u7684\u5411\u91cf(embedding)\u3002<\/p>\n\n\n\n
5.\u603b\u7ed3\uff08\u719f\u8bfb\u5e76\u80cc\u8bf5\uff09<\/h2>\n\n\n\n
\n
DeepFM trains a deep component and an FM component jointly. It gains performance improvement from these advantages:\u00a01) it does not need any pre-training;2) it learns both high- and low-order feature interactions;\u00a03) it introduces a sharing strategy of feature embedding to avoid feature engineering.\u00a0<\/p>\n<\/blockquote>\n\n\n\n
\u6309\u7167\u539f\u8bba\u6587\u7684\u8bf4\u6cd5\uff0cDeepFM\u7684\u4f18\u52bf\u662f\u53ef\u4ee5\u5b9e\u73b0end-to-end\u7684\u8bad\u7ec3\uff0c\u4e0d\u7528\u5bf9\u7279\u5f81\u8fdb\u884c\u989d\u5916\u5904\u7406\u3002\u800c\u4e14\uff0cFM\u90e8\u5206\u548cDeep\u90e8\u5206\u5171\u4eab\u8f93\u5165\uff0c\u53ef\u4ee5\u66f4\u6709\u6548\u7684\u8bad\u7ec3\uff0c\u4f46\u662f\u4e3a\u4ec0\u4e48\u5c31\u6709\u6548\u4e86\u5462\uff1f\u8fd9\u5927\u6982\u8fd8\u662f\u56e0\u4e3aEmbedding\u5c42\u5b66\u4e60\u7684\u597d\uff0c\u5bf9\u7279\u5f81\u7684\u8868\u8fbe\u66f4\u9760\u8c31\u3002<\/p>\n\n\n\n
\u5c31\u56e0\u4e3a\u540c\u65f6\u4ece\u539f\u59cb\u7279\u5f81\u91cc\u9762\u83b7\u53d6\u9ad8\u9636\u4f4e\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u5c31\u66f4\u6709\u6548\u4e86\uff1f\u8fd8\u771f\u662f\uff0c\u5728\u539f\u6587\u5b9e\u9a8c\u90e8\u5206\u7ed9\u51fa\u4e86\u4e0d\u5171\u4eab\u8f93\u5165\u7684\u60c5\u51b5\uff0c\u628awide & deep\u7684LR\u6362\u6210\u4e86FM\uff0c\u4ece\u7ed9\u5b9a\u7684\u6548\u679c\u770b\uff0c\u5171\u4eab\u8f93\u5165\u786e\u5b9e\u8981\u66f4\u597d\uff0c\u76f4\u63a5\u7528FM\u66ff\u6362LR\u7684wide&deep\u6a21\u578b\u6548\u679c\u8fd8\u4e0d\u5982\u4e0d\u66ff\u6362\u7684\u597d\uff0c\u4e0b\u56fe\u4e2dFM&DNN\u5c31\u662f\u521a\u624d\u8bf4\u4e0d\u5171\u4eab\u8f93\u5165\u7684\u60c5\u51b5\u3002<\/p>\n\n\n
\n<\/figure>\n<\/div>\n\n\n\u53c2\u8003\u6587\u732e\uff1aHuifeng Guo, Ruiming Tang, et al. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. In Proceedings of the 26th International Joint Conference on Artificial Intelligence. 1725\u20131731.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"
\u4e4b\u524d\u8bb0\u5f55\u4e86FM\uff0c\u5927\u6982\u5728\u5ea7\u7684\u5404\u4f4d\u5df2\u7ecf\u60f3\u5230\u540e\u9762\u4f1a\u804aDeepFM\uff0c\u7eb5\u7136FM\u5230DeepFM\u4e2d\u95f4\u9694\u4e86\u4f17\u591a\u7ecf\u5178\u7b97\u6cd5\uff0c\u4e3a\u4e86\u7167\u987e\u76f4\u89c9\u4e0a\u7684\u611f\u53d7\uff0c\u8fd8\u662f\u8bf4\u4e0bDeepFM\u3002 DeepFM\u4e5f\u662f\u4e00\u79cdwide&deep\u5f62\u5f0f\u7684\u6a21\u578b\uff0c\u6a21\u578b\u540c\u65f6\u8003\u8651\u4f4e\u9636\u548c\u9ad8\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0cFM\u90e8\u5206\u8d1f\u8d23\u4f4e\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0cdeep\u90e8\u5206\u8d1f\u8d23\u9ad8\u9636\u7279\u5f81\u7ec4\u5408\u3002 1.\u603b\u7ed3\u6784 \u76f8\u4fe1\u8fd9\u4e2a\u56fe\u4f60\u4e5f\u770b\u5410\u4e86\uff0c\u4e8e\u662f\u6211\u4eec\u8fd8\u662f\u91cd\u65b0\u753b\u4e00\u4e0b\u5427.. \u8fd9\u6837\u770b\u8d77\u6765\u6709\u6ca1\u6709\u5f88\u6e05\u6670\uff5e\uff5e\u7b80\u76f4\u6e05\u6670\u5230\u5bb6\u4e86\uff5e\uff5e\uff5e\u975e\u5e38\u7a00\u758f\u7684\u9ad8\u7eac\u7279\u5f81$x$\u8f93\u5165\u5230\u6a21\u578b\uff0cEmbedding\u5c42\u4ed6\u4eec\u5904\u7406\u6210\u7a20\u5bc6\u7684\u4f4e\u7eac\u7279\u5f81$D$\uff0c\u4e4b\u540e\u9001\u5165FM\u90e8\u5206\u548cDeep\u90e8\u5206\uff0c\u8fd9\u6837FM\u90e8\u5206\u8f93\u51fa\u7684\u7ed3\u679c\uff1a\u00a0 $$out_{FM} = FM(D)$$ Deep\u90e8\u5206\u7684\u8f93\u51fa\uff1a $$out_{Deep} = Deep(D)$$ \u6a21\u578b\u6700\u7ec8\u7684\u8f93\u51fa $$out = out_{FM} + out_{Deep}$$ \u6574\u4e2a\u6a21\u578b\u7684\u53c2\u6570\uff0c\u5305\u62ecEmbedding\u5c42\u7684weights\u662f\u4e00\u8d77\u66f4\u65b0\uff0c\u5927\u6982\u56e0\u4e3aEmbedding\u5c42\u4e0d\u5149\u5f97\u8003\u8651\u5e95\u9636\u7684\u7279\u5f81\u7ec4\u5408\uff0c\u8fd8\u5f97\u8003\u8651\u9ad8\u9636\u7684\uff0c\u6240\u4ee5\u4f1a\u6709\u66f4\u597d\u7684\u7279\u5f81\u8868\u8fbe\u3002 2.\u8f93\u5165\u5c42+Embedding\u5c42 \u8f93\u5165\u5c42\u662fonehot\u7684\u7279\u5f81\uff0c\u6bcf\u4e2a\u989c\u8272\u8868\u793a\u4e00\u4e2afield\uff0c\u8fd9\u4e2afield\u4e2d\u7684\u6bcf\u7ef4\u6570\u636e\u53ea\u8ddf\u4e0a\u9762Embedding\u5c42\u7684\u5bf9\u5e94field\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff0c\u5168\u94fe\u63a5\u64cd\u4f5c\u3002\u8fd9\u6837\uff0c\u6bcf\u4e2afield\u90fd\u4f1aembedding\u5230\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\uff0c\u8fd9\u4e2a\u5411\u91cf\u7684\u957f\u5ea6\u5728FM\u4e2d\u6807\u8bb0\u4e3a$k$\u3002 \u7eff\u8272\u7684\u7279\u5f81\u53ea\u8ddf\u7eff\u8272\u7684\u795e\u7ecf\u5143\u94fe\u63a5\uff5e\u7eff\u8272\u795e\u7ecf\u5143\u8f93\u51fa\u76844\u4e2a\u503c\u5c31\u662f\u7eff\u8272\u8fd9\u4e2afield\u7684Embedding\u5411\u91cf\u3002 \u6211\u4eec\u770b\u4e0b\u8fd9\u4e2a\u795e\u7ecf\u5143\u7684\u64cd\u4f5c\uff1a \u8fd9\u7eff\u8272\uff0c\u8fd8\u771f\u662f\u751f\u673a\u76ce\u7136\u7684\u989c\u8272~ \u7eff\u6cb9\u6cb9\u7684\uff0c\u6df1\u7eff\u8272\u90a3\u4e2a\u662f1\uff0c\u5176\u4ed6\u6d45\u7eff\u8272\u90fd\u662f0\uff0c\u6240\u4ee5\u8fd9\u4e2a\u795e\u7ecf\u5143\u53ea\u6709$w_6$\u7559\u4e0b\u4e86\uff0c\u4e4b\u540e\u5176\u4ed6\u4e09\u4e2a\u7eff\u8272\u795e\u7ecf\u5143\u4e5f\u53ea\u6709\u4ed6\u4eec\u7684$w_6$\u7559\u4e0b\u3002 \u6309\u7167\u8fd9\u4e2a\u601d\u8def\u5b9e\u9645\u4e0a\uff0c\u6bcf\u4e2afield\u4e2d\u53ea\u6709\u4e00\u4e2a\u4e3a1\u7684\u53c2\u4e0e\u5b9e\u9645\u8fd0\u7b97\uff0c\u5176\u4ed6\u90fd\u662f\u6253\u9171\u6cb9\u7684\uff0c\u4e8e\u662f\u6211\u4eec\u5f88\u591amodel\u7684Input\u5230Embedding\u4ee3\u7801\u5b9e\u73b0\uff0c\u662f\u76f4\u63a5\u4f7f\u7528\u6846\u67b6\u7684Lookup Table\uff0c\u6bd4\u5982[pytorch\u7684Embedding\u5c42]\u3002 \u5982\u679c\u4f7f\u7528torch.nn.Embedding\u6765\u5b9e\u73b0\uff0c\u8fd8\u8282\u7701\u4e86\u5c06\u7c7b\u522b\u7279\u5f81onehot\u7684\u64cd\u4f5c\uff0c\u6846\u67b6\u76f4\u63a5\u968f\u673a\u751f\u6210$\\text{field_length} […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[29],"class_list":["post-629","post","type-post","status-publish","format-standard","hentry","category-machine-learning","tag-29"],"_links":{"self":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=629"}],"version-history":[{"count":6,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629\/revisions"}],"predecessor-version":[{"id":642,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/629\/revisions\/642"}],"wp:attachment":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=629"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=629"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=629"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}