科研人员近期发现基于负采样技术的CBOW模型,虽然公式推导没有问题,但无论是word2vec的作者发布的代码或者主流第三方库的复现过程中全部存在梯度计算错误,导致同等条件下CBOW在实际应用中与SkipGram相比效果较差,因此在万物2vector的年代,CBOW被雪藏至今。具体可参考
https://arxiv.org/abs/2012.15332https://arxiv.org/abs/2012.15332https://www.reddit.com/r/MachineLearning/comments/uv6mtz/d_why_is_the_skipgram_model_used_in_deepwalk_and/https://www.reddit.com/r/MachineLearning/comments/uv6mtz/d_why_is_the_skipgram_model_used_in_deepwalk_and/