NNN:面向营销测量的下一代神经网络
机器学习
2025-06-05 v3 应用统计
摘要
我们提出了 NNN,这是一种基于 Transformer 的神经网络方法,用于营销测量。与依赖标量输入和参数衰减函数的营销混合模型(MMM)不同,NNN 使用丰富的嵌入向量捕捉营销和有机渠道(如搜索查询、广告创意)的定性与定量方面。结合其注意力机制,NNN 可能能够建模复杂的相互作用、捕捉长期效应,并提高销售归因的准确性。我们展示,L1 正则化允许在数据受限的典型设置中使用此类具有表达力的模型。通过在模拟数据和真实数据上的评估,证明了 NNN 的有效性,尤其是在预测能力方面取得了显著提升。除了营销测量,NNN 框架还可以通过模型探测提供有价值的、补充性的见解,例如评估关键词或创意的有效性。
引用
@article{arxiv.2504.06212,
title = {NNN: Next-Generation Neural Networks for Marketing Measurement},
author = {Thomas Mulc and Mike Anderson and Paul Cubre and Huikun Zhang and Ivy Liu and Saket Kumar},
journal= {arXiv preprint arXiv:2504.06212},
year = {2025}
}
备注
The title was updated to reflect broader scope. We clarified that our method is not an MMM and emphasized its experimental nature. R2 values in Tables 1 and 2 were corrected after fixing a bug that inflated NNN scores. Notation in the unrolling section now uses 0-indexing. We noted NNN has not been tested with many channels, added an acknowledgment, and improved grammar