加入新推出的
Discord 社区,展开实时讨论,获得同行支持,并直接与 Meridian 团队互动!
从 LightweightMMM 迁移
使用集合让一切井井有条
根据您的偏好保存内容并对其进行分类。
Meridian 是 Google 营销组合建模分析 (MMM) 方法的官方演变版本。它是 LightweightMMM 的更新版本。这两个版本均基于 Google 自 2017 年以来在贝叶斯 MMM 方面的研究。
Meridian 的主要功能包括覆盖面和频次建模、有效处理付费搜索以及实验校准。
如何迁移到 Meridian
若要从 LightweightMMM 迁移到 Meridian,您需要安装 Meridian,并按照任何 Meridian 新用户所用的流程导入数据。如需了解详情,请参阅安装 Meridian。
功能对比
这两种模型的输入数据是相同的。
下图概述了这两个项目之间的主要功能差异:
功能 |
LightweightMMM |
Meridian |
语言 |
Python |
Python |
贝叶斯库 |
Numpyro |
Tensorflow Probability |
实验校准 |
可能有,但需要手动操作 |
有 |
覆盖面和频次建模 |
无 |
有 |
优化器 |
有 |
有 |
模型的投资回报率公式 |
无 |
有 |
纳入 GQV 混杂因素 |
可能有,但需要手动操作 |
有 |
国家级和地理位置级模型 |
有 |
有,国家级以及更多地理位置 |
趋势和季节性 |
直线 + 正弦曲线重复形状(每天、每周) |
结 |
自定义先验 |
有 |
有 |
滞后与饱和度转换 |
有 |
有 |
输入缩放 |
手动 |
自动 |
模型规范的差异
LightweightMMM 提供三种不同的模型架构:Adstock、Hill-Adstock 和 Carryover。Meridian 使用 Hill-Adstock 架构的变体,不支持其他架构。您可以选择对 Meridian 基准模型应用 Hill 转换和 Adstock 转换的顺序。Meridian 覆盖面和频次模型具有固定的 Hill-Adstock 顺序:首先是 Hill,然后是 Adstock。
Meridian 和 LightweightMMM 之间的其他差异包括:
在这两个项目中,媒体渠道都是按地理位置分层的。不过,在 LightweightMMM 中,地理位置层次结构不会添加其他自由形参。然而,LightweightMMM 使用一个媒体系数来指定超先验和各个地理位置级的媒体渠道先验。Meridian 还有一个附加形参 eta_m
,用于指定各地理位置的媒体系数标准差。Meridian 还允许分层变体呈正态或对数正态分布。
非媒体功能(在 Meridian 中称为“控制变量”)在 Meridian 中也是分层的,但在 LightweightMMM 中,它们在各地理位置之间是不分层的。Meridian 模型形参 xi_c
指定了此地理位置层次结构的标准差。
借助 Meridian,您可以根据 Beta 值(与 LightweightMMM 相同)或投资回报率指定媒体先验。
Meridian 中的基准表达方式与 LightweightMMM 有所不同。在 Meridian 中,用户可以同时指定地理位置级和时间级固定效应,而基准是这两种固定效应的总和。
MCMC 抽样时间的预期差异
由于 Meridian 中的模型形参更多,模型复杂度更高,因此 Meridian 中的 MCMC 抽样时间预计要比 LightweightMMM 长。不过,由于模型相对类似,预计 Meridian 所需的时间不会比 LightweightMMM 长很多。若要确切估计需要多长时间,需要考虑计算环境、地理位置数量、模型调优形参、先验、数据和其他因素。虽然 Meridian 的模型复杂度可能会导致 MCMC 抽样时间更长,但预计结果会更准确。
如未另行说明,那么本页面中的内容已根据知识共享署名 4.0 许可获得了许可,并且代码示例已根据 Apache 2.0 许可获得了许可。有关详情,请参阅 Google 开发者网站政策。Java 是 Oracle 和/或其关联公司的注册商标。
最后更新时间 (UTC):2025-08-04。
[null,null,["最后更新时间 (UTC):2025-08-04。"],[[["\u003cp\u003eMeridian is the updated version of Google's LightweightMMM, representing the evolution of their Bayesian MMM research.\u003c/p\u003e\n"],["\u003cp\u003eKey features of Meridian include reach and frequency modeling, effective handling of paid search, and experiment calibration.\u003c/p\u003e\n"],["\u003cp\u003eMigrating to Meridian involves installing it and importing data using the same process as a new user, as detailed in the installation guide.\u003c/p\u003e\n"],["\u003cp\u003eMeridian utilizes Tensorflow Probability as its Bayesian library, while LightweightMMM uses Numpyro, although the input data for both models remains the same.\u003c/p\u003e\n"],["\u003cp\u003eMeridian offers improved features like a ROI formulation, incorporating GQV confounders, and automatic scaling of inputs, along with a more complex model architecture compared to LightweightMMM.\u003c/p\u003e\n"]]],[],null,["# Migrate from LightweightMMM\n\nMeridian is the official evolution of the Google MMM approach. It is the\nupdated version of LightweightMMM. Both versions are based on Google's Bayesian\nMMM research since 2017.\n\nThe key features of Meridian are reach and frequency modeling, handling\npaid search effectively, and experiment calibration.\n\nHow to migrate to Meridian\n--------------------------\n\nTo migrate from LightweightMMM to Meridian, you install Meridian\nand import your data using the same process as any new user to Meridian.\nFor more information, see [Install\nMeridian](/meridian/docs/user-guide/installing).\n\nFeature comparison\n------------------\n\nThe input data for both models is the same.\n\nThe following chart gives an overview of the key feature differences between the\nprojects:\n\n| Feature | LightweightMMM | Meridian |\n|---------------------------------------|------------------------------------------------------------|------------------------------|\n| Language | Python | Python |\n| Bayesian library | Numpyro | Tensorflow Probability |\n| Experiment calibration | Possible but manual | Yes |\n| Reach and frequency modeling | No | Yes |\n| Optimizer | Yes | Yes |\n| ROI formulation of the model | No | Yes |\n| Incorporating GQV confounder | Possible but manual | Yes |\n| National- and geo-level models | Yes | Yes, national plus more geos |\n| Trend and seasonality | Straight line + sinusoidal repeating shape (daily, weekly) | Knots |\n| Custom priors | Yes | Yes |\n| Lagging and saturation transformation | Yes | Yes |\n| Scaling of inputs | Manual | Automatic |\n\nDifferences in the model specifications\n---------------------------------------\n\nLightweightMMM offers three different model architectures: Adstock,\nHill-Adstock, and Carryover. Meridian uses a variation of the\nHill-Adstock architecture, and does not allow other architectures. You can\nchoose the order in which the Hill- and Adstock-transformations are applied for\nthe Meridian baseline model. The Meridian reach and frequency\nmodel has a fixed Hill-Adstock order: Hill first, and then Adstock.\n\nOther differences between Meridian and LightweightMMM include:\n\n- Media channels are hierarchical across geos in both projects. However, in\n LightweightMMM, the geo hierarchy doesn't add additional free parameters.\n Instead, one media coefficient is used to specify both the hyper-prior and\n the individual geo-level media channel priors in LightweightMMM.\n Meridian has an additional parameter `eta_m` that specifies the\n standard deviation of the media coefficient across geos. Meridian\n also allows the hierarchical variation to be either normal or log-normal in\n shape.\n\n- The non-media features, called *control variables* in Meridian, are\n also hierarchical in Meridian, whereas they are non-hierarchical\n across geos in LightweightMMM. The Meridian model parameter `xi_c`\n specifies the standard deviation of this geo hierarchy.\n\n- Meridian lets you specify media priors either in terms of beta (the\n same as LightweightMMM) or in terms of ROI.\n\n- The baseline is expressed differently in Meridian, compared to\n LightweightMMM. With Meridian, users can specify both geo-level and\n time-level fixed effects, and the baseline is the sum of both fixed effects.\n\nExpected differences in the MCMC sampling time\n----------------------------------------------\n\nDue to more model parameters and model complexity in Meridian, MCMC\nsampling in Meridian is expected to take longer than in LightweightMMM.\nHowever, because the models are relatively similar, Meridian is not\nexpected to take much longer than LightweightMMM. Precise estimates on how much\nlonger depends on the compute environment, number of geos, model tuning\nparameters, priors, data, and other factors. Although Meridian's model\ncomplexity likely leads to longer MCMC sampling time, more accurate results are\nexpected."]]