近年来，随着移动设备的发展和普及，线上购物用户规模迅速扩张，面对如此大规模的用户，如何进行个性化营销是目前迫切需要解决的命题。传统的响应模型使用机器学习方法，通过用户以往浏览、消费等行为预测用户接受营销后的转化概率，对转化率更高的用户优先分配营销策略，这种方法会同时命中营销敏感群体和自然转化群体。而智能营销模型使用因果推断方法，通过估计因果增益评价营销的效果，可以精准识别出营销敏感人群，不仅使具有一定偏好的消费者得到优惠，同时使商家收获更高的经济效益。

首先，本文介绍了互联网营销策略的发展过程及经典的因果推断模型。其次，使用MineThatData数据集进行实证分析，对数据进行特征编码以使特征具有更好的解释性，对实验组进行随机重采样使组间数据量达到平衡，对模型的前置性假设(正值假设、无混淆性假设和可忽略性假设)进行检验，对响应变量显著性进行检验以探究协变量和处理变量对响应变量的影响。再次，运用双机器学习、元学习和双重鲁棒学习方法建立条件平均处理效应估计的因果推断模型。对比模型结果表明，使用双机器学习模型估计营销行为对用户是否访问的因果效应效果最优，使用元学习模型估计营销行为对用户是否消费及消费金额的因果效应效果最优。最后，通过因果推断模型输出的特征解释器结果和平均处理效应，对营销敏感人群进行分析。分析结果表明，历史消费金额、末次消费间隔、是否新用户和是否购买过某品类用品是影响用户转化的关键因素，最终得出营销敏感人群圈选规则。

本文建立基于因果推断的智能营销模型，采用MineThatData数据集进行实证分析，运用双机器学习、元学习和双重鲁棒学习的方法估计条件平均处理效应并依据模型输出的结果进行营销敏感人群分析，验证了因果推断模型可以解决智能营销中营销敏感人群的识别问题。

In recent years, with the development and popularization of mobile devices, the scale of online shopping users has rapidly expanded. Faced with such a large number of users, how to carry out intelligent marketing is currently an urgent issue that needs to be solved. The traditional response model uses machine learning methods to predict the conversion probability of users after receiving marketing based on their past browsing, consumption, and other behaviors. It prioritizes marketing strategies for users with higher conversion rates. This method will hit both marketing sensitive groups and natural conversion groups simultaneously. The intelligent marketing model uses causal inference method to evaluate the effectiveness of marketing by estimating causal gains. It can accurately identify marketing sensitive groups, not only allowing consumers with certain preferences to receive discounts, but also enabling businesses to reap higher economic benefits.

Firstly, this article introduces the development process of internet marketing strategies and classic causal inference models. Secondly, empirical analysis was conducted using the MineThatData dataset to encode the data features for better interpretability. Use random under-sampling method for experimental group to achieve a balanced data volume between groups. The model's preconditioning assumptions (positivity assumption, unconfoundedness assumption and ignorability assumption) were tested, and the significance of the response variables was tested to explore the effects of covariates and processing variables on the response variables. Once again, a causal inference model for estimating conditional average treatment effect is established using double machine learning, meta learning, and doubly robust learning methods. The comparison model results show that using a double machine learning model to estimate the causal effect of marketing behavior on whether users visit is the best, while using a meta learning model to estimate the causal effect of marketing behavior on whether users consume and the amount of consumption is the best. Finally, the feature interpreter results and average treatment effect output from the causal inference model are used to analyze the marketing sensitive population. The analysis results indicate that historical consumption amount, last consumption interval, whether new users have purchased a certain category of products, and whether they have purchased a certain category of products are key factors that affect user conversion, ultimately leading to the selection rules for marketing sensitive groups.

This article establishes an intelligent marketing model based on causal inference, uses the MineThatData dataset for empirical analysis, estimates the conditional average processing effect using double machine learning, meta learning, and doubly robust learning, and analyzes the marketing sensitive population based on the output of the model. It verifies that the causal inference model can solve the identification problem of marketing sensitive population in intelligent marketing.