MIT+ Fintech Use of Machine Learning Methods in Personalization and Targeting

Use of Machine Learning Methods in Personalization and Targeting

• When the methods are useful? Robustness
• Method adaptations: Dynamic Targeting
• Social impact. :Algorithems bias

Robustness

• Rebustness:Data Challenges

  • Covariate shift(stable)

  • Concept shift（Mapping changes)

    • Timechange?customers’ response.
    • not well studied

  • Aggregation of targeting variables

    • what is the aggregation?

    • Use individual level to train the model

    • In practice: data is very sparse:

      i.e. their purchase behavior is just very occasionally.

      Use sparse data to train the model, performance may not be very well.

    • use data to accomodation

    • 这里的Sparse:finite parameters 说明顾客的行为是不太好预测的。

    • 在传统机器学习的，pushnishing 或正则化 湿因为becasue of sparsity，来防止过拟合

• Robustness: Targeting Methods
  • Distance driven methods:
    • kernel regression
    • k-NN
    • HC Trees
    • when predicting ,least distance
  • model driven methods
    • lasso regression: a very powerful method.
    • Finite mixture models:parameter assumptions
  • classification methods.
    • CHAID
    • SVM
  • why not neuron nets: these methods are easy to express. and neuron nets need good train data. Managers need meaning, they don’t need maths.
• Robustness: Data
  • A large retailer in US
  • 2 types of promotional offers
  • Use 2 large-scale field experiments.
    • train data
    • use the first 1 and marketing strategy on the second .

Why model-driven methods are best?

Cool data-set : distance: don’t have data close to them.

Model-driven:smooth the data in this dataSet.

Consider the robustness

How to adjust the methods to make them stable.

过拟合问题。

How to adjust? I don’t know.

---

Dynamic targeting:constructing Markov states for reinforcement learning

BG

Firms want to target their customers with a sequence of marketing actions

1. Marketing actions focus on *long-term *implications
2. Standard targeting models are short-angled.

Can be modeled as Markov Decision Process

• To be solved by a range of Reinforcement Learning algorithms

• State:summarization :classify customers into different groups, customers meet in different space and time.
• mail/not mail: c.f.招商： 实际上有很多的marketing actions
• R：从state 到reward 的映射
• P：转移概率，

目标当然是 ：maxmize profits in the long run .

什么是reinforcement learning?

引例子

让智能体在环境里学习，每个行动会对应各自的奖励。

c.f 人类学习中与环境进行交互，

A Typical Reinforcement Learning Algorithm

1. States: what is your state

   

consumers are very hard to study.

what is your criteria for states?

implicit assumption

Proposed Method

a mapping:

Map(historical transaction space) to (Probability space of future purchasing events)

This means we do predictions based on your marketing actions.

If prediction is same , then you are the same type of customers. we care about the future.

Why markov

Non-Markov distortion

Disadvantages

• too rely on data value.
• dp , not too much on deep reinforcement learning: it is very unstable.
• Bench reinforcement learning.

Algorithm bias：Social impact

Is women less likely than men to receive STEM ads given gender neutral targeting.

STEM need be gender-neural

Estimated: gender’s likelihood receiving STEM ads.

• algorithm doesn’t do gender-neural.

What is the cause?

因为women的点击率高，所以不需要发这么多。

discrimination v.s. results.

Key reasons: competitive spillover: Men-impression is cheaper

Ads的投放是竞价，need to pay more in order to win women’s impression.(bid for there impression)

Don’t focus on algorithms too much .

Conclusion

• When the methods are useful?
• Method adaptations
• Social impact.

赏

谢谢你请我吃糖果❤️

支付宝