Starbucks Promotion Analysis

ここをクリックして日本語での要約を展開

概要

本プロジェクトではスターバックスのオンラインクーポンに対する顧客の反応について分析した。あるクーポンを見た時（まず見る前提で考える）に、その「顧客の属性」、「送られたクーポンの属性」、「その顧客の過去の購買行動」をもとにユーザーが送付されたオンラインクーポンを達成されるか否かを予測する機械学習モデルを構築した。

クーポンは３種類ある。

1. BOGO(ある会計において規定額に達したら一定額の割引がなされる)
2. Discount(クーポンが送られてからの累計使用金額が規定額に達したら一定額の割引がなされる)
3. Information(ある一定期間閲覧可能なインフォページが送られる)

注意点として

1. クーポンの有効期間はクーポンが送られた瞬間からであり、見られた瞬間からではない
2. そのためクーポンを閲覧できる時間はクーポンにより異なる
3. クーポンは見られたかに関わらず規定額に達したら割引がなされる

関連URL

• 解析を行ったJupyter notebook2本

工夫した点

各個人の過去のクーポンに対する行動を、その人間の新たな「属性」として落とし込む点。 過去に送られたクーポンの総数、各クーポンの種類の数、クーポンを見た時のクーポンの残り有効時間は各顧客によってばらばらである。これらを各顧客に有限個の変数に落とし込むことに苦戦した。

これは最終的に、

1. 過去にその顧客に送られたクーポンが開封されるのにかかった時間の分布
2. 過去にその顧客がクーポンを開封した際のその後の達成率
3. 各顧客がクーポンを受け取った時の状況を「有効期間までの残り時間が5日8時間以内 or 5日8時間より多い」✖️「達成までの残り必要使用額が$10未満 or $10以上」の４パターンに分け、それぞれの状況での過去の達成率(0~1)

という３つの観点から変数に落とし込んだ。なお、今回はデータ数の制限から3.で4つの状況に分類したが、これに加え、「達成時の割引額の大きさ」に応じて状況をさらに２分割し合計８パターンの状況に分割することでより精度が改善されることが期待される。

過去に顧客のクーポンを見た瞬間の状況を４つのセグメントに分けて考えるというアイデアは以下のポストを参考にした。

https://note.mu/hik0107/n/n854ff66b2621

どう役に立つのか

「顧客に見せるクーポンを顧客がクーポンを開封した瞬間に決定できる」という前提を置くとする。

この時、表示するクーポンの有効残期間、どれだけの達成基準額、どれだけのクーポン額によって顧客がクーポンを達成できるかどうかが推測できる。これをもとに達成が予想されるクーポンを表示させることができる。（なお今回はモデルの解釈可能性を優先して、ランダムフォレスト分類器を用いたが、ロジスティック回帰などによって達成確率を０〜１までの確率として出すことによって、クーポンを各個人に応じてより細かく最適化することが可能になる）

なおこの方法では問題が２点ある。

1. 有効残期間 現在の仕様ではクーポンの消滅期間はクーポンの中身に依存している。以上の手法を取り入れるにはクーポンの消滅期間は開封されなかった場合５日、などの新たな前提を置く必要がある。
2. 利益になるか 今回のモデルではクーポンの開封を予想するだけであり、例えばクーポンの特典額を$1000にした場合当然であるが達成率は跳ね上がるだろう。しかしながらこれは会社にとって利益になるわけではない。どのクーポンを送ると利益を生み出すことになるのかは、顧客がクーポンを受け取っていない時の日常的な使用額と、クーポンを送った際の使用額期待値の大小比較によって決定できる。

Project Overview

This is my final project in Udacity Data Scinentist Nanodegree. In this project, I will create a predictive model to predict whether a customer would complete a coupon when they first view it. (It is tricky that even if a customer does not view or display the coupon in the shop, the coupon can be automatically used once he spends the amount needed.)

Data set used in this project is a simplified version of the real Starbucks app.

Relevant URLs

• two Jupyter notebooks where this project was done

Repo Structure

.
├── 190315StarbucksPromotionAnalysis.ipynb
├── 190403StarbucksPromotionAnalysis.ipynb
├── README.md
└── data
    ├── firstly_viewed_offers.csv
    ├── interaction.csv
    ├── merged_df2.csv
    ├── portfolio.json
    ├── portfolio_clean.csv
    ├── profile.json
    ├── profile_clean.csv
    ├── transcript.json
    ├── transcript_clean.csv
    └── model.pkl

Data Overview

This data records 17k customers' behaviors, including receiving offers, opening offers, and making purchases. As a simplification, there are no explicit products to track. Only the amounts of each transaction or offer are recorded.

There are 3 types of offers that can be sent. Offers can be delivered via multiple channels.

 * buy-one-get-one (BOGO)
 (a user needs to spend a certain amount at a time to get a reward equal to that threshold amount)
* discount
 (a user gains a reward equal to a fraction of the amount spent)
* informational.
 (there is no reward, but neither is there a requisite amount that the user is expected to spend.)

Dataset Details

1. portfolio.json : Coupons detail (10 types of coupons x 6 fields)

id (string) - offer id
offer_type (string) - type of offer ie BOGO, discount, informational
difficulty (int) - minimum required spend to complete an offer
reward (int) - reward given for completing an offer
duration (int) - time for offer to be open, in days
channels (list of strings)

2. profile.json : Rewards program users (17000 users x 5 attributes)

gender: (categorical) M, F, O, or null   
age: (numeric) missing value encoded as 118   
id: (string/hash)   
became_member_on: (date) format YYYYMMDD   
income: (numeric)   

3. transcript.json : Event log (306648 events x 7 fields)

person: (string/hash)
event: (string) offer received, offer viewed, transaction, offer completed
value: (dictionary) different values depending on event type
offer id: (string/hash) not associated with any "transaction"
amount: (numeric) money spent in "transaction"
reward: (numeric) money gained from "offer completed"
time: (numeric) hours after start of test

Project Goal

The goal of this project is to create a machine learning model to predict whether a customer would complete the offer WHEN THEY VIEW THEIR OFFERS FIRST. (Even though it is suggested that　coupons can be completed without being viewed in this scenario.)

In simple words, I made a model to answer this question.

When a person view a coupon, would the person accomplish it?

[NOTE] This model can only predict whether a customer would complete the coupon or not. Whether issuing coupon to the customer would be beneficial to the company is not discussed here. For example, If a coupon which makes $1000 rewards is issued, almost all customers would accomplish it. It has to be decided what kind of coupon would attract a customer and also still beneficial to the company based on the behavior taken by the customer without any coupon (which is not given in this dataset).

Project Approach Overview

In this project, I took these steps below.

(In 190315StarbucksPromotionAnalysis)

1. Exploratory analysis of the dataset
2. Make a dataframe whose each row represents each coupon sent in this survey period.

(In 190403StarbucksPromotionAnalysis)

3. Make a dataframe whose each row represents the last coupons which were sent to each customers.  
(this dataframe contains only customers to whom at least 2 coupons had been sent before the last one.)
4. create a predictive model from dataframe of (3)
5. interpret the model

Detail of Predictive Model

Model variables

When customer A receive a coupon X, my predictive model uses 3 types of variables as X.

1. Type of coupon X (BOGO or Discount, how much more did he had to spent when he saw the coupon)
2. Attributes of a customer A (age, income, gender...)
3. Behaviors taken by customer A when he got coupons before this coupon X
4. Situation when the customer viewed coupon X (how many hours are left etc.)

And this model predicts y as 1 (will complete) or 0 (will not complete).

Dataset size used for creating the model

Since I decided to predict the customer's latest reaction to the coupon,

2844 customers who received 2 coupons before the latest one
873  customers who received 3 coupons before the latest one
111  customers who received 4 coupons before the latest one

These 3828 customer's action to the latest coupon is my concern.

I took 10% testing data and create a model from 90% dataset. I used Random Forest Classifier and tuned hyperparameters by using Grid Search.

The final model tuned predicted testing data with micro f1 score 0.80. This model is pickled in a file "data/model.pkl".

Model variables

'age', 'income', 'gender_F', 'gender_M', 'gender_O', 'became_year', 'became_month_sin', 'became_month_cos', 'became_day_sin', 'became_day_cos', 'became_dow_sin', 'became_dow_cos', 'offer_bogo', 'offer_disc', 'difficulty', 'duration', 'reward', 'email', 'mobile', 'social', 'web', 't_received', 't_viewed', 'amt_till_viewed', '0.0', '6.0', '12.0', '18.0', '24.0', '30.0', '36.0', '42.0', '48.0', '54.0', '60.0', '66.0', '72.0', '78.0', '84.0', '90.0', '96.0', '102.0', '108.0', '114.0', 'past_completion_rate', 'short_little_comp_rate', 'short_lot_comp_rate', 'long_little_comp_rate', 'long_lot_comp_rate', 't_left_when_viewed','amt_needed_when_viewed'

Features ['short_little_comp_rate', 'short_lot_comp_rate', 'long_little_comp_rate', 'long_lot_comp_rate'] have null values. I tried 2 ways of imputation.

1. Replace null with the mean completion rates of that column (that is, the mean rate in each of situation)

2. Replace null with the mean completion rates of that person's "past_completion rate" column

With the data given in this project, the models created by these 2 ways recorded almost the same f1 score in testing after their tuning. (The 1st model was saved in "model.pkl")

Conclusion (interpretation)

These are top 10 important features of this model.

1   past_completion_rate   0.10823659095118901
2   long_lot_comp_rate   0.08595954918332138
3   short_lot_comp_rate   0.08367514878558703
4   long_little_comp_rate   0.07800267360062899
5   short_little_comp_rate   0.07629602818899717
6   income   0.05515854732093211
7   became_year   0.04978646626207604
8   t_left_when_viewed   0.04372529933213142
9   age   0.032500881997055905
10   reward   0.028411643826730427

[1]Features representing past completion rate of the customer

past_completion_rate(1)  
long_lot_comp_rate(2)  
short_lot_comp_rate(3)
long_little_comp_rate(4)   
short_little_comp_rate(5)  

are about past completion rates and it makes sense that these are the most important features of all. Feature engineering in this notebook was worth doing!

[2]Features representing attributes of the customer

became_year(6)  
income(7)  
age(9)  

are about attributes of the customer. It is concluded that the persons' attributes are more relevant to whether he would achieve the coupon rather than the coupon's type.

[3]Features representing the coupon

t_left_when_viewed(8)  
reward(10)  

As was expected, remaining time and and the reward of a coupon have affect on customers' behavior. It is interesting that how much more amount of purchase the customer have to make to achieve the coupon is not important as these 2 factors.