Convex-opt for advertorial pricing

Convex-opt for advertorial pricing is a project for optimizing the KOL selection problem. This project is based on scipy minimize method.

It is primarily designed to provide effective, robust and explainable optimization architecture for firms who want to select promotion KOLs on social media platforms for their products.

⚙️ Build Virtual Environment

We recommend you to build a virtual environment to run this project.

python -m venv your_venv_name
source your_venv_name/bin/activate
pip install -r requirements.txt

⚒️ Module Structure

root/
  |──src/
  |   |── __init__.py
  |   |── main.py
  |   |── opt_test_v3.py
  |   |── plot.py
  |   |── preproc.py
  |   |── property_scoring.py
  |   |── utils.py
  |   └── property_scoring.sql  
  |── data/
  |   |── param_mat/
  |   |── plotly_fig/
  |   |── opt_data_v2.csv
  |   |── preproc_data_v3.csv
  |   |── scoring_data_v3.csv
  |   |── use_data.csv 
  |   └── voice_data.csv/
  |── tests/
  |   |── __init__.py
  |   |── test_opt_test_v3.py
  |   |── test_preproc.py
  |   └── test_property_scoring.py
  |── .gitignore
  |── README.md
  |── .coveragerc
  |── pytest.ini
  |── clean.sh
  |── demo.sh
  |── requirements.txt
  └── run_test.sh (optional)

🔔 Key Features

Optimization Procedures

The main goal of this project is to minimizing the cost of promotion, so we design a simple objective function to achieve this goal.

$$ \text{Obj function} = \min\mathbf{Count}^\mathbf{T} \mathbf{Price} $$

Where $\mathbf{Count}$ is the amount of total posts and $\mathbf{Price}$ is the price of each post.

Due to the issue of over concentration of post, we add penalty term to the objective function to split the post count of each KOL.

L1 Regularization (Ridge Regression)

$$ \text{Obj function} = \min\mathbf{Count}^\mathbf{T} \mathbf{Price} + \lambda \cdot \mathbf{Count}^{|\cdot|} $$

L2 Regularization (Lasso Regression)

$$ \text{Obj function} = \min\mathbf{Count}^\mathbf{T} \mathbf{Price} + \lambda \cdot \mathbf{Count}^{2} $$

ElasticNet

$$ \text{Obj function} = \min\mathbf{Count}^\mathbf{T} \mathbf{Price} + \lambda \cdot \left(\alpha \cdot \mathbf{Count}^{|\cdot|} + \frac{(1-\alpha)}{2} \cdot \mathbf{Count}^{2}\right) $$

Voice Decreasing Rate

To fit the real world situation, we've designed a voice decreasing rate function to simulate the voice decreasing rate of KOLs. In this project, we use the following function to simulate the voice decreasing rate.

$$ \text{Voice Decreasing Rate} = \frac{1 - e^{-\frac{n}{k}}}{1 - e^{-\frac{1}{k}}} $$

Where $n$ is the number of posts and $k$ is the degree of voice decreasing.

Other Features

Working... writing document is so hard 😭

To Execute

To execute the optimization procedure, I use shell script instead of python script to pass arguments. Regarding the sophisticated optimization procedure, we suggest everyone to first exxecute the demo.sh to see the possible scenarios.

Noticed that you need to change the execution permission of all .sh files by typing chmod +x *.sh in your terminal.

chmod +x demo.sh
./demo.sh

Then you will see something like below:

Select demo scenario:
1. Scenario with plain objective function and budget + voice constraints
2. Scenario with regularized objective function and budget + voice constraints
3. Scenario with regularized objective function and budget + voice constraints + post constraints
4. Scenario with regularized objective function and budget + voice constraints + post constraints + KOL constraints
5. Scenario same as 4 but with unsatisfied constraints that need to adjust the weights.

Enter your choice:

Now, you can choose one of the scenarios to execute, just type the number and press Enter. Then you will see the result of optimization procedure like below:

Output Cost: 665.0 (Thousand dollars)
Output Voice: 6031.192183401445
   platform live_count post_count short_count vid_count live_price post_price short_price vid_price live_voice  post_voice short_voice    vid_voice
0        fb        0.0        0.0         0.0       0.0       25.0       46.0       187.0     250.0        0.0    5.917391         0.0  1194.974967
1        fb        0.0        1.0         0.0       0.0       11.0       10.0       144.0     113.0        0.0   120.45432         0.0   741.681559
2        fb        0.0        1.0         0.0       0.0       11.0        8.0       215.0     105.0        0.0  132.596772         0.0          0.0
3        fb        0.0        0.0         0.0       0.0        5.0        2.0       284.0      93.0        0.0   35.593277         0.0     0.999999
4        fb        0.0        0.0         0.0       0.0       99.0      120.0       202.0     368.0        0.0   281.29993         0.0          0.0
5        fb        0.0        0.0         0.0       0.0       17.0       21.0       203.0     151.0        0.0   78.353186         0.0    38.949981
============== Optimize Done, output information ===============
Selected rows: 20 rows
Objective function types: l2
Budget Constraint: 700 (Thousand Dollars)
Target Voice: 6000
Regularization lambda: 0.5
Minimum count: Live: 5, Post: 7, Short: 5, Vid: 7
Choosed KOLs: KOL ['1', '2', '7', '8', '9']

To Test

To test the optimization procedure, we use pytest to test the functions in src/ folder. To execute the test, you can simply execute the shell script run_test.sh in the root directory.

chmod +x run_test.sh
./run_test.sh