TikTok Hackathon 2023- Optimize Advertisement Moderation

The ScoreMatcherzz

Developer:

Bryan Aurelius Tanaja - Students from Singpore University of Technology and Design majoring in Engineering Systems and Design. Soloing the project, did the whole project and developed both ranking and matching algorithm

Development Tools Used

• Python
• Gurobipy
• Sklearn
• Pywraplp

Assets Used

• Ads data:p_date ad_id, delivery_country, queue_market, punish_num, latest_punish_begin_date, ad_revenue, avg_ad_revenue, start_time, baseline_st, product_line, task_type_en -Moderator data: moderator, market, Productivity, Utilisation %, handling time, accuracy
  

Libraries

• NumPy & Pandas : For data processing and numerical operations.
• Sci-kit Learn : Used for linear regression models and random forest
• Gurobipy : For optimization problems.
• Pywraplp : For optimization problems.

Inspiration

We recognized the untapped potential in the world of social media advertising. While the current advertisement moderation systems work reasonably, the balance between monetization and user safety wasn't optimized. This became our muse, urging us to create a dynamic model that not only enhances the efficiency of ad moderation but also optimizes the matching of ad content to moderators.

What it does

Our solution optimizes the social media advertisement moderation process through a dynamic scoring system. This system prioritizes ads for review based on their value and risk. The highest priority ads get reviewed first, ensuring swift monetization of valuable content and rapid intervention on high-risk ones. Further, it matches each ad with the best fitting moderator, factoring in aspects like language proficiency and industry expertise. The dual objective is clear: reduction in user risk and maximization of traffic monetization.

How we built it

1. Data Processing and feature engineering
2. Scoring Model : Use gurobipy to optimize and random forest to predict for future use. Matching Mechanism: build 2 model using gurobipy and pywraplp to match and use linear regression machine learning model

Challenges we ran into

-Ensuring data consistency from various teams was initially a challenge, given the diverse formats and systems used. -Optimizing the weight of each variable -Building a seamless matching mechanism that minimized mismatch rates proved to be a complex task.

• Improving the accuracy of the model -integrating machine learning model to optimization model.

Accomplishments that we're proud of

Developing a model that can be dynamically updated as new data and parameters emerge, making our solution future-proof.

What we learned

• Learn how to process and feature engineer data in deeper level
• learn to define objective as the beginning steps
• understand optimization model deeper

What's next for TikTok Hackathon 2023- Optimize Advertisement Moderation

-Model Refinement: Continuously feed back data to refine and enhance the model. -Queueing theory: Try to implement more of queueing theory -Integrated Learning: Implement machine learning techniques to auto-update the model based on evolving ad trends and moderator performance.

Built With

• gurobipy
• python
• pywraplp
• sklearn