[KSSCI 2021] 인과추론의 데이터 과학 - Session 12

인과추론의 데이터과학. (2021, Oct 21). [Session 12] 머신러닝을 통한 비정형 데이터 분석 (마케팅 연구사례) [Video]. YouTube.

 

Session 12

기존 review 영향 연구의 한계점

• customers read rich review content, instead of only considering aggregated volume, rating, and variance metrics → 모델에 어떻게 address해야할까
• review content are significantly different across product categories
  • ex. Watch: “Accurate” and “Waterproof” → 카테고리별 다른 단어를 어떻게 반영할까

 

Large-Scale Cross-Category Analysis of Consumer Review Content on Sales Conversion Leveraging Deep Learning

Research Questions

• RQ1. Do consumers read review content beyond summary statistics?
  • When do they read?
  • Who reads it?
  • Where do they read it?
  • What information in the reviews do they pay attention to?
• RQ2. What’s the impact of the reviews “read” on conversion?
  • Challenge1. Data on review-content-reading behaviors are not available → 데이터 부족 ⇒ UK에서의 아마존 데이터 (300개 이상 카테고리)를 활용
  • Challenge2. Review content is represented by unstructured words or sentences → 어떻게 high-level feature로 가져올까 ⇒ 어떤 high-level feature가 중요한지 정의함 + ML 활용

출처: 인과추론의 데이터과학

Identification Strategy

• Using Regression Discontinuity
  • 소비자들이 online platform에서 review가 discontinuity하게 노출될 것

출처: 인과추론의 데이터과학

 

First Law of Motion: Influencer Video Advertising on TikTok

• Contents 소개

 

출처: 인과추론의 데이터과학

 

Research Question

• How to predict the incremental effect of video ads on product sales
• Challenges
  • Challenge 1: No systematic way to predict sales lift based on video ads content
  • Challenge 2: Dimensionality of Average TikTok Video
    • 15 Seconds * 60 Frames per Second * (1080p * 1920p) per Frame
    • 1,866,240,000 Pixels per Video as the independent variables to the algorithm

 

Overcome technical challenge (“M-Score”)

• Step1. Compute a content engagement heatmap over the video ad
• Step2. Compute a product placement heatmap over the video ad
• Step3. Compute m-score as the normalized inner product of the two heatmap

출처: 인과추론의 데이터과학

 

Engagement Heatmap

• Content Engagement Heatmap
  • Step 1. Train 3D Convolutional Neural Network
    • to predict video-level engagement (e.g., # likes) with a large sample of videos ads on TikTok using cross-video variation in content and engagement
  • Step 2. Extract 3D Saliency Map over the video that assigns an engagement score to each pixel
• Product Placement Heatmap
  • Scale Invariant Feature Transform (SIFT)
  • object detection by constructing key points of reference (product) and target image (video frame) then match based on some distance metrics

출처: 인과추론의 데이터과학

 

Empirical Results

출처: 인과추론의 데이터과학