Song Recommendation Engine Tianlai Karaoke App Gao Wa, Cui - - PowerPoint PPT Presentation

song recommendation engine
SMART_READER_LITE
LIVE PREVIEW

Song Recommendation Engine Tianlai Karaoke App Gao Wa, Cui - - PowerPoint PPT Presentation

Song Recommendation Engine Tianlai Karaoke App Gao Wa, Cui Xiaoting, Zheng Feng Advisor: Professor Li Yanhua Technical support: Xu Hengyu (Tianlai) Tianlai Karaoke App What is Tianlai? Your own karaoke platform Record your singing,


slide-1
SLIDE 1

Tianlai Karaoke App

Gao Wa, Cui Xiaoting, Zheng Feng Advisor: Professor Li Yanhua Technical support: Xu Hengyu (Tianlai)

Song Recommendation Engine

slide-2
SLIDE 2

Tianlai Karaoke App

What is Tianlai?

  • Your own karaoke platform
  • Record your singing, and post your songs on

your channel to share with others

  • Add friends and build your interest circle

Tianlai is popular

  • Award for the best creative app and most popular music app
  • 35 million TV installation & 0.4 billion mobile installation
slide-3
SLIDE 3

Problem Statement

  • How to increase target users engagement with the App?
  • Build a good recommendation system
  • Identify the factors that most affecting users song preference
  • Find users’ features that are most useful for song recommendation
  • Whether user would sing along on recommended songs?
  • The accuracy of our recommended songs
slide-4
SLIDE 4

Sample Data

  • Total of 33,489,549 records of 95,109 users and 329,789 songs (2013-2018)
  • 50 Features : User features + song features + singer features
slide-5
SLIDE 5

Data

slide-6
SLIDE 6

Data Preprocessing

  • Missing Values
  • Features selection
  • Create new feature
slide-7
SLIDE 7

Data Preprocessing

  • Combining records

Total 5000 records 20 users 4 types of interactions 100 songs

slide-8
SLIDE 8

Methodology

Engine

Data Source Training Data Data Preparater Prepared Data

Model 2

Algorithm 2 Algorithm 1 Algorithm 3

Model 1 Model 3 Recommendation

slide-9
SLIDE 9

Recommending Process

slide-10
SLIDE 10

Algorithms

1. Predictive Model a. XGBoost b. Random Forest c. SVR 2. Collaborative Filtering a. User-user based b. Item-item based

slide-11
SLIDE 11

Collaborative Filtering

slide-12
SLIDE 12

Collaborative Filtering

User- item matrix User-feature matrix

user id song id 1 2 3 1 14 19 23 2 36 89 49 User id Singer - male Singer - female Song - type 1 1 14 19 23 2 36 89 49

slide-13
SLIDE 13

Results

Algorithm Root Mean Square Error Algorithm Root Mean Square Error Random Forest 51.46 User-user CF 26.85 XGBoost 49.93 Item-item CF 28.39 Support Vector Machine 51.68 User-user CF (user-feature matrix) 29.29

slide-14
SLIDE 14

Results

XGBoost Random Forest

slide-15
SLIDE 15

Future Work

  • Optimize the accuracy of models
  • Add new algorithms such as deep learning
  • Deploy recommendation engine on Karaoke App

We are looking for students to join our team! You are welcome to join through independent study or volunteering work!

slide-16
SLIDE 16