Data Source

• From Kaggle Anime Recommendations Database
• Recommendation data from 76,000 users at myanimelist.net
• https://www.kaggle.com/CooperUnion/anime-recommendations-database
• with Two files : anime.csv rating.csv

Data Source

anime.csv

Data Source

anime.csv

• anime_id - myanimelist.net's unique id identifying an anime.
• genre - comma separated list of genres for this anime.
• rating - average rating out of 10 for this anime.

Data Source

rating.csv

Data Source

rating.csv

• user_id - non identifiable randomly generated user id.
• anime_id - the anime that this user has rated.
• rating - rating out of 10 this user has assigned (-1 if the user watched it but didn't assign a rating).

Data Visualization

Data Preprocessing

• There are 41 kinds of genres for all animes.
• For each anime, we create a 41 * 1 vector to represent.
• And for each user,  we create a 41 * 1 vector from the ratings.

Data Preprocessing

How to deal with the missing values?

for each user, replace -1 with

      5 (If there are no ratings from this user)

      10 - average rating from this user (otherwise)

Data Preprocessing

User Instance

Evaluation

NDCG@50

CG

• Cumulative Gain
• Example:
  gain = {3, 5, 9, 6}
  CG = 3 + 5 + 9 + 6 = 23

DCG

• Discounted Cumulative Gain
• Example:
  gain = {3, 5, 9, 6}
  DCG = 3 / lg(2) + 5 / lg(3) + 9 / lg(4) + 6 / lg(5) = 13.24

NDCG@k

• Normalized Discounted Cumulative Gain
• k: number of items in DCG
• NDCG = DCG / iDCG
• iDCG: ideal DCG
• Example:
  gain = {3, 5, 9, 6}
  DCG = 3 / lg(2) + 5 / lg(3) + 9 / lg(4) + 6 / lg(5) = 13.24
  iDCG = 9 / lg(2) + 6 / lg(3) + 5 / lg(4) + 3 / lg(5) = 16.58
  NDCG@4 = 13.24 / 16.58 = 0.80

Pros & Cons

• Offline: needless of user response
• Relevance could be any real number (in contrast to MAP which only allows binary relevance)
• The returning value itself is nontrivial and thus can only be used to distinguish between models. 

Usage

Reference

• https://www.microsoft.com/en-us/research/publication/evaluating-recommender-systems/

Model

Model 1: KNN

• Put anime vector to build a k-d tree
• Fit user vector with k-d tree with cosine distance
• Recommend some nearest animes

Model 2: Anime Clustering

• Group animes into several clusters
• Fit user instance into one of the clusters
• Recommend animes in that cluster with high rating

Model 3: User Clustering

• Group users into several clusters
• Fit user instance into one of the clusters
• Recommend animes in that cluster with high rating