Bayesian Personalised Ranking or One Class Collaborative Filtering for predicting whether a user reads a particular book. Details explained here
Solution mainly comprised of 3 steps:
Defined an embedding U for each userID and embedding B for each bookID. For each epoch, iterated through Train and for each (userID,bookID) generated N negative samples (userID,bookID’). Maximised Sigmoid( Dot(U[userID].B[bookID) - Dot(U[userID].B[bookID’] ) for all pairs.
(Latent Factor Size = 6, Gradient Descent (Adam Optimizer) with LR = 0.01, Lambda = 0.0001, ~250 epochs with LR reduced to 0.001 after 200 epochs, N = 20)
To make predictions for (userID,bookID) in Val, I looked at all other books read by the particular user and calculated the minimum score assigned (say s). I predicted 0 if the score assigned to (userID,bookID) < 0.50*s (if s > 0) or < s (if s < 0). Otherwise, I predicted 1. (Leaderboard score ~ 70-72%)
Feature Type 1: Look at the Jaccard Similarity between bookID and other books read by this user. Take the mean of the top K highest Jaccard similarities. Do this for K = 1,3,5,7.
Feature Type 2: Similar to Feature Type 1, but calculates Jaccard similarities between users. Do this for K = 1,3,5,7.
Feature Type 3: Popularity of books. Defined as the number of users who read this book divided by the total number of interactions.
Feature Type 4: Activity of users. Defined as the number of books read by this user divided by the total number of interactions.
Trained a logistic regression model to predict 0/1 which takes as input the predictions of OCCF and features extracted in step 2. (Leaderboard score ~ 74-75%). This was fit on Val and evaluated on Test to avoid over-fitting (note that all features were scaled to [0,1] for training).