fnumegha / booklook-2 Goto Github PK

Welcome to BookLook! BookLook is a recommender system for book readers which aims to assist users in the task of finding the next book to read. BookLook will take into account both the dislikes and likes of the reader and his/her friends to recommend the “Next book to read”.

Let's take an example to explain this! Suppose Cav is friends with Trump and Parisa. Now, Cav hates 3 books which Trump also hates & Cav loves 3 books which Parisa also loves. Now if we know that Trump likes "Twilight" and Parisa likes "Life of Pie", then , is it more likely for Cav to like "Twilight" or "Life of Pie"? This is exactly the question we are trying to answer with BookLook!

There are existing book recommending systems in the market. Goodreads is one such example, it provides personalised recommendations based on user interests and collaborative models. These recommendations are based on books the user recently shelved, and from the overall ratings of books provided by all the users on the platform.

However, Goodreads does not seem to take into account the likes and dislikes of a user’s friends while recommending books to the user. BookLook does take this dislike/hate feature into account. [1] talks about how the most powerful signal in predicting a user’s high rating of an item comes from what items he or his friends dislike.

We plan to use the Book-crossing dataset [2]. For implementation we are referring to [1] and will implement the Love-Hate Square counting Method. This method basically uses user-content collaborative filtering using network-based techniques.

In the Love-Hate Square counting method, we would first create a bipartite rating network with two types of nodes (i.e. the users represented by circular nodes and the books represented by square nodes); the edges carrying scores which represent the users’ ratings to items; or all edges can be categorized as an I-love-it or an I-hate-it edge based on whether the rating score for the edge is above or below a threshold.

For a target user-item (u − i) pair, all squares involving the given user-item pair (i.e. paths that start from user u and end at the item i) are identified; the counts for each particular love-hate square configuration are tabulated; the counts for each love-hate configuration are then used as features in a machine learning framework for training and prediction [1].

We ran multiple classifiers for the Love-hate Square Counting Method (SQM), SVM and Logistic Regression performed the best. We also used Matrix factorization and Neural Networks (Restricted Boltzmann Machine) for classification. The results are tabulated below.

Machine Classifier     RMSE   Recall(Love)  Precision(Love)  Recall(Hate)  Precision(Hate) F-measure(Love) F-measure(Hate)
  Naive Bayes          0.664  0.276          0.419            0.744         0.606           0.333           0.668  
  SVM                  0.631  0.083          0.558            0.955         0.607           0.145           0.742
  Logistic Regression  0.628  0.093          0.570            0.952         0.609           0.160           0.743
  Decision Tree        0.633  0.159          0.512            0.897         0.613           0.242           0.728
  KNN                  0.665  0.384          0.460            0.696         0.626           0.419           0.660
  SVD                  0.643  0.949          0.644            0.143         0.636
  RBM                  0.589

We can see that the RMSE for ML Classifiers is on an average 0.6, which is pretty decent. Though if we look at the Precision and recall values we see that the RMSE is good because a lot of entries are being classified as Hate. We see that SVD performs same as well. RBM does give us a better RMSE.

There could be two reasons for the odd precision and recall values : The data is heavily skewed for a threshold value of '6' or even '7'. Secondly, Predicting ratings is a complex multi class problem which we have further reduced to a binary classification probelm.

In this work, we tried to show that hate is as powerful a signal as is love towards a particular commodity and helps in indicating a user's preferences. We showed this by comparing the Love-Hate square counting method against industry used methods like SVD and Neural networks, and we can see that this method performs at par with the industry specific methods in terms of RMSE.

1 Love-Hate Square Counting Method for Recommender Systems.

2 Book-Crossing Dataset.