Fix a bug

Fix a bug

Author: shah314

README.md

@@ -12,7 +12,7 @@ BPR puts a zero mean Gaussian prior on the learnt latent factors (embeddings) wh
 
 The implementation requires python3, pandas and numpy. The dimensionality can be changed in util.py. Currently it uses a dimensionality of 50 and 10 iterations. For larger dimensionality and more number of iterations, it might be useful if there is an access to a cluster of servers or a GPU.
 
-The hit rate @ position 10 is around **71%** (on a 30% test set) (the hit rate was initially 51% but I was able to do some parameter tuning and improve this to 71%) on the MovieLens [2] data set with 600 users, 9000 movies and 100,000 ratings. The data is uploaded to the data folder for convenience.
+The hit rate @ position 10 is around **51%** (on a 30% test set) on the MovieLens [2] data set with 600 users, 9000 movies and 100,000 ratings. The data is uploaded to the data folder for convenience.
 
 The MovieLens data sets can be found on the [MovieLens Web Page](https://grouplens.org/datasets/movielens/).
 

python/bpr.py

@@ -12,14 +12,15 @@
 
 @author: shah
 """
-from util import m_normal, learning_rate
+from util import m_normal, learning_rate, get_lambda
 from classes import ret
 import random as random
 import numpy as np
 import math
 def bpr_update(users, movies):
     count = 0
     lr = learning_rate()
+    lam = get_lambda()
     for u1 in users:
         u = users[u1]
         userid = u.userid
@@ -38,20 +39,20 @@ def bpr_update(users, movies):
                 diff = Vi - Vj
                 d = firstterm * diff
                 derivative = d
-                Vu = Vu + lr * (derivative + 0.1 * np.linalg.norm(Vu))
+                Vu = Vu + lr * (derivative + lam * np.linalg.norm(Vu))
                 users[u1].factor = Vu
 
                 # ITEM POSITIVE FACTOR
                 d = firstterm * Vu
                 derivative = d
-                Vi = Vi + lr * (derivative + 0.1 * np.linalg.norm(Vi))
+                Vi = Vi + lr * (derivative + lam * np.linalg.norm(Vi))
                 movies[rand_pos].factor = Vi
 
                 #ITEM NEGATIVE FACTOR
                 negvu = -1 * Vu
                 d = firstterm * negvu
                 derivative = d
-                Vj = Vj + lr * (derivative + 0.1 * np.linalg.norm(Vj))
+                Vj = Vj + lr * (derivative + lam * np.linalg.norm(Vj))
                 movies[rand_neg].factor = Vj
 
 def calculate_first_term(Vu, Vi, Vj):

python/filereader.py

@@ -48,12 +48,11 @@ def read_ratings(filename):
                     user1.movies_train[movieid] = rating1
                     users[userid] = user1
                     trainuserdict[userid] = 1
-
+            else:
                 testcount = testcount + 1
                 if userid in users.keys():
                     user1 = users[userid]
                     user1.movies_test[movieid] = rating1
-
                 else:
                     user1 = user(userid)
                     user1.factor = random_vector()

python/util.py

@@ -20,6 +20,9 @@ def min_rating():
 def learning_rate():
     return 1
 
+def get_lambda():
+    return 0.1
+
 def random_vector():
     dim = dimension()
     cov_mtx = cov_matrix()