Main recommend system algorithms are listed below.
- based on user
- based on item
idea: find the top similar user/item, recommend what they related with weight(similarity)
what this repo implementes
idea: user/item matrix decomposition, with feature dim
With all item features know, find the relationship bewteen user and features, based on user's behavior history.
Different from neighbor based collaborative filtering algorithm. Content based recommend algorithm take user behavior history into consideration, while neighbor based collaborative filtering algorithm ignore this.
Combine collaborative filtering and content based recommend algorithm, diminishes the shortage of each algorithm. It works like making recommend from each algorithm, and combine them with weights.
load the MovieLens 100K Dataset, parse, devide into train/validation dataset, transform to a USER_NUM * MOVIE_NUM matrix with train dataset, ie train_matrix in code.
Set up tf_train_matrix, tf_user_features, tf_item_features tensor, which could be understood as:
tf_train_matrix = tf_user_features * tf_item_features.T
In one word, train_loss, which is the training objective, is defined as,
train_loss = 0.1 * train_diff_loss + 0.1 * train_user_loss + 0.1 * train_item_loss
train_user_loss and train_item_loss are the regulation terms, root/sum/square of the tf_user_features and tf_item_features, which prevents overfitting.
train_diff_loss is the root squared error of tf_train_matrix and tf_user_features * tf_item_features.T. You might notice that tf_train_matrix doesn't cover all, since it
s quite common(even mostly) that some users didn't see some movies. So These blanks are fill with the average rating of this movie. If nobody see this movie, rating it 3.
Their weights are all equal, 0.1, which is empirical and observation. Why not all 1, because I want loss smaller which makes me feel good. Big loss is always scary, isn't it?
train_avg_dist and val_avg_dist means the average abs rating error between prediction and ground truth.
train_equal_num, train_equal_ratio, val_equal_num, val_equal_ratio stands for the right prediction num and it's ratio of train/test dataset.
It won't take too much time. I set a lot of scalars to watch in Tensorboard during training. Names of scalars need well organized to utilize the group scalar in Tensorboard.
The best accuracy is 33.59%, the best average abs error is 0.9294 before overfitting, which means 1/3 prediction are right, and average prediction is 1 rating distance from ground truth, seems to be acceptable in suggestion area.
