This time we'll talk about the hybrid recommendation system. I'm Jaegwon Kim from Yangon University. Let's talk about what we did content. Firstly, I want to talk about advantages of hybrid recommender systems. Before we talk about the CF and MF hybrid system, let's see with the principle of hybrid recommender system. What would it be the advantage of hybrid RSs? When we use the recommendation by combining multiple recommendation algorithm, the combining multiple recommendation algorithms often more accurate than using a new one algorithm. The performance is often affected by how multi-algorithms are combined. The main principle of hybrid system for recommend-er systems is like this. This structure is the whole sequence for the hybrid systems. Firstly, we can see the input for this whole structure. Every input would be entered to the each hybrid systems, in this case, hybrid system 1, 2, and 3. In each hybrid system, recommend-er their recommendation wrist to the combiner. Then the combiner combine older result. Then finally makes a recommendation list as your final result. Let's get some example for it. For this implementation, we also did the same input for the main libraries like this. This stage also the data load and pre-processing. For calculating the RMSE index, we can implement this method RMSE2. Here is two different recommender algorithms as a method. One is our recommender 0, the other is recommender 1. Actually these are the tummy recommendation algorithms. They just made a random generation using random method. Each of them, they made a recommendation this as a number array. Finally, we can combine the each result as a hybrid recommendation result. In here, we need to decide the weight for each hybrid systems. With 0.8 and 0.2, there are different weight for each hybrid systems. Our prediction 0 and prediction 1 is the result for each hybrid system. Finally, the recommendations gets the final recommendation list with weighted sum. The RMSE tool measured the final prediction result for the topmost evaluation. This is the example of the performance. But as I said, this result is based on the tummy recommendation algorithms, so the performance is so bad. Then what would it be the hybrid system with Collaborative Filtering and Matrix Factorization? With this implementation, we also did the same things as we did, which are firstly, import the libraries and the data load and pre-processing, and train and test split on here. Also, we need the RMSE measurement for evaluation. In the main part, we firstly did the rating matrix with the train dataset. Then this part is for the calculating cosine similarity for train set users. With the cosine similarity method, we can get the user similarity and we can modify this user similarity with the Pandas DataFrame. We can also get the rating mean and rating biases age with these two lines. I want to use def CF knn bias vessel like this. This method gets some inputs from user id, movie id, neighbor size equals zero. We get some condition check on here and there. The first one is for the movie id and the second one is when the neighbor size is general. With this whole sequence, we can get the prediction, and then the method returns the prediction value. This is for the matrix factorization. As I explained before, this int method is for setting all the hyper-parameters and the initial values for matrix factorization. We can also use the RMSE method as a performance evaluation. The lower part of this code stands for the get prediction. With this method, we can get the rating for user i and item j. Also, we can use the SGD method for stochastic gradient descent to get optimized P and Q matrix. This set test method is for selecting the test set. We can use the full roof during the ratings till the end of ratings test. This method returns the test set. For calculating the RMSE for test set, we implemented the test on the RMSE method. We firstly initiate the error equals zero and do the full roof to calculate all the errors based on the RMSE. Firstly, we can calculate the accuracy of the test set while training. This test method, firstly, initialization for the HA matrix such as P and Q and then initialization about bias terms like this. The list of training samples can be on here. The last part is for the stochastic gradient descent for a given number of iteration. With all of these sequences, we can get and return the training processes. Rating for given user_ID and user item_ID, we can get one prediction based on this method. This method calculate for the one predictions based on the model, though this method returns the prediction result. We also can calculate the full matrix, which is user-movie rating matrix. Finally, matrix factorization class can be created with this line, and we can get the result from mf.test method. These iterations are from iterate 10-250. During the process, the Test RMSE comes down till about 0.9. We've already seen the dummy hybrid systems, but this is not a dummies because we can get the recommendation list from these lines, and one each for the matrix factorization, the other is for collaborative filtering based on KNN bias algorithm. By balancing of the two predictions, prediction0 and prediction1, we can calculate each of performance based on the RMSE. Here is the static weight for each algorithms, 0.8 and 0.2. We can simply weight the sum for the result of the hybrid recommender system, and let's get some result with this hybrid system. The result is about 0.9. This performance is the best performance we've seen. With executing this code, we can get all the combinations of variation weight. The weight value start from the zero, and increase with 0.01 margin till the 0.99. As the variation of weight factor, the RMSE value also changing. We can check all the parameters for the best RMSE value on here. The weight is 0.88, and 0.12. This is the optimal parameter for this hybrid recommender system. Let's sum up with this. I said about the advantage of hybrid recommender systems and basic principle of hybrid recommender systems, and we look at the main implementation example about hybrid recommender system with CF and MF. Thank you.
