선택한 단어 수는 입니다.
최소 단어 이상 선택하여야 합니다.
최소 단어 이상 선택하여야 합니다.
선택한 단어 수는 30입니다.
최대 10 단어까지만 선택 가능합니다.
최대 10 단어까지만 선택 가능합니다.
문제 정의
- This paper focuses on memory-based algorithms, especially on the application of movie recommendation. One may use various characteristics related to movies, such as actors, director, genre, year of release, etc.
제안 방법
- Following the observation from the previous subsection, the proposed method takes difference between genre preferences of users into account, in calculating similarity between users. The genre preference is measured through the number of ratings; high number of ratings to those movies in a specific genre implies that much preference for that genre.
- The accuracy of prediction is usually measured through several metrics in literature such as MAE (Mean Absolute Error) and RMSE (Root Mean Squared Error) [15]. In this paper, we use MAE to estimate the prediction precision of our method. MAE=# for N predictions # for the corresponding real ratings ri.
- However, these works are usually integrated into the traditional similarity measures or require extra process to convert the user-item rating matrix. The proposed method does not require extra information on user-preferred types of movies, but simply utilizes the number of ratings made by users for each genre, which is interpreted as the genre preference of the users. With this simple treatment, i.
- 1] shows the changing patterns of the two metrics according to similarity. Three popular similarity measures are used in this experiment on MovieLens dataset; cosine similarity, Pearson correlation, and Mean squared difference (MSD). It is observed that there seems no big difference between the behavior of JACgr and NRcgr, This is an obvious result, considering that a user gives ratings to more movies in his interested genres than to movies out of his interested genres.
- The dataset consists of 1,000,209 ratings of 1~5 scale made by 6,040 users on 3,900 movies approximately. To secure the reliability of the result, we implemented 5-fold cross validation of the results where each experiment was repeated 5 times with different training and testing data. 80% of each user's ratings were used for training, i.
- Its distinctive property that is differentiated from the classic measures is that it reflects only users’ preference for genres and number of co-ratings per genre onto similarity calculation. With such approach, it is quite notable that the proposed measure significantly outperforms previous similarity measures in terms of both prediction and recommendation qualities in our experiments, considering that it does not utilize differences between ratings made by two users as in traditional similarity measures. This implies that there is much room for further improvement of performance in collaborative filtering methods.
대상 데이터
- As our idea is for movie recommendation, the MovieLens dataset is selected, as it has been popularly used in the related study and publicly open for research purpose. The dataset consists of 1,000,209 ratings of 1~5 scale made by 6,040 users on 3,900 movies approximately. To secure the reliability of the result, we implemented 5-fold cross validation of the results where each experiment was repeated 5 times with different training and testing data.
- , whether genre interests between similar users are more in common than between less similar users. The dataset used for this experiment is MovieLens (http://www.grouplens.org), where information on genres which a movie belongs to is also provided. Since no specific information regarding genre preference of users can be obtained from the dataset, we assume that the number of ratings made by a user to movies belonging to a genre indicates the level of preference of the user for that genre; the more ratings obviously implies the higher interest.
성능/효과
- However, COS demonstrates far lower results than all of the others, while GPRF yields much higher performance compared to even MSD, where around 9~13% improvement has been achieved throughout varying number of nearest neighbors. Consequently, all of these results demonstrate that considering only preference for movie genres and number of co-rated items in each genre to estimate similarity, as suggested in this research, surely enhances both prediction and recommendation accuracy in collaborative filtering systems.
- To investigate F1 performance of the measures, it is seen that the results are much similar to those of MAE. However, COS demonstrates far lower results than all of the others, while GPRF yields much higher performance compared to even MSD, where around 9~13% improvement has been achieved throughout varying number of nearest neighbors. Consequently, all of these results demonstrate that considering only preference for movie genres and number of co-rated items in each genre to estimate similarity, as suggested in this research, surely enhances both prediction and recommendation accuracy in collaborative filtering systems.
- , considering ratings not as a whole, but grouped into each genre, we confirmed that performance of recommendations is greatly enhanced, as proved through several experiments. Its main contribution is to reveal that utilization of only the features other than rating differences between users significantly improves performance of CF methods which have been solely dependent on rating differences.
※ AI-Helper는 부적절한 답변을 할 수 있습니다.