- Wai-Kin Tsui
- Ethan (Yi) Lee
./animerec.py <username>prints the top five recommendations for a user../animerec.py <username> <series_ID>predicts the rating a given user would give to a series. The series ID is the number in the URL of the MyAnimeList page for the series. This may be used for evaluation purposes.
If this stops working, chances are myanimelist.net may have changed their website format. Unfortunately this repository is not actively maintained and we aren't prepared for that possibility.
For our final project, we plan to set up a recommendation system for anime television series based on the watch histories of users who share similar tastes to the target user.
In order to achieve this, we will utilize a vector-based search model. Each user is represented as a vector, and each series they have watched is represented as a component of their vector, with the value of the component being the rating they gave for that series. We will retrieve user ratings from Myanimelist.net. The generation of recommendations will be done by finding the user vectors that are closest neighbors to the target vector, and then predicting the rating that the target user might give to series they haven’t watched based on the (normalized) average rating given by those neighbors to those series. The method of calculating similarity will be dot-product, as it is not sensitive to the norm of the vector, which represents the number of series a user has watched.
Additionally, for predicted ratings, we will implement a critical popularity feature that checks to see how many of the neighboring users have watched the given series. If the number of neighbors that have watched the series is below 10, this might cause the predicted rating to be less accurate, as there is less data from which to calculate the predicted rating. As a result, the program will output a warning that tells the user that the series is comparatively unpopular, and that those results should be taken into context.
Docstrings are available to describe the functions of each source file.
We will evaluate the success of the project by repeatedly picking a random target user, removing one of their watched series’ ratings, and then predicting their rating using the algorithm outlined above. We will then compare the predicted rating and the actual rating. Please see RESULTS.md for the results.
We are planning on crawling user profiles on Myanimelist.net, which are not forbidden from being crawled by the website’s robots.txt file. We will aim to extract approximately 5000 profiles over 4 days, which would be equivalent to roughly 50 profiles per hour, or less than one user per minute on average, which is comparable to, if not slower than, what a human would be expected to be able to do. We will access the website once every 10 minutes, each time extracting at most 20 profiles.
If time permits, we could explore something similar to IDF-weighting. This could be done by assigning popular series lower weights for the closest-neighbors algorithm to see if it improves the evaluation. We could also explore additional options for people to choose recommendations either from all users or only users who have watched a similar number of series, multiple target users in case people might want to look for recommendations for series to watch with friends, or genre filters that will allow people to find recommendations of a certain type.