CINERECCS
Project Rundown
The moment we formed our team, all of us were on the same page regarding the nature of our group project. We all wanted to work on something that college students would relate to and use. Among the many things that struck us was the idea of having a source that could suggest us good movies based on our favorite movies as it is hard to find good recommendations that cater to our interests. It was challenging to discuss with the group on how to execute this idea smoothly. With our team’s efforts, we figured out the logistics of this project. After a lot of team meetings and test runs, we found a way to put the diverse talents of our teammates into running this project.
Key Aspects
Ideation
Initial Ideas and Process
Before we landed on this idea, we had hundreds of ideas. Each was better than the other, however, the issue was implementation. We had a brilliant idea that we felt would benefit the FLAME community. However, when we analyzed the work that had to be done, we realized there were issues when we actually sat down to work on it. It was extremely disappointing that we had to let go of that idea but after some research and several debates, we landed on a better idea!
Social Benefits
Cinereccs not only creates perfectly customized recommendations for one person but for all those you want to enjoy a movie with. You no longer have to spend hours deciding and arguing about a film to watch. Cinereccs does it for you. Simply input a movie (or two) that each of you likes and press submit and that's it, a carefully curated list just for you and your friends.
Social Benefits
Working of the Algorithm
Specifics of the Algorithms
Collaborative filtering using matrix factorization is a technique used in recommendation systems to predict a user's preferences based on the preferences of other users. The basic idea is to represent the preferences of users and movies as a matrix, where the rows represent users, the columns represent movies, and the entries of the matrix represent the ratings
that each user has given to each item. However, this matrix is often sparse (mostly filled with empty ratings), as users have only rated a small subset of movies.
Matrix factorization is a mathematical technique that can be used to decompose this sparse matrix into two smaller matrices: one that represents user preferences and another that represents movie properties. These matrices are obtained by minimizing the difference between the predicted ratings and the actual ratings in the original matrix. We used Singular Value Decomposition offered by the scikit-surprise python library to achieve this.
Once the factorization is complete, the user preference matrix and movie property matrix can be multiplied together to predict the ratings that a user would give to a movie that they have not yet rated. These predicted ratings can then be used to make personalized recommendations to the user.
When you input your favorite movies, your user vector is compared to all the existing users using a mathematical technique known as cosine similarity. Based on this, we can find the most similar user in the existing dataset and use it to recommend you movies.
Once we had the model trained in python, we uploaded it to Microsoft Azure Machine Learning Endpoints and used JavaScript API calls to give the model inputs and produce outputs.
Future of the Program
1. In the future we plan on expanding our movie base; we currently have as many as 58k films available on our database, however, globally, around 12.5k films get released every year and we hope to gather as many as possible for our audience. The mission is to venture into niche films and various languages. Finding hidden gems and serving them to our audience is our biggest goal as an organization.
2. In order to cater to our Indian audience, we aspire to focus on movies in the vernacular languages of India, considering the fact that there has been a growth in the viewing of regional movies.
3. Improving customer search convenience is the next aim. Currently, the database requires exact wording in order to match the search. We’d like to make sure that the audience has to type in minimum text (even if grammatically incorrect) to obtain the search they are looking for.
4. The program takes about fifteen seconds to run, ideally the response should be faster so that is definitely something to look forward to improving on.