Using video analytics to slove the cold start problem in recommendation systems | |
| Author | Panta, Subigya Jyoti |
| Call Number | AIT Thesis no.CS-18-01 |
| Subject(s) | Recommender systems (Information filtering) Interactive videos Digital video |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Computer Science, School of Engineering and Technology |
| Publisher | Asian Institute of Technology |
| Series Statement | Thesis ; no. CS-18-01 |
| Abstract | In this thesis, I use video analytics to obtain age and gender of a person and use this information to help bootstrap a recommendation system to overcome the well-knkown cold start problem. I perform a comparative study of recommendations produced with age and gender alone versus recommendations produced by collaborative filtering with data collected manually by survey. One hundred and thirty participants took part in the survey, in which they gave ratings to items on scale of 1 to 5. Collaborative filtering is a standard method for recommendations based on similarity of users and items using avariety of models. For the casestudy, I focus on purchase of drinks such as tea,coffee, frappes and smoothies at a coffee shop at AIT. The coffee shop has 41 different menu items; the system predicts users’ ratings for menu items by combining traditional collaborative filtering measurements obtained through survey methods with the demographic data available from video analytics. The top n predicted ratings are given as recommendations to new customers. To evaluate each method, I use an approach I call “Pop and Predict,” in which I remove a known rating from the rating matrix, predict the rating, and then calculate the root mean squared error between predicted ratings and the original ratings. For baseline error, I use root mean squared error obtained by using global and local means and predicted ratings. For comparison, I use root mean squared error obtained by collaborative filtering. Then I use models such as neural networks and SVMs in conjunction with video analytics to obtain the root mean squared error for each model, and I compare the error obtained by each approach. I find that, using only age and gender for prediction produces poor results - it was worse than global mean. However, age and gender combined with information such as time of the day (morning, afternoon, evening, night) and ingredients in the menu item lead to much better results. |
| Year | 2018 |
| Corresponding Series Added Entry | Asian Institute of Technology. Thesis : no. CS-18-01 |
| Type | Thesis |
| School | School of Engineering and Technology (SET) |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Computer Science (CS) |
| Chairperson(s) | Dailey, Matthew N.; |
| Examination Committee(s) | Mongkol Ekpanyapong;Phan Minh Dung; |
| Scholarship Donor(s) | Asian Institute of Technology Fellowship; |
| Degree | Thesis (M. Eng.) - Asian Institute of Technology, 2018 |