I am currently preparing for submitting a poster application for Grace Hopper 2015; the submission is due 3/20.
I started by writing an introduction, which I initially based off of our CREU project proposal which Dr. Bansal wrote. I began with broad context information about Big Data and focused in on semantic technology. It was at this point that I realized how narrow my project has become; while my project indeed falls within the realm of Big Data and Semantic Technology, there is so much more to it than that. In building videogame suggestion software, I am diving into the huge field of recommender systems, of which I have no knowledge. I immediately began researching this field to better introduce my project, and the result of that research is the subject of this post.
Recommender systems became a field in response to a huge amount of entertainment content. Consumers simply didn't know where to start. I haven't done that much research into the beginnings of this field, but I do know that it is a large one and definitely still has some open problems. I found this paper (accessible when logged in to IEEE Xplore via ASU libraries) which has been cited thousands of times. It describes the current state of the field of recommender systems in 2005, including current issues and possible extensions. This was a great source for me, as it gave me a good introduction to the topic while maintaining technical rigor. The paper classifies recommender systems into three categories:
It also includes some very specific problems with recommender systems such as:
Now, from my previous work researching the current state of videogame recommender systems, it is clear that my domain is not one in which much effort has been directed towards creating accurate videogame suggestions. After all, it is a relatively new domain. Also, the phenomenon of women playing videogames in such high numbers is also pretty new. Thus, it makes sense (to me, at least) that this problem hasn't been identified by many other people. I tried to determine how to describe what the problem is in more academic terms so that I could search the relevant literature, and I decided upon "stereotyping of demographic segments." I think that is an accurate phrase for the problem I have identified. However, my review of literature has not turned anything up. In my opinion, this is a subset of the overspecialization problem, which recommends items which are too similar to previously chosen items. In this case, the recommender systems are classifying a user based on their demographic segment and recommending games that are similar to others in that demographic segment. I posit that this produces inaccurate recommendations.
I would like to solve this problem using semantic technology. I decided that if my particular domain isn't addressed in the literature, I might be able to find some ways that semantic technology has addressed other issues with recommender systems. I found this paper, this one, and this one (again, accessible through IEEE Xplore). They describe how the use of ontologies aided in creating a better picture of a user because the user data was multidimensional. As the author of the initial paper stated, "...most of the recommendation methods produce ratings that are based on a limited understanding of users and items as captured by user and item profiles...". I feel that my proposal, of obtaining user metadata other than demographic data, will develop a better picture of a user and thus better recommendations. I will finish my poster application in the next week and will post the finished product here.
I started by writing an introduction, which I initially based off of our CREU project proposal which Dr. Bansal wrote. I began with broad context information about Big Data and focused in on semantic technology. It was at this point that I realized how narrow my project has become; while my project indeed falls within the realm of Big Data and Semantic Technology, there is so much more to it than that. In building videogame suggestion software, I am diving into the huge field of recommender systems, of which I have no knowledge. I immediately began researching this field to better introduce my project, and the result of that research is the subject of this post.
Recommender systems became a field in response to a huge amount of entertainment content. Consumers simply didn't know where to start. I haven't done that much research into the beginnings of this field, but I do know that it is a large one and definitely still has some open problems. I found this paper (accessible when logged in to IEEE Xplore via ASU libraries) which has been cited thousands of times. It describes the current state of the field of recommender systems in 2005, including current issues and possible extensions. This was a great source for me, as it gave me a good introduction to the topic while maintaining technical rigor. The paper classifies recommender systems into three categories:
- Content-based recommendations: the user will be recommended items similar to the ones the user preferred in the past
- Collaborative recommendations: the user will be recommended items that people with similar tastes and preferences liked in the past
- Hybrid approaches: a combination of the above two methods
It also includes some very specific problems with recommender systems such as:
- overspecialization, when the system recommends items too similar to those already recommended or rated
- new user problem, the user has to rate a sufficient number of items before a content-based recommender system can provide reliable recommendations
- new item problem, until a new item is rated by a substantial amount of users the recommender system will not recommend it
- sparsity, the number of ratings already obtained is usually very small compared to the number of ratings that need to be predicted
- intrusiveness, the recommender system requires explicit feedback from the user and often at a significant level of user involvement
- flexibility, the user cannot customize recommendations
- other issues which were named but not described, such as explainability, trustworthiness, scalability, and privacy
Now, from my previous work researching the current state of videogame recommender systems, it is clear that my domain is not one in which much effort has been directed towards creating accurate videogame suggestions. After all, it is a relatively new domain. Also, the phenomenon of women playing videogames in such high numbers is also pretty new. Thus, it makes sense (to me, at least) that this problem hasn't been identified by many other people. I tried to determine how to describe what the problem is in more academic terms so that I could search the relevant literature, and I decided upon "stereotyping of demographic segments." I think that is an accurate phrase for the problem I have identified. However, my review of literature has not turned anything up. In my opinion, this is a subset of the overspecialization problem, which recommends items which are too similar to previously chosen items. In this case, the recommender systems are classifying a user based on their demographic segment and recommending games that are similar to others in that demographic segment. I posit that this produces inaccurate recommendations.
I would like to solve this problem using semantic technology. I decided that if my particular domain isn't addressed in the literature, I might be able to find some ways that semantic technology has addressed other issues with recommender systems. I found this paper, this one, and this one (again, accessible through IEEE Xplore). They describe how the use of ontologies aided in creating a better picture of a user because the user data was multidimensional. As the author of the initial paper stated, "...most of the recommendation methods produce ratings that are based on a limited understanding of users and items as captured by user and item profiles...". I feel that my proposal, of obtaining user metadata other than demographic data, will develop a better picture of a user and thus better recommendations. I will finish my poster application in the next week and will post the finished product here.