На заседание на семинара “ИНФОРМАТИЧНО МОДЕЛИРАНЕ”,
на 27.04.2018, петък, от 15:00 часа, в зала 578 на ИМИ,
доклад на тема:

“Exploiting Additional Information for Improving the Performance of Collaborative Filtering Recommendations”

ще изнесе проф. Иван Ганчев – ИМИ-БАН, University of Limerick, Limerick, Ireland

Резюме:
This presentation will cover novel approaches for collaborative filtering (CF), as the most popular and widely deployed technique used in recommender systems, by utilizing intrinsic information existing within the user-item rating matrix, or besides it, to build effective and efficient recommendation tools for assisting users in finding high-quality personalized service/item recommendations from massive information resources.

Two lines of research are presented, addressing the main problems CF suffers from, i.e. the scalability, data sparsity, and cold-start. In first, two novel types of CF approaches are suggested and elaborated, which generate recommendations solely based on the user-item rating matrix.

The first one is designed from rating prediction perspective, aiming to predict as accurate as possible the rating value for a user to user’s still unrated items. Incorporation of user representations, learned from user rated items, into a matrix factorization (MF) framework, is proposed thereby generating more accurate recommendations with good scalability and computational efficiency. Two recommendation models are developed within this approach –  UserMF and UserReg –, which extend MF at the rating-model level and regularization level, respectively.
The second type of approach is designed from item-ranking perspective, aiming to find the top-N ranked items that would be most interesting to a particular user and have not been viewed by this user yet. Most of the existing ranking-based prediction approaches consider items as having equal weights which is not always the case. Different weights of items could be regarded as a reflection of items’ importance, or desirability, to users. Thus, it is proposed to integrate variable item weights with the regular ranking-based CF, utilizing only user implicit feedback data. A novel ranking-based recommendation model, WeightedSLIM, is proposed for this approach, with the consideration of various item weights.

The second research line presented is to adopt additional item information to especially address the item cold-start problem. In recent years, additional information of users and items has been observed as being useful to alleviate both the data-sparsity and cold-start issues and to improve recommendation performance. However, in many real-life recommendation scenarios, user information (e.g. demographic and social connections) may not always be available, and users are usually reluctant to have their information accessed by third-party agents, whereas the side information of items is much easier to collect and merchants are willing to offer it. A novel MF model, FeatureMF, is developed and described, which takes into account item features and thus addresses the item cold-start problem. More specifically, item latent vectors are extended with item representation learned from item metadata.

In summary, three new CF approaches are elaborated, along with four novel recommendation models, in order to better deal with the problems of data sparsity and cold start, and to improve the overall recommendation performance in terms of prediction accuracy with good scalability and computational efficiency.

Поканени са всички интересуващи се.

Стоян Порязов