Computer-based multimedia learning environments support the idea that people learn better and more deeply when appropriate pictures (i.e., animations, video, static graphics) are added to text or narration. There are many adaptive learning systems that adapt learning materials to student properties, preferences, and activities. Adaptive learning environments mostly support only traditional concepts of learning. There is a need to design and develop an e-learning system that embodies principles of constructivist learning approach. The solution is in recommenders systems, which suggest items of interest to users based on their preferences (i.e. previous ratings). If there are no ratings for a certain user or item/object, there is a situation called a cold start problem, which leads to unreliable recommendations. Researchers mostly avoid tackling the absolute cold start in recommender systems.
The topic of presented dissertation is designing a recommender system with a novel approach to avoid cold start problem. Approaches for solving the new user cold start problem can be divided into two main groups: the first group performs additional inquiries to gather more information about the users; and the second group uses dedicated algorithms for users in the cold start state.
The first group of approaches aims at performing additional inquiries about the user. According to this approach, we relate combinations of different learning styles (taking into account four different learning styles models) to preferred multimedia types. We explore a decision model aimed at proposing learning material of an appropriate multimedia type. The study includes 272 student participants. The resulting decision model shows that students prefer well-structured learning texts with colour discrimination, and that the hemispheric learning style model is the most important criterion in deciding student preferences for different multimedia learning materials. To provide a more accurate and reliable model for recommending different multimedia types more learning style models must be combined. Kolb’s classification and the VAK classification allow us to learn if students prefer an active role in the learning process, and what multimedia type they prefer. The results also shows that there is an obvious need to combine learning styles model in order to get a wider view of the student’s characteristics: an approach to problem solving problems, cognitive modes, way of thinking, and a dominant mode of perceiving information. On another hand, model recommends same multimedia material regardless of the learning topic.
In the second part of our research, we have designed and developed a novel approach for alleviating the cold start problem by imputing missing values into the input matrix, thereby improving recommendation performance. Our approach has three steps: 1) finding similar users to given user in cold start state; 2) selecting relevant attributes for the imputation process; 3) aggregate ratings to input matrix for a user in the cold start state. We separate our approach for solving cold start problem into solving absolute cold start problem and solving partial cold start problem. According to the results of our experiments (solving absolute cold start problem), the results indicate that all our proposed methods improve recommending for non-negative matrix factorization with stochastic gradient descent (NG). For semi-non-negative matrix factorization with missing data (SN), combinations FR-ME (imputing attribute's mean value into the attributes that have the highest frequency of the most frequent values) and SD-MF (imputing attribute’s most frequent value into attributes that have the lowest standard deviation) improve recommendations for users in the absolute cold start state. For non-negative matrix factorization with alternating least squares (NS) and matrix factorization by data fusion (DF), none of variations of proposed parameters (methods) improves recommending in absolute cold start state. In the next stage of our research, we evaluated our approach for solving partial cold start problem.
Statistical analysis of experimental evaluation of our approach on the artificial domain showed that each parameter significantly improved recommending of matrix factorization methods. The methods that yield improvements in recommendation accuracy compared with the raw matrix factorization are methods that consider 25 % of similar users ($25$-*-*-*), select an attribute according to the frequency (*-FR-*-*) or RReliefF (*-RR-*-*), and impute a value aggregated by mean value (ME) or predicted by using regression trees (RT). For further investigation we chose two method combinations (25-FR-ME-* and 25-RR-RT-*), which were expected to work well, and compared them with other strategies on real domains. Among all approaches evaluated on the artificial domain, we chose the best performing method with the highest average rank – a method that considers 50 % of similar users, selects an attribute for imputation according to the RReliefF, and imputes a value predicted by linear regression (50-RR-LR-*). All three combinations of the selected methods were evaluated on two real domains: Jester in PEFbase. An evaluation showed that method 25-FR-ME-* combined with matrix factorization NG performed statistically better than the raw matrix factorization algorithms (DF, NG, NS in SN) on real domains for users in the partial cold-start state. The results demonstrated the advantage of using imputation approaches in terms of better recommendation accuracy. At the same time, the results have shown that imputing of missing values has no negative impact for recommending to the users, which are not in the cold start state.