Web applications of today are dealing with huge amounts of data. Developers need tools to manage those data efficiently and they need be able to present the most important information to users. Modern applications offer personalization of content and some of the applications are also capable of delivering adapted information based on user needs. One of the possibilities for adapting information presented to the user is collaborative filtering. Information is filtered based on the preferences of similar users. In our work we are developing a general adaptive web model. One of our experiments focused on collaborative filtering algorithms and their application in adaptive systems. We present our approach, experiments and results of our work.
Collaborative filtering1 is the process of filtering for information or patterns using
techniques involving collaboration among multiple agents, viewpoints and data
sources. We live in the age of information explosion and we need tools to process the
large amounts of information and offer users the comfort of dealing only with relevant
information. The collaborative filtering techniques proved themselves to be very
useful for this. Many existing applications use filtering techniques to recommend
items such as music [
reasoning layer (Fig. 1). Using artificial intelligence algorithms we can derive some
user characteristics that were not stored in the user model. Similarly, we could add new rules to the multidimensional matrix in the storage layer based on the rules of similar users.
Collaborative filtering is based on the assumption that similar users have similar
preferences [
There are many approaches and algorithms to compute the similarity of user
preferences [
Neighbor (k-NN) algorithm [
and calculate the unknown rating of item i for the user u0 as the arithmetic mean of rating of the K nearest users:
K ∑ Pi (o) P0 (o) = i=1 (2)
K
n
correlation coefficient [
and their standard deviations:
r = covs(XXsY,Y ) = E(( X i − E(sXX )s)Y(Yi − E(Y ))) = n 1−1 ∑i=n1 X is−X X Yi s−Y Y (4) where X and Y are sample means and sx and sy are sample standard deviations.
Similarity can be also calculated using Spearman correlation coefficient [
¬∃i, j : i ≠ j( X i = X j ∨ Yi = Yj ) (5) then ρ is given by: ρ = 1− 6 ∑i=n1 di2 (6) n(n2 −1) where di is the difference between the ranks of corresponding values Xi and Yi and n is the number of values in each data set.
The last algorithm that we used for our experiments uses the Kendall coefficient
[10]. Kendall τ coefficient is defined as:
sim(ua ,ub ) = 1 −
d (ua ,ub )
(3)
(7)
τ =
nc − nd
1 n(n −1)
2
where nc is the number of concordant pairs and nd is the number of discordant pairs.
We performed experiments in the field of music recommendation [
preferences.
user “soustruh“ user ”filipesta”
users has significant influence on the results. Optimalization could be also achieved by changing the k value – number of nearest neighbors.
The algorithms were implemented as a java library. For the experiments we used a graphical user interface. However, the library could be used separately, e.g. in the adaptive web portal backend. This will be the next challenge and we will perform experiments corresponding to the scheme of the GOMAWE that we mentioned in the text earlier.
adaptive web systems. The algorithms computing the recommendations will be part of the reasoning layer of our General Ontological Model for Adaptive Web
algorithms and achieved promising results.
approach allows dealing with adaptation techniques as black box components.
component. An evaluation of adaptation based on the collaborative filtering library in the system that we are developing will be subject of our future work. Our complete work should lead to developing of a universal framework for adaptive web applications.