With the World Wide Web moving from passive to active, the role of recommender systems as an aid to make decisions play a very prominent role. This enables its users to nd new items of high personal interest, which they were previously unaware of. While traditional approaches have shown the generation of high quality recommendations, the additional use of background knowledge to describe the items and their preferences on a more granular level is still lacking. Furthermore, these approaches do not take into consideration the contextual information, wherein the dimension 'time' plays a signi cant role. In this paper, we propose a new approach for recommending movies, which semantically enriches the process of generating recommendations by using a taxonomy derived out of di erent data sources from the LOD-Cloud. Furthermore, the paper also addresses the interplay between the rating behavior of the users and the dimension 'time'.
The amount of information in the world wide web grows continuously, causing users to be overwhelmed by the sheer volume of data. Users need a mechanism which aid them in their decisions when choosing the most useful item. Recommender systems collect their actions and then infer on their preferences. The system then generates an internal user representation, the user pro le. With a search space reduction based his preferences, items of low personal interest will gradually be removed from the displayed items, improving his search results and simplifying the navigation.
Adomavicus et. al. [
Recommendations based on traditional algorithms don't consider any
contextual information like date, place, companion and mood. Considering the di erent
contextual aspects, the dimension 'time' can be identi ed as the most important
one [
The proposed algorithm follows the "collaboration via content" approach. User pro les contain the preferred attributes, derived from the item representations. They are further extended by using a attribute-taxonomy, which assigns the attributes to appropriate super-class relations, hence allowing to regard connections between di erent but related attributes. An exponential dampening factor is applied, to re ect the generality of attributes nearer to the root. Following the assumption, that old ratings have a lesser in uence on the preferences than newer ones, they are weighted less by applying a decay rate.
The approach is formally described as follows: Let U be the set of users, I the set of items and R the systems rating scale as an totally ordered set of values. Then ut : U I ! R is the utility function, that calculates the usefulness of a single item i to the user u. In order to create a representation A(i) of an item i 2 I, its direct attributes D(i) := faj jaj 2 T g are determined and weighted by the function w(i; aj ), which speci es the signi cance of an attribute and follows the T F IDF metric. Beside the direct attributes aj , the representation contains their indirect attributes sad(aj ). This hierarchy of attributes forms the attribute-taxonomy T . Based on its structure, the path pa(aj ) = (sa0(aj ) sad(aj ) ; ra) for the attribute sa0(aj ) aj to the root ra is determined. Each attribute is weighted according to its distance d from aj by the height-function h(sa(aj )) = d , whereby the hierarchy in uence parameter > 0 adjusts the in uence of indirect attributes. To re ect the di erent weighting schema of direct and indirect attributes, the weighting function w(i; aj ) takes possible multiple occurrences of the indirect attributes into account. The representation A(i) of an element i is then de ned as: A(i) := f(aj ; w(i; aj ))g [
[ aj2D(i)
f(saj ; w(i; saj ))jsaj 2 pa(aj )g with w(i; a) = 8 1 < jA(i)j
P : aj2D(i)^a2pa(aj) log n(a) jIj
if a is direct attribute h(a) w(i; aj ) otherwise with n(aj ) = is the number of items, containing the attribute aj (1) (2) Based on the item representations A(i) of the users rated items I(u), his pro le p(u) := f(aj ; pr(u; aj ))jaj 2 A(i) ^ i 2 I(u)g can be constructed. It contains the items attributes aj and their preference weights pr(u; aj ; t), which show the degree of interest on it at the point in time t. Pro les are generated in an iterative way. Starting with the oldest rated item A(i)0, it is lled with attributes and their weights. Only those items are taken into account, which are rated higher than the mean of the systems rating scale. In order to re ect the preference changes over time, the time decay factor 0 1 is applied to the contained attributes when a new item A(i)t) is rated at a later point at time t. pr(u; aj ; t) = (1 (1 ) pr(u; aj ; t ) pr(u; aj ; t 1) + 1) w(i; j) if aj 2 A(i)t otherwise (4)
By applying the Pearson correlation between the user pro les, the similarity sim(u; u0) of two users u and u0 can be calculated. Hereby are the preference weights pr(u; a) of the common attributes ca(u; u0) := f j a a 2 p(u) \ p(u0)g the used variables for the correlation. The prediction ut(u; i) for a single item i is de ned as the weighted sum of the k most similar users u0, which have rated i, their similarities sim(u; u0) to u and their ratings ru;i on i. By using ut(u; i), the recommendation set RS(u; N ) for a user u can be generated. It contains the N unseen items, which has the highest predicted rating. (5) (6) (7)
P a2ca(u;u0)
P a2ca(u;u0) P (pr(u; a) pr(u)) (pr(u0; a)
pr(u0)) (pr(u; a) pr(u))2 r
P a2ca(u;u0) (pr(u0; a)
pr(u0))2 ut(u; ij+1g sim(u; u0) = r
sim(u; u0) ru0;i ut(u; i) = u02N(u;k)
P ksim(u; u0)k u02N(u;k)
RS(u; N ) := fiijii 2 InI(u) ^ ut(u; ij ) 3
The MovieLens project provides a dataset, consisting of 3.500 users, who rated 6.000 movies with 1.000.000 ratings and was enriched with URIs, which identify the movies in DBpedia. A top-down approach starting with the DBPedia URI for Category:Film and going down the tree via skos:broader of until it reached the categories assigned to the movies was applied. The created taxonomy T consisted of 3.804 direct attributes a of the movies and furthermore 691 indirect attributes sa while having a height h of 12.
Other knowledge bases such as Freebase extract information form multiple data sources. One interesting aspect of Freebase is the way it categorizes lms into genres by the property lm. lm genre. The lm.genre taxonomy contains a total of 700 genres which are related to each other via child genre relations. In order to use the freebase lm genres in our approach we, rst had to extract all the genres and their relations, and then proceeded to infer a genre taxonomy where the genres are related to each other via subclass relations. This taxonomy contains 700 direct attributes a assigned to the movies and 238 indirect attributes sa which are contained inside the taxonomy T with height h of 4.
The DBpedia taxonomy contains more distinct attributes due to the higher number of lm related categories in comparison to the genres in Freebase. Furthermore the Freebase taxonomy is rather shallow in comparison to the DBpedia taxonomy. The number of distinct indirect attributes represents the nodes that are not present in the direct item description but are generated due to taking into account the superclasses of those categories/genres. For example for movie that has cyberpunk as a genre we would also take sci- into account as a genre due to it being the superclass of cyberpunk. 4
To evaluate the system, a 5-fold cross validation was performed with each set
containing 1.200 users. Each of the 5 evaluation runs use 4 di erent training sets
Ut and the remaining set Ue is the test set. The system was rst evaluated to nd
the optimal values for k, and , using the F1@N metric [
As baseline algorithm, the widely used item-to-item collaborative ltering
algorithm using the adjusted cosine similarity [
0:26 50
100 neighborhood size k 0:2 0:4 0:6
0:8 time decay
Di erent neighbourhood sizes k have a signi cant in uence on the accuracy. Figure 1 shows the in uence of k. Thereby achieved the di erent approaches their highest accuracy at k 120. Smaller values for k result into a degradation of the accuracy. If the neighbourhood is to small, the few contained users have a strong impact on the predictions. Larger neighbourhoods don't have an in uence on the accuracy. For the taxonomic approach was the value of k higher, because more users with shared general preferences can be considered, hence achieving a higher accuracy. 0:31 0:2 0:4 0:6
0:8 hierarchie in uence Fig. 3. Impact of hierarchy in uence on the accuracy of the recommendations 0:32 0:3 0:26
0:32 0:28 0:29
0:26
According to the height h of the used taxonomy, has to be adjusted accordingly. Figure 3 shows its impact on the accuracy of the system. While to small and to large values lowers the accuracy, as optimal value could = 0:6 be determined. When is chosen to low, the additional information gain by the taxonomy has a negligible e ect and the system behaves similar to the concept only recommender. The taxonomy enables to nd users, which has related preferences in movie genres, e.g. in dystophic sci- and in cyberpunk. But if the value for is set to large, the concepts nearer the root gain a higher in uence resulting in a loss of precision in the preference representation of the users. Thereby users are treated equally, even if they are only loosely connected.
Figure 4 shows the accuracy according of the used F1@5 metric. Each algorithm used the parameters for k, alpha and , where it achieved its highest accuracy. Since both tononomix approaches behaved similar according to the time in uence, its value was set to = 0:7 and for the attribute only recommender = 0:1. For the taxonomy using system was = 0:6, and k = 120.
The paper proposed a new approach for semantically enriching the process of recommending items by using a taxonomy derived out of the LOD-cloud. It outperforms the baseline algorithm to a signi cant level. In addition to this, there was a positive in uence of the taxonomy on the accuracy of recommendation. As the DBpedia-recommender uses the same concepts as the attribute-recommender and follows the same paradigm, its accuracy is signi cantly increased by the use of the taxonomy. In a nutshell, the proposed approach seems to be well suited to work with the structure, given by the DBpedia-category taxonomy.
Following to our assumption, the degradation of each users' preference over his usage period has an positive impact on the accuracy. But some preferences tend to exist in present, even if they rst were captured at the beginning. Therefore, the degradation of preferences has to be considered individually for each user and each of his preference. A deeper analysis on the in uence of the aspect 'time' on the proposed approach falls out of scope of this paper and is addressed independently in the forthcoming papers.