This paper presents an approach of recommending a ranked list of books to a user. A user pro le is de ned by a few liked and disliked books. To recommend a book, we calculate semantic relatedness of the given book to the liked and disliked books by using Wikipedia. Based on the obtained scores, we predict ratings of the book. We evaluate our approach on a dataset that consists of 6,181 users, 8,171 books and 67,990 user-item pairs to predict the rating.
Wikipedia provides a valuable source of background knowledge about millions of entities such as movies, actors, places and books. This knowledge can be exploited to build the Top-N recommendation system which deals with nding a set of N items that best match a user pro le. The user pro le can be de ned by liked and disliked items. We assume that user might like the items that are similar or related to his/her liked ones. Therefore, the recommendation task can be re-modeled as to nd out a ranked list of items which are more related to the liked items than the disliked ones.
In recent years, there have been several e orts in utilizing external knowledge
bases such as Wikipedia and DBpedia 1 for recommendation. Most of the work
focuses on boosting collaborative ltering approach or to improve content-based
systems [
User pro le is de ned by a few liked and disliked books. The task is to nd out a ranked list of N other books that a user might like. We compute the relatedness scores of a given book with the liked and disliked books. We recommend the book only if the score for like prediction is greater than dislike prediction. Since user can like or dislike more than one book, we need to aggregate the relatedness scores to obtain nal con dence scores for like and dislike predictions. Therefore, we use three methods to aggregate the relatedness scores.
Average: We calculate relatedness scores of the given book with all the books liked by the user, and obtained a con dence score by taking an average of these scores. Similarly, we calculate the con dence score for disliked items. Maximum: We calculate relatedness scores of the given book with all the books liked by the user. Unlike to the Average case, we choose the relatedness score of the most related pair as the con dence score. Similarly, we calculate the con dence score for disliked items.
Random: We randomly select one book from all the books liked by the user, and one from all the disliked ones. We calculate relatedness scores of the given book with the randomly selected liked and disliked books. The obtained relatedness scores are considered as the con dence scores for the corresponding classes. 2.2
DiSER generates a high dimensional vector by taking every Wikipedia article as a dimension, and considers the associativity weight of an entity with the article as the magnitude of the corresponding dimension. To measure the semantic relatedness between two entities, it computes the cosine score between their corresponding DiSER vectors.
DiSER retrieves a list of relevant Wikipedia articles and rank them according to their relevance scores with the given entity. It considers only the human annotated entities in Wikipedia, thus keeping only the canonical entities that appear with hyperlinks in Wikipedia articles. The tf-idf weight of an entity with every Wikipedia article is calculated and used to build a semantic vector. The semantic vector of an entity is represented by the retrieved Wikipedia concepts sorted by their tf-idf scores. For instance, there is an entity e, DiSER builds a semantic vector v, where v = PiN=0 ai ci and ci is ith concept in the Wikipedia concept space, and ai is the tf-idf weight of the entity e with the concept ci. Here, N represents the total number of Wikipedia articles. 3 3.1
In order to evaluate our approach, we perform experiments on a dataset provided in \Linked Open Data-enabled Recommender Systems"2 challenge. There were three di erent tasks, where task 2 was \Top-N recommendation from binary user feedback". The dataset consists of 6,181 users, 8,171 books and 67,990 user-item pairs to predict the rating. All the books contain their corresponding DBpedia and Wikipedia links.
Entity Relatedness
ESA Context-VSM
DiSER Average
ESA Maximum Context-VSM
DiSER
Precision Recall
F1
We performed experiments with three di erent relatedness measures: ESA,
ContextVSM and DiSER. Similar to DiSER, ESA computes the relatedness score by
taking distance between two high dimensional vectors built over Wikipedia.
However, unlike DiSER, it does not perform any speci c feature selection for
entity relatedness. Thus, it considers only the surface form of an entity and do
not di erentiate between two entities with the same surface forms. For instance,
ESA builds the same vector for \Harry Potter ( lm series)" and \Harry
Potter (book series)" as it generates the vector for their surface form i.e. \Harry
Potter". We compute ESA score between the book titles. Context-VSM is
similar to KORE [
We presented our approach of Top N books recommendation. We reported the results of three di erent relatedness measures. DiSER outperformed other two methods of computing relatedness scores. Further, we showed that random aggregation achieves comparable scores. Thus, we can conclude that Wikipedia is a valuable resource for obtaining the recommendation of popular books in a coldstart scenario. Future work will include the investigation of other relatedness measures to boost the conventional recommendation methods like collaborative ltering.