Recommending Scientific Literatures in a Collaborative Tagging environment 2 Recommending scientific Literatures in a Collaborative Tagging environment This paper proposes a hybrid recommendation algorithm similar to the user-user CF algorithm for a collaborative tagging environment. The difference lies in the user interest modeling, the user similarity computation and the user rating simulation. 2.1 The Representation of User Interest and Literature The user's interest keywords have three sources: the user tags of literatures, keywords of the tagged literatures and their citations. To distinguish the importance of these three sources, different weights are assigned to them respectively. Then, the keywords frequencies are used to form an m-dimension user interest vector as follows Here u, denotes the weighted word frequency of the ith keyword Similarly, the model of a single literature consists of its keywords, keywords of its citations and all users' tags on it D=<d1…,dm> Here d, denotes the relative weighted frequency of the ith keyword summed to one 2.2 The Computation of User Interest Degree The user rating is simulated by user interest degree which is not directly from the user, but measured by similarity between vectors of the user interest and the literature. The formula for interest degree is as follows, where dot-product-based similarity is used instead of cosine similarity since the length of user interest vector is meaningful 2.3 The Computation of User Similarity and Prediction Once the set of most similar users is isolated with the correlation-based similarity [6]. the adjusted weighted sum approach is used to obtain prediction [7] Formally, we can denote the prediction Paas P=R+ (4) Here NSet denotes the nearest neighbor set, sim(u, n)denotes similarity between user u and n. R denotes user ns rate on item i, r denote the average rating of user u.Recommending Scientific Literatures in a Collaborative Tagging Environment 479 2 Recommending Scientific Literatures in a Collaborative Tagging Environment This paper proposes a hybrid recommendation algorithm similar to the user-user CF algorithm for a collaborative tagging environment. The difference lies in the user interest modeling, the user similarity computation and the user rating simulation. 2.1 The Representation of User Interest and Literature The user’s interest keywords have three sources: the user tags of literatures, keywords of the tagged literatures and their citations. To distinguish the importance of these three sources, different weights are assigned to them respectively. Then, the keywords frequencies are used to form an m-dimension user interest vector as follows. 1,..., Uuu =< > m (1) Here i u denotes the weighted word frequency of the ith keyword. Similarly, the model of a single literature consists of its keywords, keywords of its citations and all users’ tags on it. 1,..., Dd d =< > m (2) Here i d denotes the relative weighted frequency of the ith keyword summed to one. 2.2 The Computation of User Interest Degree The user rating is simulated by user interest degree which is not directly from the user, but measured by similarity between vectors of the user interest and the literature. The formula for interest degree is as follows, where dot-product-based similarity is used instead of cosine similarity since the length of user interest vector is meaningful. 1 (,) m i i i R U D ud = = ∑ (3) 2.3 The Computation of User Similarity and Prediction Once the set of most similar users is isolated with the correlation-based similarity [6], the adjusted weighted sum approach is used to obtain prediction [7]. Formally, we can denote the prediction Pui as (,) ( ) (,) ni n n NSet ui u n NSet sim u n R R P R sim u n ∈ ∈ × − = + ∑ ∑ (4) Here NSet denotes the nearest neighbor set, ( , ) sim u n denotes similarity between user u and n. Rni denotes user n’s rate on item i, Ru denote the average rating of user u