This paper presents an extension of a multifactor recommendation approach based on user tagging with term neighbours. Neighbours of words in tag vectors and documents provide for hitting larger set of documents and not only those matching with direct tag vectors or content of the documents. Tag popularity, tag representativeness and tag similarity are applied similarly as in the original approach but also to neighbours. By doing so, we treat the documents which have been added to the result set by considering word neighbours in the same way as the others. This provides an advantage in the situations where the quality of tags is lower. We discuss the approach on the examples from the existing Medworm system to indicate the usefulness of the approach.
Search systems typically return a ranked list of web pages on di®erent aspects
of the same topic in the returned list in a response to a users request. Recently,
social activities such as tagging have emerged mostly to help people to organize
resources of their personal interest on the web. The tagging information has been
applied to help information retrieval and recommender systems. Although some
successful applications have been developed (see, for instance,[
Tags in Medical bookmarking systems such as M edworm are usually assigned to organize and share resources on the Web. Tag clouds are weighted lists of tags. The relative importance of a tag is visualized with bigger font size, bolder letters, and is measured as a count of the popularity of the tag i.e. how many times users have used it to describe a resource. Tags are generally submitted by any user in bookmarking services such as for example http : ==medworm:com. The data source tags provide useful information with su±cient background and context, even though the set of tags are quite limited to provide an accurate degree of relatedness between tags and neighbours.
Furthermore, the tags themselves as they appear in theM edworn system, for
example, represent multiple domains. Therefore, it is not directly applicable to
consider only tag measures. In our approach, we focus our e®orts on neighbour
objects of the tags for search and retrieval systems. Our approach combines basic
similarity calculus with external factors such as a tag popularity, tag
representativeness and closest neighbours semantic similarity, document score, semantic
similarity of tags as described in [
The rest of the paper is structured as follows. Section 2 discusses the problem and proposed solution on an example. Section 3 de¯nes our approach to multi-factor recommendation with word neighbours. Section 4 discusses related work and positions our work in this context. Section 5 discusses experimental evaluations and outcome results. Section 6 explains analysis of experiment and conclusions .Section 7 discusses future work. 2
Let's consider the following scenario. If the user submits a query to the system, the system evaluates which documents are relevant to the query and returns a rank ordered list of documents to the users. Normally, the system considers content of the documents or collaborative user activities as factors to judge the relevance. Recently, tag-based approaches have been coined in the literature as well but only in the general situations. The domain speci¯c searches by experts usually do not hit the most relevant documents in the ¯rst top n results. For example, the medical M edworm system gives almost 14470 records just based on swine °u tag. However, the system does not know what exactly user is looking for, or user doesn't know the proper words to describe what it is that he wants. Then the returned results are often unsatisfactory.
However, we could resolve this problem by proposed extension of tag and neighbour based recommender systems for medical events. This approach searches for the most trusted information. The traditional approach based on simple tag based recommendation factors are not e±cient enough for domain speci¯c systems such as M edworm due to lower relevance of user generated tags used. Therefore, we also apply neighbours to consider wider set of documents.
Figure 1 shows which kind of information M edworm can o®er to improve the search to ¯nd information with respect to a certain aspect of a document. One just needs to refer to its associated tags and predicted neighbours in the corresponding documents. There are two columns depicted in the ¯gure which represent the positions of neighbours in the text or the tag vector. By considering one side or both sides of the neighbourhood, we can target wider space of documents than with original query. This allows for expansion of the document set hit by the query. Therefore, we ensure that the user will not miss important medical events documented in a document or a blog even when it does not match exactly the query. By applying personalization factors as in original query to extracted neighbours, we achieve similar ranking and therefore provide a means to access the most relevant documents.
The main concept is to achieve recommender system in medical events. The proposed recommender system combines associated neighbours with di®erent aspects of similarity and tags. Consequently, a hybrid recommender system obtained integrate many independent recommendations by applying tag popularity, tag representativeness, tag similarity and neighbours. It exploits neighbours that are dynamically re-calculated according to the e®ectiveness of the recommendation. The system also uses the semantic similarity between the neighbours to calculate neighbours set. These results in°uence the ¯nal recommendation list to re-order the rank. The process of the calculation looks as follows. First, a users click on a tag or he submits a query. Then, the recommendation algorithm is applied to produce a set of recommended resources. Second, this set is then sorted by taking the user interest and tag neighbours into account and re-ranks the results accordingly. 3.2
The tag and neighbour based recommender approach based on [
Tag Similarity (TS). It combines the classical cosine similarity (CosSim) from
user query and information retrieval ¯eld with a semantic similarity (SemSim)
which is de¯ned in [
Cosine similarity (CosSim). The cosine similarity in our approach is a measure between, a user query Q transformed into a tag vectore and a set of tags or words (represented as a vector) of particular web page (W¹ ). Each word, w(ti), in each dimension corresponds to the importance of a particular tag ti.
The Cosine similarity (Q; wi) is calculated for every tag word resource wi 2
W¹ . As an output, this stage of the algorithm will produce a subset of resources
W 0, that have some similarity to the query tag and similarity scores for each
[
Let us assume that the user interacts with the system by selecting a query
tag and expects to receive resource recommendations. Therefore, a query is a
unit vector consisting of a single tag, and the equation is (adapted from [
W¹ (QDi ;WDii ) t2T W¹ (QDi;WDii ) where T is the set of tags, the similarity of the selected tag to each resource and recommends the top n., W¹ is the set of words of that particular visited web page, Di and Dii are a particular documents from a set of web pages/documents. Semantic Similarity (SemSim). The semantic relation between two tags is de¯ned as follows:
SemSim(s; t) = M DSim(s; t) £ OntoSim(s; t); 8s; t 2 T
Where s and t are particular tags of set of tags T . MD Sim (s,t) is the Medical Dictionary similarity score and OntoSim (s,t) is the similarity score achieved from ontologies.
Neighbour Semantic Similarity (NS). Calculates closest neighbour in°uence for
personalized recommendation by vector space models [
The similarity between word w(ti) and w(tj) is sim(w(ti); w(tj)) sim(w(ti); w(tj)) = cos w(t¹i); w(t¹j) = kww(t¹(it¹)ik)¤¤kww(t(¹jt¹)j)k
In this view, the solution to be addressed includes how to represent the tags
and their closest neighbours and how to use it to in°uence the activation of
user preferences. The approach is to predicted neighbours, the current visited
context is represented as (is approximated by) a set of words concepts from the
domain ontology. Ultimately, the perceived e®ect of neighbours extended hybrid
approach is that user interests that are in focus for a current context, and those
that are in the semantic scope of the ongoing user activity are considered for
personalization [
Tags have been recently studied in the context of recommender systems due
to various reasons. Tags are signals or labels that particular resource was
interesting for a user, and he bookmarked it as well as tagged it with a speci¯c
tag relevant for a particular situation a user was encountered in.
Recommendations of relevant events should be based on the su±cient occurrences for similar
signals expressed by tags. Therefore, di®erent similarity measures need to be
studied in this context for e®ectiveness and e±ciency. [
[
Other works such as [
We have conducted an experiment to preliminary assess the performance of the recommender approach proposed in this paper. The nature of the experiment was based on a simulation a mix of scenarios regarding the amount of pages, tags and their neighbours. The proposed scenarios were created aiming at simulating realistic usage of M edworm. The variables addressed by each scenario are: { Amount of Pages: each page has a set of tags that are compared for processing the recommendations. Therefore, the more pages exist then more time will be spent to calculate the similarity between the pages. { Amount of Tags: the similarity of the pages is given by their tags. The whole set of tags of each page must be compared to verify which ones are similar. { Set of neighbours: set of neighbours are depending on similarity of the tags.
The whole set of neighbours of each page must be compared with semantic meaning of the tag.
These variables were chosen because we are using them for calculating the
recommendations. This process is time consuming and invariably a®ects the
system performance. However, it does not mean that other factors such as page
size should not be considered[
We found that the choice of tf ¤ idf played an important role to ¯nd tag representativeness. In our evaluation, tf idf have identical trends, but tf idf always provides superior results, so we have reported only results found based on those weights. We were able to extract tag neighbours. Some samples were taken from each dataset of neighbours to ¯nd neighbours semantic similarity using cosine similarity. The validation was performed to measure the improvement in recommendation. We used MedicineNet1 online free medical dictionary to measure semantic similarity.
Let's assume that given semantic similarity is '(s), where S is the semantic similarity. We also have to de¯ne cosine similarity between the user's query and tags ti, tj, tk, tl, etc. These tags could be in vector node of current webpage: W¹ = f: : : ; ti; : : : ; tj; : : : ; tk; : : : ; tl : : : : : : g, where '(s) µ w¹ and '(s) \ w¹ = ;
The neighbour similarity depends on the similarity of the tag words. So, neighbour semantic similarity (NS) is subset of cosine similarity. NS can be computed as follows: '¹(N S) = f: : : ; ti ¡ 1; ti + 1; : : : ; tj ¡ 1; tj + 1; : : : ; tk ¡ 1; tk + 1; : : : ; tl ¡ 1; tl + 1; : : : : : : g, '¹(N S) µ w¹ and '¹(N S) » '(s):
In order to test the e®ectiveness of the algorithm, we compute the factors with a collection of documents from M edworm data source about 90 pages. The documents were encoded with xml format, so we decided to make 466 tags manually. Content of the pages and some tags were extracted from web sites on the Internet. Similarly, we utilized manually generated neighbours to assign tags to M edworm pages tagged by a particular user. Due to certain constraints,
we had to limit the number of user queries. We needed just adequate number of satisfactorily di®erent pages and su±ciently di®erent assignment of tags to them. Each test case consists of tag, neighbours, semantic factors and resource. We consider this resource as the target results, since we know that the user is interested in it.
Here, we have one user interest query \avian °u pandemic". A simple way to start out is by eliminating documents that do not contain all three tags \avian", \°u", \pandemic", but in general it hits many documents. Our algorithm distinguished relevant and irrelevant documents and tags like \avian", \°u", \pandemic" that occur rarely and good keywords. After performing a recommendation using both the tag and neighbours, the rank of the target resource in the recommendation set was recorded and shown in table1.
ReturnPos Document# Document score(DS) Recommedation Score(HS)
As seen at Table 1 our extended hybrid based algorithm accepts a user interest, a set of neighbours, and a selected tag. The recommendation and document scores are from the interval between 0 and 20. As we can see, the top recommendation score is 14.53 (72.65% accuracy). It means that the particular document is the most relevant to users query. Result positions 2, 3, 4 are also relevant to user query but not most relevant when comparing to position 1. Position 5 document got mixed results and it is partially related to the query. This result was not considered excellent but satisfactory since our recommendations relied basically on syntax similarity of the tags. 6
Categorized M edworm is the application of medical RSS feeds 2 to better target the delivery of health care, facilitate the discovery of new products, and helps to determine a person's predisposition to a particular disease or condition through web. In this recommender systems, the extended hybrid algorithm
can perform tasks such as discovering documents (much like the web robots),
ranking documents, ¯ltering them, and automatically routing useful and
interesting information to users and it has learning and adaptation capabilities. In
fact, the extended hybrid algorithm perfectly suits information discovery and
retrieval in the web. For example, information discovery and ranking can be
handled by document score which depends on the tag popularity and tag
representativeness. Another tag similarity can specialize in indexing, yet another
like neighbours similarity can implement an information retrieval, and so forth.
Applying these techniques to the web pages/documents, retrieved by a search
tool could substantially weed out unrelated documents and improve the ranking
quality of the remaining page/documents [
As per our experiment, the user has tagged several medical object web resources with the tag \°u". If a user selects that tag, the system should recommend resources concerning the number of records about the \°u". Certainly in addition, another \°u" related documents may have been tagged with alternative tags: disease, swine, H1N1, avian, bird °u, in°uenza, °u attacks, respiratory problems, pandemic, symptoms, lung infection, etc. These resources may have been tagged with \°u" and may not have been \°u" but they should still be made available to a user. The recommendation strategies must also be adapted to deal with the neighbours. Typically, recommender systems have dealt with two dimensions: users query and object neighbours semantic similarity. Consider again the \°u" related topics. After selecting \°u" the system may recommend resources only related by semantic similarities of neighbours. User may notice as of his query \avian °u pandemic", the \avian", \°u", and \pandemic" are related tags. These are generated by the closest neighbour semantic objects.
Finally, user may notice resources in Medworm's pro¯le dealing with the medical neighbour objects, and view one of those resources as his priority is most trusted information in the top position. 7
The evaluation of the system provided some supplementary conclusions, namely, a recommendation performed with association neighbours appeared to be the most useful but only positive in°uence neighbours. Future work will focus on positive and negative neighbours and tag similarity i.e. sentiment analysis. It would bene¯t the system to utilize such opinions and to lower the score of bad results, even if other strategies show them as recommendable.
This work is partially supported by the European Union under the IST project M-Eco (http://www.meco-project.eu).