Article Preview
Top1. Introduction
Nowadays, a large volume of healthcare related information spread over several websites on the Internet makes it difficult for individuals to locate useful information. There are many healthcare related journals online and good recommendations are always required for researchers. Artificial intelligence, machine learning and natural language processing have been essential for feature selection in user profile learning for helpful recommendations. User profiles are constructed by analysing the healthcare related content referred by users in the past and extracting the prime keywords. These keywords are used to query matching healthcare articles or journal papers or documents to make useful suggestions for the users. In this work, the existing natural language processing techniques have been explored and a framework for content recommendation of healthcare related journals has been proposed.
Content-based recommender systems in general analyse the description of content-based items rated high or/and viewed by a user and constructs a structured user representation of interest based on the features of the items of interest to make recommendations to the user. The features of the user representation are matched with the features of a target content-based item. If the user profile reflects the user interests accurately, it is of immense utility. The user's interests can be accurately matched up with target web pages or journals or other information-based sources.
Content-based Information Filtering (IF) systems have techniques to represent the user profile features and the items and have approaches to compare the user profiles or user feature vector with the target item feature vector or representation. A high-level architecture of a content-based recommender system has been given in Fig. 1 (Lops et al., 2011). It has three components i.e. Content Analyzer is the component which extracts relevant structured information from unstructured data such as text and stores it in the repository. For this, techniques are borrowed from the Information Retrieval domain. It serves as the pre-processing step.
Feature extraction techniques extract the relevant features from the items such as relevant keywords, n-grams, etc into a keywords vector. Profile learner is the module or component that collects data specific to the user interests and generalizes it for user profile construction. The generalization strategy infers a profile of user's preference from items or objects liked/referred or not previously. User profile construction and user profile enhancement are the tasks of this component.
Filtering Component is the component that suggests relevant items to the user by matching the user representation with the target item's feature set. The result is a ranked list of possible candidate items computed by some similar measure or metrics such as Pearson correlation, Euclidean distance measure, Jaccard coefficient measure, cosine-similarity metrics depending upon the suitability (Lops et al., 2011).
Personalized healthcare related content can be recommended in real-time to users which encourage consumption of appropriate content. Recommendations made by the proposed approach can be consumed by researchers and scholars who are doing research or studying healthcare related topics. Medical professionals such as healthcare workers and doctors can also be benefitted by this approach when they want to refer literature about some healthcare related topic and need some useful recommendations.
There have been efforts in past to understand user inclination by term occurrences i.e. term frequencies and inverse document frequencies in the content referred by users. But this approach lacks intelligence and suffers from the problems of natural language ambiguity (Lops et al., 2011). There is scope of work to be done on getting deep into the semantics of terms and looking into the explicit as well as implicit relationship between terms which gives a better insight of user preferences. The context of terms is often ignored which does not give a complete view of user inclination.
Work on unstructured data has various challenges and complexities associated. In the traditional bag-of-words model, the weights of terms are assigned based on their occurrence. But these terms have associations of different kinds and intensity with other terms. The semantics and coupling of terms with other terms need to be analysed in depth to understand the meaning and context of terms.