PROJECT TITLE :

Item Concept Network Towards Concept-based Item Representation Learning

ABSTRACT:

Leveraging textual information is a common approach taken when item concept modeling is being done. However, the vast majority of currently available models do not make use of the inferential property of concepts in order to determine the meanings of words. As a result, these models ignore the relatedness that exists between correlated concepts, a phenomenon that we refer to as conceptual "correlation sparsity." In this paper, we differentiate between word modeling and concept modeling, and we propose a framework for item concept modeling that centers on the item concept network (ICN). ICN models and further enriches item concepts by leveraging the inferential property of concepts. This addresses the correlation sparsity issue, which was previously mentioned. To be more specific, the proposed framework consists of two stages: the first is the construction of the ICN, and the second is the embedding of learning. In the first step, we propose a method for the generalized construction of networks in order to build ICN, which is a structured network that deduces expanded concepts for items through the use of matrix operations. In the second stage, item and concept embeddings are learned through the utilization of neighborhood proximity. The resulting embedding, which is made possible by the ICN that was proposed, makes it easier to perform homogeneous and heterogeneous tasks such as item-to-item and concept-to-item retrieval, and it provides related results that are more diverse than those that are delivered by traditional keyword-matching-based methods. The framework encodes useful conceptual information, and as a result, it performs better than traditional methods in a variety of item classification and retrieval tasks, as shown by our experiments on two real-world datasets.