نگرش پیشنهادی اعتماد-معنایی و مبتنی بر ترکیب برای برنامه های کاربردی کسب و کار الکترونیکی

Collaborative Filtering (CF) is the most popular recommendation technique but still suffers from data sparsity, user and item cold-start problems, resulting in poor recommendation accuracy and reduced coverage. This study incorporates additional information from the users' social trust network and the items' semantic domain knowledge to alleviate these problems. It proposes an innovative Trust–Semantic Fusion (TSF)-based recommendation approach within the CF framework. Experiments demonstrate that the TSF approach significantly outperforms existing recommendation algorithms in terms of recommendation accuracy and coverage when dealing with the above problems. A business-to-business recommender system case study validates the applicability of the TSF approach.

Recommender systems are considered the most popular forms of web personalization and have become a promising and important research topic in information sciences and decision support systems [9], [10], [17], [19], [20], [24], [46] and [49]. Recommender systems are used to either predict whether a particular user will like a particular item or to identify a set of k items that will be of interest to a certain user, and have been used in different web-based applications including e-business, e-learning and e-tourism [8], [22] and [31]. Currently, Collaborative Filtering (CF) is probably the most known and commonly used recommendation approach in recommender systems. CF works by collecting user ratings for items in a given domain and computing similarities between users or between items in order to produce recommendations [1] and [31]. CF can be further divided into user-based and item-based CF approaches. In user-based CF approach, a user will receive recommendations of items that similar users liked. In item-based CF approach, a user will receive recommendations of items that are similar to the ones that the user liked in the past [1]. Despite their popularity and success, the CF-based approaches still suffer from some major limitations; these include data sparsity, cold-start user and cold-start item problems [1], [3], [36] and [37]. The data sparsity problem occurs when the number of available items increases and the number of ratings in the rating matrix is insufficient for generating accurate predictions. When the ratings obtained are very small compared to the number of ratings that are needed to be predicted, a recommender system becomes unable to locate similar neighbors and produces poor recommendations. The cold-start (CS) user problem, which is also known as the new user problem, affects users who have none, or a small number of ratings. When the number of rated items is small for the CS user, the CF-based approaches cannot properly find the user neighbors using rating similarity, so it fails to generate accurate recommendations. The CS item problem, which is also known as the new item problem, affects items that have none, or only a small number of ratings. With few ratings for CS items, CF-based approaches cannot appropriately locate similar neighbors using rating similarity and would be unlikely to recommend them [1], [33], [36] and [37]. In view of these limitations, researchers have commonly decided to opt for trust-based [11], [16], [26], [27] and [44] and semantic-based [2], [15], [22], [35] and [45] recommender systems to tackle such limitations. These systems can deal with the trust relations between users and semantic features of items, which cannot be well handled in traditional CF-based recommendation approaches, to support the recommendation process. These systems have proved to be successful in solving some limitations of CF-based approaches by allowing the recommender systems to make inferences based on an additional source of knowledge. We believe that, by considering information extracted from the users' trust network and the items' semantic domain knowledge, a fusion-based recommendation approach that takes into account both trust and semantic information should provide more effective recommendations. Based on this notion and following our previous work [22], [23], [38], [39], [40] and [41] where we addressed some limitations of CF-based recommendation approaches, this paper proposes a fusion-based recommendation approach that fuses the trust and semantic information of users and items within the CF framework to achieve yet more effective results in terms of recommendation accuracy and coverage, especially when dealing with data sparsity, CS user and CS item problems. The proposed approach, called TSF (Trust Semantic Fusion), fuses two hybrid recommendation approaches; the user-based trust-enhanced CF, and the item-based semantic-enhanced CF. The user-based trust-enhanced CF approach utilizes the intuitive properties of trust and trust propagation to address the data sparsity and CS user problems. The item-based semantic-enhanced CF approach employs the underlying semantic relationships between items to help reduce the effect of data sparsity and CS item problems. We also define and introduce the notion of an item's reputation weight into the item-based semantic-enhanced CF to further improve the quality of predictions. This paper is organized as follows. In Section 2, research background and related work are described. Section 3 presents the components of the TSF approach. A case-based mathematical example for illustrating the procedure of the TSF is given in Section 4. Section 5 demonstrates the experimental evaluation and results using MovieLens and Yahoo! Webscope datasets. Section 6 describes a case study to validate the feasibility of applying the TSF approach into real e-business applications. Finally, the contributions of this study are summarized, and future research is presented in Section 7.

This paper proposes the TSF recommendation approach, which provides a far higher quality of recommendations in terms of recommendation accuracy and coverage, compared to the benchmark trust, semantic and CF-based recommendation algorithms. The TSF approach fuses the user-based trust-enhanced CF and the item-based semantic-enhanced CF approaches. The user-based trust-enhanced CF approach utilizes the intuitive properties of trust and trust propagation to address the sparsity and CS user problems. The item-based semantic-enhanced CF approach employs the underlying semantic relationships between items to reduce the effect of sparsity and CS item problems. The experimental results verify the effectiveness of the TSF approach, by significantly achieving better recommendation accuracy and more coverage when dealing with data sparsity, CS users and CS items compared to the benchmark recommendation algorithms. Also, the results of the validation performed, using as a case study a B2B recommender system, allows us to conclude that the proposed TSF approach is a feasible and effective method for building recommender systems in real e-business applications. In future work, we plan to: (1) study and evaluate the impact of using different Uninorm functions to aggregate trust on the recommendation quality of the user-based trust-enhanced CF recommendation approach; (2) study and evaluate the impact of different fusion strategies on the recommendation quality of the fusion-based recommendation approach; (3) design and develop an efficient method for updating and rebuilding the implicit trust social network. Once the implicit trust social network has been built, it will be difficult to instantly reflect new information for user preferences. Updating the implicit trust social network has not regularly been considered because of the expensive computational time required for this process. Accordingly, an efficient method of updating and rebuilding the implicit trust social network is required; and, (4) further extend our ‘BizSeeker’ system to incorporate the use of the proposed fusion-based recommendation approach in real-world practice for e-business applications.