Recommender Systems (RS) provide customized suggestion to users for specific item across the bulk of identical data such as media recommendations, electronic commerce web pages & social networks. RSs are being developed using the methods such as Collaborative Filtering (CF) and Contents Based Filtering (CB). However, CF suffers from sparseness problem wherein user-to-item data is sparse and CB filtering depends on feature extraction methods for item descriptions that require knowledge of content semantics and context of RS. In order to deal with the sparsity problem, various matrix factorization techniques embedded with pre-processed auxiliary information are used. On the other hand, currently employed techniques of feature extraction lack in deep semantics of items textual information as they individually cover either the semantic details or topic information. This paper proposes a hybrid RS model called Deep Semantic based Topic driven Hybrid RS (DST-HRS) that employs item description semantics influenced by its topics information. The proposed model extracts the embeddings by capturing the semantics of textual information and incorporates topic details into it. It further integrates these embeddings into Probabilistic Matrix Factorization (PMF), thus efficiently exploiting the semantics of items textual information such as reviews, synopsis, comments, plots etc to overcome the sparseness issue. The proposed DST-HRS can easily be deployed with lesser computation and time complexity. The model is validated on freely available datasets including Amazon Instant Video (AIV) and Movielens (1M & 10M). The validation exhibited a better performance for sparse user-to-item ratings as compared to the state-of-the-art.

推荐系统（RS）向大量相同数据（例如媒体推荐，电子商务网页和社交网络）中的特定项目的用户提供定制建议。使用诸如协同过滤（CF）和基于内容的过滤（CB）之类的方法来开发RS。但是，CF存在稀疏性问题，其中用户到项目的数据稀疏，而CB过滤依赖于需要描述内容语义和RS上下文知识的项目描述的特征提取方法。为了处理稀疏性问题，使用了嵌入有预处理辅助信息的各种矩阵分解技术。另一方面，当前采用的特征提取技术缺乏项目文本信息的深层语义，因为它们分别覆盖了语义细节或主题信息。本文提出了一种混合RS模型，称为基于深度语义的主题驱动混合RS（DST-HRS），该模型采用受其主题信息影响的项目描述语义。所提出的模型通过捕获文本信息的语义来提取嵌入并将主题详细信息纳入其中。它将这些嵌入进一步集成到概率矩阵分解（PMF）中，从而有效地利用了项目文本信息的语义，例如评论，提要，注释，情节等，从而克服了稀疏性问题。提议的DST-HRS可以轻松地以较少的计算和时间复杂性进行部署。该模型已在包括Amazon Instant Video（AIV）和Movielens（1M＆10M）在内的免费数据集中进行了验证。与最新技术相比，该验证对稀疏的用户对项目的评级表现出更好的性能。