Complements and substitutes are two typical product relationships that deserve consideration in online product recommendation. One of the key objectives of recommender systems is to promote cross-selling, which heavily relies on recommending the appropriate type of products in specific scenarios. Research on consumer behavior has shown that consumers usually prefer substitutes in the browsing stage whereas complements in the purchasing stage. Thus, it is of great importance to identify the complementary and substitutable relationships between products. In this article, we design a neural network based framework that integrates the textual content and non-textual information of online reviews to mine product relationships. For the textual content, we utilize methods such as LDA topic modeling to represent products in a succinct form called “embedding.” To capture the semantics of complementary and substitutable relationships, we design a modeling process that transfers the product embeddings into semantic features and incorporates additional non-textual factors of product reviews. Extensive experiments are conducted to verify the effectiveness of the proposed product relationship mining model. The advantages and robustness of our model are discussed from various perspectives.

Social media has become very popular nowadays by spreading plentiful human-centric data to a large number of audience. A rich body of literature has studied the influence maximization problem in social networks under certain propagation models. However, existing models suffer from the assumption of unlimited propagation time and some of them are nondeterministic. Influence maximization algorithms within those frameworks are often trading off between performance guarantee and computational cost. In this paper, we try to formulate the influence maximization problem in another way, where we limit the propagation time to a predefined propagation round and in each round maintain the probability to make the propagation process more tractable. We introduce a new diffusion model called a probability propagation model and formulate this optimization problem as a holistic probability maximization problem. We show that information diffusion estimation in the proposed framework is not NP-hard. Hence, any algorithm within it will be more efficient. However, the maximization problem is still NP-hard. After proving the submodularity in the proposed framework, we design a partial-updating greedy algorithm and its heuristic extension to solve the maximization problem. Extensive experiments on four synthetic data sets and four real-world data sets from Facebook, Wikipedia, arXiv, and Epinions demonstrate the effectiveness of the proposed algorithm.