INTRODUCTION
WITH the emerging of the Internet era, online social networks such as
Twitter and Face book continue to be popular. People’s communication and
lifestyle have brought about tremendous changes. The generation and
dissemination of hot topics in social media are constantly affecting the daily lives
of people. The social hotspots refer to news or topics that are concerned or
interested by the public at present. The social network topology and the user’s
reads and replies to messages in the network promote the dissemination and
evolution of information related to the hot topic, that is, the propagation of the
network topics [1].
Therefore, mastering user-forwarding participation behavior is
important for evaluating the influence of a microblog topic [2], monitoring public
opinion through networks [3], [4], and information retrieval [5]. At present, the
prediction of user behavior in social networks mainly includes the following two
approaches. The first approach analyzes the structural topology map used for
information dissemination in social networks. This approach predicts the path and
range of the information propagation [6] and, hence, the user’s participation
behavior. Which users will participate in the micro blog is commonly predicted by
dynamic propagation [7] or an infectious disease model [8], [9].
Such predictive models typically classify network nodes as unknowns,
communicators, and immunizers [10], [11]. However, this modeling has two main
shortcomings. First, it creates a complex topology diagram requiring a large
number of calculations. Second, it considers only the relationships of interest
among the users, ignoring the differences among users and the frequent changes of
topics in social networks. The second modeling approach considers user activity,
the number of fans, and the number of messages [3], [12], [13]. Some scholars also
make predictions based on the user’s micro blog interest and micro blog
information [14]–[16]. The influence of social media platforms (such as Weibo)
and the behavior of users are then predicted by machine learning.
The forwarding behavior of online social networks has been extensively
studied in recent years. Focusing on the different aspects of the predicted content,
prediction models using both approaches have been established. However, despite
significant progress in this area of research, there are still some challenges. 1) The
Complexity of the Multi message Interaction: Most studies predict either the micro
participation behavior during single messaging or the macro popularity perception
during multi message topics. These studies ignore the complexity of interactions
among multiple messages under hot topics that occur in actual situations. 2) The
Ambiguity of Multi message Mutual Impact Metrics:
The user participation behavior is closely related to the multi
message interaction under a topic. Traditional micro participation behavior mostly
starts from a single message, generally, only analyzes user attributes or network
topology, and does not accurately measure the interaction of multiple messages. 3)
The Accuracy of the Predicted Model: Traditional mod models cannot correctly
capture the nonlinear relationship between the topic data input and user behavior
prediction output. In addition, ordinary neural networks are usually over fitting and
prone to local minimums, thus reducing the accuracy of predictions. When
predicting user participation behavior, the model should consider the personal
characteristics of users. In addition, the interactions among multiple messages
under the same hot topic are vital for improving the prediction results. Multi
message interaction mechanisms and nonlinear relationships can be handled by a
back propagation (BP) neural network model. The BP neural network is a
multilayer feed forward network trained by an inverse error propagation algorithm.
It can learn and store a large number of input–output mode mapping relationships,
without the need to derive mathematical equations for the relationship in advance.
However, as mentioned earlier, multiple messages exert an iterative guiding effect
on user behavior, which causes over fitting of the neural network.
To avoid the over fitting problem, this article applies a simulated
annealing algorithm to the BP neural network, which assists the local
miniaturization solution of the algorithm and greatly improves the accuracy of the
prediction results. The simulated annealing algorithm is derived from the principle
of solid annealing and has greatly improved the prediction results in many past
instances [17]. The main innovation of this article is that we study the user
behavior of social hotspots from the perspective of multi message interaction at the
micro level. The specific contributions of this article are as follows.
1) A user participation behavior prediction model based on multi
message interaction is constructed. Based on the mapping relationships between
the basic user information and participation behavior under the traditional single
message, the multi message interaction-driving mechanism improves the
completeness of the prediction results. Meanwhile, it is more realistic to describe
the process of message dissemination. 2) A quantization mechanism based on
multi message interaction is proposed. This article can more accurately measure
the multi message selection process within the user community by quantitatively
evaluating the mutual influence of messages from the perspective of topics.
Meanwhile, the hidden influence under the same topic can be qualitatively
measured, which leads to user’s participation behavior. 3) The BP neural network
was improved by the simulated annealing algorithm. This method fits well with the
nonlinear relationship between the topic data input and the user behavior prediction
output. Moreover, the neural network over fitting problem is solved by the
simulated annealing algorithm, and the prediction accuracy is further improved.
This article is organized as follows. This section introduces the background and
status of the research. Section II discusses the work related to our study, and
Section III formalizes the research question. Section IV describes the proposed
method and its learning algorithms. Section V experimentally evaluates our
method oon a real-world data set, and Section VI concludes this study.
