홈으로ArticlesAll Issue
ArticlesA Topic-Rank Recommendation Model Based on Microblog Topic Relevance & User Preference Analysis
  • Fuguang Bao1,2,3, Wenqian Xu2, Yao Feng2, and Chonghuan Xu1,3,4,5,*

Human-centric Computing and Information Sciences volume 12, Article number: 10 (2022)
Cite this article 2 Accesses
https://doi.org/10.22967/HCIS.2022.12.010

Abstract

There are many well-known social networking platforms in China, such as Sina Weibo, WeChat, and Tik Tok. With the explosive growth of the amount of information on social networking platforms, it becomes difficult for users to find the efficient information they need. Thus, extracting the high-quality content to provide users with personalized recommendations to meet users’ needs is quite significant research for a microblogging platform. This paper proposes a topic-rank recommendation model, which is based on user topic relevance and user preference. Firstly, we construct a topic screening model using the term frequency-inverse document frequency (TFIDF) and latent Dirichlet allocation (LDA) methods. Secondly, we present a model of user influence. The user’s topical influence is obtained by detailed calculation of the direct and indirect influences on the user. Finally, we build a topic-rank recommendation model through comprehensive consideration of topic relevance and user preference. In this regard, experimental results show that the topic-rank recommendation model has a wilder coverage and higher accuracy compared with a traditional personalized recommendation model, while the research results will provide certain references to online platforms, enterprises, and governments.


Keywords

Topic-Rank, Recommendation Model, Sina Weibo, Topic Relevance, User Preference


Introduction

With the rapid development of social media, increasingly more users choose to communicate and share opinions on social networking platforms, while the emergence of new social networking platforms breaks the situation where traditional portal websites master most information and enables everyone to become the creator and disseminator of information. Meanwhile, the rapid development of platforms such as Sina Weibo and Twitter have expanded information dissemination channels and accelerated the speed of information dissemination among people. As an information publishing platform based on user relationships, Sina Weibo has functions such as reposting, commenting, and following. Its large number of active users and constantly changing themes make it the fastest and most influential social networking platform for public discourse and opinion sharing today. However, with the proliferation of internet users and information, users face massive amounts of messages, so knowing how to efficiently and precisely find the information they need is becoming increasingly more difficult. Accordingly, extracting high-quality content to provide users with personalized recommendations and improving user retrieval efficiency is one of the key issues that the Sina Weibo platform needs to study. The main reasons for the above problems are information overload, information redundancy, and that users cannot obtain other information except from friends. Amidst the wave of information explosion, users are keen to immediately find users, topics, and popular or trending events that they are interested in from the internet, and filter out information that is useless without them. As an important part of Sina Weibo, a recommendation system helps users save time and reduce the problem of information overload to a certain extent. In addition, in this virtual network, various recommendation models are widely used in various fields, and its research is highly significance.
Existing documents lack research on recommendation systems from the perspective of topic evolution and user information such as background information and user preference. Moreover, the coverage rate of existing research is not high enough, while the effect of personalized recommendation is not accurate enough, resulting in the inability to meet the needs of people amid the wave of information explosion. Therefore, in order to improve the efficiency of people’s search and satisfaction with personalized recommendations, we have proposed a topic-rank recommendation model, using recommendation scores instead of clustering to improve accuracy. We have associated users’ likes, comments, reposts, and published content on Sina Weibo with similarities between other users in the node network, and use term frequency-inverse document frequency (TFIDF) to measure user preference. Finally, the recommendation system is obtained by comprehensively considering the theme influence and user preference.
The main contributions are summarized as follows:

We propose an improved topic recommendation system based on topic selection, topic ranking, and user preference.

We introduce the latent Dirichlet allocation (LDA) model to calculate user similarity and propose an improved model to measure the characteristics of topic focus, activity, and authenticity.

The thematic impact is subdivided into two types as follows: direct impact and indirect impact. Then, we use an improved susceptible-infectious-removed (SIR) model to calculate indirect influences, specifically influencing fans through strong and weak connections, resulting in more accurate and credible result.

In particular, we consider both topic influence and user preference and weight them in the process of generating recommendation results.


Related Work

The existing researches can be divided into two subjects as follows: personalized topic recommendation based on topic evolution and personalized topic recommendation based on user information.

Research on personalized topic recommendation based on topic evolution
Zhou and Yang [1] have used the co-word analysis model to analyze the hotspots of topics. Bao et al. [2] has proposed a KNN-SMOTE-LSTM algorithm to study the nearest neighbors between nodes, which could calculate the similarities among nodes. Wang and Xia [3] have proposed a content recommendation model based on social tag mining, whereas Li et al. [4] have proposed a comprehensive recommendation model, which used the LDA topic model and matrix decomposition to find the user’s interest orientation and achieve a personalized recommendation. In terms of content recommendation, He et al. [5] have proposed an association rule-LDA topic model content recommendation model to recommend specific topics to users, whereas, Dai et al. [6] have proposed an online topic model based on popular factors, incorporating popular factors of topic content discovery and evolution by considering the topic’s timeliness and accuracy. However, it ignored the number of readers and other popular factors.

Research on personalized topic recommendation based on user information
Zhong et al. [7] have combined text semantics with user background for microblogging information recommendation. Zeng et al. [8] have proposed a sorting algorithm based on the improved Apriori recommendation friends based on the user’s forwarding and commenting behavior information. Whereas, fixating on the sparseness of user location and user attribute information data, Xin and Wu [9] have proposed a recommendation system that mainly used the topic model on machine learning to extract the user's hidden topics. Then, Guo et al. [10] have proposed the strength-MF algorithm to model the trust relationship strength and user interest, while Xiao et al. [11] have introduced a trust mechanism to calculate the trust weight between users.

Summary
It can be examined from the existing recommendation systems and researches that although the topic recommendation model has made preliminary progress, its coverage rate is relatively low. Therefore, this paper proposes a topic recommendation system that integrates the topic influence and user preference, and we consider two different objects, namely the topic and user each.


A Model based on Topic Influence & User Preference

For social network services, the model of providing users with recommendations can be measured with multiple metrics. In the traditional recommendation system research, various models have been proposed to the process of capturing data onto articles and recommended models for users. Fig. 1 shows the overall process framework of our model as follows.

Fig. 1. Flowchart of user topic influence scoring model


Construction of Topic Selection Model
Microblogging topic crawler model
As most of the data on Sina Weibo is open, and there is no strict permission setting, the crawler technology can be better implemented. Since users in Sina Weibo have various interests and hobbies, we use crawler technology to screen users on specific topics based on topic relevance. Accordingly, the theme crawler first locates by the theme, obtains the character node related to it, and puts it into the node pool. Then, the theme crawler takes out the character node from the node pool based on the corresponding crawling model and uses the model of data collection to obtain the basic information about the character node and relationship information (including follow information, fan information), and then pre-process and store the related data until the node pool is empty. Fig. 2 shows the microblogging crawler flowchart.
In the process of data cleaning, it is necessary to carry out noise removal processing, and through manual browsing and screening models, the collected data is processed with irrelevant information such as URLs, duplicate data, garbled codes, etc., to obtain the value data of analysis. Based on this, the above data is processed further as follows. Firstly, remove the special symbols of the data, including double quotation marks (""), dots (.), substitution signs (~), spaces, line feed symbols, and other characters that are not useful for article research. Then, use the R language to construct a corpus based on the word segmentation of the custom dictionary, and delete the numbers, letters, stop words, etc. in the corpus. Finally, we get the corpus that needs to be built.

Fig. 2. Flowchart of microblogging topic crawler.


Construction of word document matrix
In order to construct a screening model for a certain topic, a term weighting model-inverse document frequency value TFIDF is used to calculate the word frequency of each word [12] after transforming the above-built corpus into a word document matrix. The following steps are to cluster the LDA topic model based on the TFIDF value of each word, calculate the keywords according to the word vector extracted by TFIDF, and use the LDA topic model to calculate the probability topic. After formatting the words of each document into the form of vectors, a word document matrix can be constructed. The columns in the word document matrix represent the set of words that appear in multiple documents in the corpus, and the rows represent the documents in the corpus.
TFIDF is a weighting technique which is commonly used in information retrieval or data mining [13]. In the field of text mining, the TFIDF value is often used to evaluate the importance of a word or word to a document in a document set or corpus. Accordingly, if this term often appears in a particular document, but it rarely appears in the entire document set detected, and thus, its TFIDF value for this particular document will be high [12]. As the frequency of various words with the document increases, the importance of these words will also increase.
In TFIDF, TF represents the word frequency, referring to the frequency of words appearing in the document. The formula for calculating TF of word $w_i$ appears in document $d_j$ is as follows:

$TF_{w_i d_j} = \frac{I(w_i,d_j)}{\displaystyle\sum_{i} I(w_i,d_j)}$(1)

In the above formula, w_i represents the i-th word after the word segmentation, $d_j$ represents the $j$ document, and $i$ represent the frequency count. IDF represents the reverse document frequency, which is calculated as follows:

$IDF_{w_i}=log⁡ \left(\frac{|D|}{(1+|{j:w_i∈d_j}|}\right)$(2)

The meaning of this calculation formula expresses the total number of documents $|D|$ divided by the sum of the number of documents containing the word $w_i$ plus 1, followed by the logarithmic operation. The term frequency represents the frequency of the words, and the reverse document frequency represents the prevalence of the words. The concept of TFIDF is that if a word appears more frequently in a document but less frequently in other documents, then this word is a feature word. Based on this, for the words in document $d_j$ in $w_i$, TFIDF is calculated as follows:

$TFIDF_{w_i d_j} = TF_{w_i d_j}*IDF_{w_i d_j}$(3)

The word document matrix DTM is composed of a list of word column vectors, and the elements of each word vector represent the TFIDF value of the corresponding word of each document. According to the above steps, we fill in the results obtained in the corresponding position of the word document matrix. The formula is expressed as follows:

$DTM_{i,j} = TFIDF_{w_i d_j}$(4)

In the above formula, $DTM_{i,j}$ represents the element located in the $i$ row and $j$ column of the word document matrix DTM.
Considering that the number of documents in the corpus is too large, it may happen that a word that appears in a document does not appear in other documents, and the TFIDF value of the word in some documents is 0. Therefore, we get the initial DTM matrix as a coefficient matrix with extremely high sparseness. In order to simplify the calculation, we reduce the sparsity of the word document matrix DTM by deleting the words with a smaller total TFIDF value.
The total TFIDF value of each column of words is calculated as follows:

$TFIDF_j = \displaystyle\sum_{i} DTM_{i,j}$(5)

Among them, $TFIDF_j$ represents the TFIDF value of the $j$ word, and $DTM_{i,j}$ represents the TDFIDF value of the $𝑖$ row and $j$ column in the word document matrix. The following steps are to arrange the TFIDF of $j$ words according to the size, to keep the words with the first 90% quantile, then reconstruct the word document moment DTM, and finally get the word document matrix after reducing the sparsity.

LDA topic model
The LDA topic model is also called the probabilistic topic model. It represents each text corpus (called a document set) through generating word segmentation with a specific probability and then determines the topic according to the word segmentation distribution [14]. The LDA topic model extracts the topic of an article based on the semantic features of words. It is a three-layer Bayesian structure and consists of two parts, namely “document-topic” and “topic-participle,” which can make the same topic words come together [15]. Using the above keywords extracted by the TFIDF model, the LDA topic model is used to segment and cluster the word vectors, and the topic of the document set is inferred with a certain probability (the topic is not unique).
The LDA model believes that each document is a random mixture of various topics according to a certain ratio, and the mixing ratio adheres to multiple distributions. According to the former research [16], the topic of a document can be represented as follows:

$Z|θ=Multionomial(θ)$(6)

Each topic is formed by mixing the words of the vocabulary according to a certain ratio, and the mixing ratio adheres to multiple distributions, denoted as follows:

$W|Z,∅=Multionomial(∅)$(7)

The probability generated $d_j$ under the comment $ω_i$ condition is expressed as follows:

$P(ω_i │d_j) = \displaystyle\sum_{s=1}^{K} P(ω_i|z=s)×P(z=s|d_j)$(8)

$P(ω_i|z=s)$ represents the probability that $ω_i$ belongs to the $s-th$ topic, and $P(z=s|d_j)$ represents the probability of $s-th$ topic in text $d_j$.

The Establishment of User Topic Influence Model
On the Weibo platform, there will be some users with ulterior motives or certain commercial agendas. Their activities will lead to the production of two abnormal users of zombie fans and navy forces. Therefore, we should eliminate the zombie fans and navy forces in Sina Weibo before calculating the user influence. Afterward, for the low topic focus rate, it would lead to an inaccurate calculation of the user's topic influence. Based on this, we need to not only calculate the direct influence of the users based on their post, but also calculate the users’ indirect influence based on their fan activity influence.

Eliminate zombie fans
Liu and Yang [17] have proposed a recent user activity calculation, but while this paper proposes different views. Not all taboo likes are zombie fans, and it is unreasonable to measure zombie fans only by the number of Sina Weibo users. Accordingly, we believe that it is necessary to conduct a comprehensive evaluation of the user’s recent activities and to consider a variety of factors that may affect the activity of Weibo users, including factors such as the kind of likes, reposts, comments, check-in status, and following users, etc. The formula for $user_m$ activity is as follows:

$A_m = \displaystyle\sum_{r=1}^6 γ_{mr} HY_{mr}$(9)

$Avg_A = \frac{\displaystyle\sum_{s=1}^{S} A_m}{S}$(10)

$A_m$ is the activity of $user_m; γ_{mr}$ is the weight of $user_m$ in the r-th influencing factor; $HY_{mr}$ is the number of actions of $user_m$ in the r-th influencing factor. $Avg_A$ is the overall activity of microblogging, and S is the total number of users. Accordingly, deleting users characterized as $A_m <<Avg_A$, and the zombie fans to filter users can make data more effective.

Eliminate cyber navy
The cyber navy refers to the employed cyber writers who publish specific information about specific content on the internet, while pretending to be ordinary netizens or consumers and influencing other ordinary users by publishing, replying, and disseminating blog posts. Since the emergence of the cyber navy has greatly affected the identification quality of user influence and disrupted the order for the internet, it is important to accurately discover and eliminate the cyber navy. It’s noteworthy that as opposed to zombie fans, the cyber is very similar to ordinary users in that they will post, follow, and like articles, seriously affecting the analysis of user influence. We find that the cyber navy pays a high level of attention to users, while ordinary users rarely return visits, which is a distinct characteristic of the cyber navy. The calculation formula for the re-inspection rate is as follows:

$R_m = \frac{follow_m}{FOLLOW_m}$(11)

$Avg_R = \frac{\displaystyle\sum_{g=1}^{G} R_m}{G}$(12)

$R_m$ is the return rate of $user_m$, while $follow_m$ is the number of return users of $user_m$. $FOLLOW_m$ is the number of all followed users of $user_m$, and $Avg_R$ is the overall average return rate of Sina Weibo. By deleting the users who meet the condition that $R_m << Avg_R$, the navy will be deleted successfully.

Topic-Rank Recommendation Model
In this paper, we propose a user influence recommendation model that uses recommendation scores rather than clustering to improve the accuracy. We associate the user’s likes, comments, reposts, and releases with the similarities between other users in the node network, and at the same time define the TFIDF value of the related users as the degree of preference. Finally, the theme influence and preference degree are weighted and comprehensively considered to form a recommendation system.

Direct influence of user topic
In the internet environment, direct influence is a summary of the influence of users who can directly have contact, such as the user’s own friends and fans [18, 19]. If a user’s Sina Weibo that’s published on a topic is reposted, commented, and liked by friends or fans in a large number, the user's direct influence on the topic will be very large. The formula for calculating the user's activity on each topic is as follows:

$Ac_{mk}=\displaystyle\sum_{l=1}^{L}(\frac{wp_{mkl}}{WP_k} + \frac{wd_{mkl}}{WD_k} + \frac{wz_{mkl}}{WZ_k}$(13)

$Ac_{mk}$ represents the activity degree of microblog $user_m$ on the topic $k$, and $k$ is the total number of tweets posted by users on the topic $k$. $wp_{mkl}$ represents the number of microblogging comments posted by microblog $user_m$ on the topic $k$, while $WP_k$ represents the number of all microblog users’ comments on a microblog in the whole dataset of topic k. Then, $wd_{mkl}$ is the number of thumb up microblogs published by microblog $user_m$ on the topic k, and $WD_k$ represents the number of thumb up microblogs of all microblog users in the whole dataset of topic k. Then, $wz_{mkl}$ is the number of microblog posts forwarded by microblog $user_m$ on the topic k, and $WZ_k$ represents the number of microblogs forwarded by all microblog users in the whole dataset of the topic k.
After studying the topic activity of Sina Weibo users, we also need to study the quality of users’ Sina Weibo. The quality of Sina Weibo is reflected as to whether the content of Sina Weibo is positive & authentic and whether it has a certain positive impact on the entire society. When selecting Sina Weibo content, poor quality content should be eliminated and content that has a positive impact on society should be retained. Therefore, in the process of screening Sina Weibo data, this paper selects data such as the kinds of comments, reposts, likes, and favorites on Weibo. If a Weibo is commented, reposted, liked, and favorited by many people, it naturally indicates that the quality of the Weibo posted is high. According to the above ideas, the microblog quality calculation is performed on the data for the topics collected by $user_m$ on the topic screening model.

$Qua_{mk} = \displaystyle\sum_{l=1}^{L} \displaystyle\sum_{h=1}^{4} \frac{si_{mlhk}}{SI_{lhk}} $(14)

$Qua_mk$ is the microblog quality of $user_m$ on the topic, and L represents the total number of microblogs related to the topic k published by $user_m$ which is screened out according to the LDA theme model. While h refers to the four kinds of operational behaviors, namely publishing a microblog, forwarding a microblog, microblog commenting, and microblog thumbs up, $si_{mlhk}$ represents the number of tweets of $user_m$ in the h operation of the topic k. Then, $SI_{lhk}$ represents the number of tweets of type h operation of all users in topic k.
Microblog users also have different categories, such as a variety of plus V users (VIP users). These specifically refer to the microblog authenticated users, which is divided into orange personal real-name authentication (actors, singers, writers, etc.), blue enterprise real-name authentication (a company, studio, etc.) and general users. Users with different authentication situations will affect the user’s influence to a certain extent due to the different fan bases. Based on this, in order to solve this problem, this paper adds $user_m$’s credit rating $Cre_m$.

$Cre_m = \begin{cases} 1,Blue plus V authenticated users \cr 0.5,Yellow plus V authenticated users \cr 0,Ordinary unauthenticated user) \end{cases}$(15)

Pulsing the direct influence of microblogging on the original basis, $user_m$ on users of the k-th topic $F_m$ is calculated as follows:

$F_{mk}=Ac_{mk}+Qua_{mk}+Cre_m$(16)

$Ac_{mk}$ represents the activity of microblog $user_m$ on the topic k, and $Qua_{mk}$ is the quality of microblog $user_m$ on the topic $k$ and $Cre_m$ is the credit degree of $user_m$.

Indirect influence of user topic
In the network society, many users can operate on the Sina Weibo posted by $user_m$ even if they are not friends or fans of $user_m$, who have no direct contact with $user_m$. With this, the behavior is only through others reposting $user_m$’s Weibo. Based on this situation, indirect interaction between users can be considered. Thus, this paper defines the indirect influence of $user_m$ as the influence on the fans and friends of $user_a$ who is a friend or a fan of $user_m$ in the social network, which refers to the influence of $user_a$ when his or her posted Weibo is forwarded, commented, and liked. It’s noteworthy that due to the similarities between information dissemination and virus dissemination, the SIR model has been widely used in the field of information dissemination in recent years [20]. At first, all nodes are in the susceptible state S, and then some of the nodes are exposed to the information and become the infection state I. Finally, the nodes that are already in the infected state I or the nodes that are still in the susceptible state S become recovery statues R due to the saturation of message transmission or disinterest. Since people related to $user_m$ will have their friends, fans, and passers-by, it is not accurate enough to only use the direct communication with the SIR model. Accordingly, we add the strong and weak relationships of the infection state on this basis.
After improvement, the number of people affected per unit time is calculated as follows:

$\begin{cases} \frac{dS(t)}{dt}=-λ_1 (t) I_1 (t)-λ_0 (t) I_0 (t) \cr \frac{dI_0 (t)}{dt}=λ_0 S(t) I_0 (t)-β_0 I_0 (t) \cr \frac{dI_1 (t)}{dt}=λ_1 S(t) I_1 (t)-β_1 I_1 (t) \cr \frac{dR(t)}{dt}=β_0 I_0 (t)+β_1 I_1 (t) )┤ \end{cases}$(17)

where $S(t)+R(t)+I_0 (t)+I_1 (t)=1$.
Relative to all nodes, the indirect influence of users is calculated as follows:

$QF_{mk}=\frac{I_{0_{mk}}+I_{1_{mk}}+R_{mk}}{P}$(18)

$I_{0_{mk}}$ represents the number of the topic k of the fans of $user_m$ that spread into the weak relationship node, and $I_{1_{mk}}$ is the number of the topic k of the fans of $user_m$ that spread into the strong relationship node. Then, $R_{mk}$ represents the number of the topic k of the fans of $user_m$ that get the message but no longer spread messages, while P is the total number of nodes.

Topic influence of user
By combining the obtained direct influence of $user_m$ on the k-th topic with the indirect influence of $user_m$ on the k-th topic, the user-topic influence of $user_m$ on the k-th topic can be obtained, which is calculated as follows:

$Q_{mk}=QF_{mk}+F_{mk}$(19)

$Qf_{mk}$ represents the indirect influence of $user_m$ on the topic $k$, while $F_{mk}$ is the direct influence of $user_m$ on the topic k.

Recommendation system of user topic influence
In terms of Sina Weibo posting, likes, comments, and forwarding, we compare the similarity of $user_m$ in the node with other users in the topic network node. Firstly, we convert the content of each operable behavior of $user_m$ into a text vector and calculate the TFIDF value, using the segments of the text vector. Then, we calculate the average of the obtained TFIDF, and finally get the similarity between $user_m$ and other users in the network node. Subsequently, we can calculate the user’s preference based on the degree of relevance [21] and then integrate the preference degree and user influence in order to obtain the user recommendation system and rank recommended users. Fig. 3 shows the flowchart of the user topic influence recommendation system as follows.

Fig. 3. Flowchart of a user topic influence recommendation system.


According to formulas (1), (2), (3), (4), we calculate the TFIDF value of each text vector segmentation of $user_m$, which is summarized as follows:

$TFIDF_{m_{i}}=\frac{\displaystyle\sum_{j=1}^{J}TFIDF_{m_{i_j}}}{J}$(20)

$TFIDF_m=\frac{\displaystyle\sum_{i=1}^{I}TFIDF_{m_i}}{I}$(21)

where $j$ represents the number of word segments of each text vector, and $i$ is the number of text vectors in the text vector set, while the obtained $user_m$ is the value of the user's preference for other users. To ensure the reliability of the data preprocessing conclusions, we also carry out a normalized calculation analysis on the user topic direction.

$recommend_m = ω*TFIDF_m+(1-ω)*Q_{mk}$(22)

Among them, $ω$ represents a factor that adjusts the user's topic influence and preference similarity. If the preference similarity is too large, the impact factor can be increased; otherwise, the preference similarity factor can be increased. After getting the recommended scores, they will be expanded by 10 times to form the recommended scores.


Experiment and Analysis

Data Sources and Metric
Due to the massive number of Weibo users, it is unrealistic to calculate all the data. Therefore, in order to save time and improve efficiency on the basis of reliable data, this paper randomly selects nearly 2 months of data for research on the Sina Weibo page. We also mainly conduct multiple experiments on Weibo content (if there are videos or pictures to selectively crawl photos or videos at the top of the document), using topic crawlers to randomly intercept Weibo user data. As a result, total of 3,478,638 Weibo posts and 74,321 users were obtained.
In experimental measurement, scholars usually use a variety of metrics to verify the performance of the recommended model. Accordingly, we choose four typical metrics of accuracy, recall, F-score and root mean square error (RMSE) to evaluate the performance of the algorithm.

Precision specifies the fraction of related instances among the instances retrieved in pattern recognition, information retrieval, etc. In recommendation measurement, we define precision as the ratio of recommended items selected by the users to the total number of recommended items. The formula for Precision can be expressed as c/L, where c represents the number of recommended items selected by the users in the test set, and L represents the total number of recommended items.

Recall specifies the fraction of the total number of relevant instances actually retrieved. In recommendation measurement, we define recall as the ratio of recommended items selected by the users to the total number of items actually selected by users. The formula for Recall can be expressed as c/M, where c represents the number of recommended items selected by users in the test set, and M is the total number of recommended items actually selected by users.

F-score specifies a comprehensive indicator that runs the weighted harmonic average of Precision and Recall, while reflecting the balance between Precision and Recall. The F-score formula can be expressed as:

$F-score = \frac{2∙(Precision∙Recall)}{Precision+Recall}$(23)

RMSE specifies the square root of the ratio of the square of the deviation between the observed value and actual value to the number of observations. Due to the limited number of observations in the actual measurement, the most reliable value (the optimal value) must be used instead of the true value. In the process of evaluating personalized recommendations, RMSE calculates all the values of the square root of the mean of the sum of squares of deviations between the true value and predicted value. The smaller the value gets, the higher a recommender’s prediction accuracy will be. The formula is

$RMSE=\sqrt{\frac{1}{T} \displaystyle\sum_{i=1}^{T} (RE_i-RE_i')^2}$(24)

where $RE_i$ stands for real value, while $RE_i'$ stands for predicted value and $T$ stands for times of observation.
We use the coverage rate as a metric to evaluate the result accuracy of different models. User coverage indicates the range of users’ influence on network nodes and indicates the spread range of users’ influence. This paper defines the user coverage rate as the ratio of the number of users interacting with users in a time period for the number of all active users in that time period. The user coverage rate set in a certain time period [$t_0,t_1$] can be expressed as:

$Coverage Rate= \frac{|user_m ∩ user_k|}{|user|}$(25)

Among them, $user_m ∩ user_k$ represents the collection of users who have interacted with the user, while $user_m$ represents the collection of users interacting with the user, and $user_k$ represents the collection of users interacting with the user. Then, the user indicates the set of active users in that time period, and the active user indicates the users who are interacting or being interactively operated.

Data Preprocessing
We get a corpus containing cleaned text data, which contains a total number of 358,641 documents. The document entry matrix constructed by the TFIDF value of documents and words converts text data into data in the form of a two-dimensional table, (that is to say, the words are converted into variables).
As shown in Table 1, for pre-optimization, there are a total number of 28,247,249 non-sparse entry documents and a total number of 38,728,957,327 sparse entry documents, showing the sparsity of 0.999271. After optimization, there are a total number of 1,932,821 non-sparse entry documents and a total number of 13,578,392 sparse entry documents, showing a sparsity of 0.857655. After optimization, the sparsity decreased by about 0.14 compared to pre-optimization.

Table 1. Post- and pre-optimization comparison table
Before After
Document Non-sparse/Sparse entries Sparse degree Document Non-sparse/Sparse entries Sparse degree
358,641 28,247,249/38,728,957,327 0.999271 358,641 1,932,821/13,578,392 0.857655
Accordingly, it seems that due to the large number of documents and words in the corpus, there also are a large number of zero matrix entry documents, (in other words, very sparse matrix entry documents). As a result, it can optimize the calculation by reducing the storage of 0 constant matrix data entry files, which is conducive to more accurate analysis.
The value calculated according to the TFIDF algorithm corresponds to a total number of 358,641 documents (that is to say, each text data is regarded as a document). By constructing a document term matrix, each matrix element of the matrix represents a word with the document corresponding to each TFIDF value. Each row represents a document (a data extraction), and each column represents a different word of the corpus. Table 2 shows part of the document entry matrix.
As an example, it takes five words such as feeling, taste, find, time, and food in the same corpus. Table 2 shows that in document 1, the proportion of feeling is 0.059899, the proportion of find is 0.020189, and the proportion of taste & hour & food are all 0. Accordingly, in document 1, removing the 0 matrix entry documents, it can result in the keywords feeling and find. Table 2 shows the remaining nine documents analogy.

Table 2. Keyword extraction in different documents
  Feeling Taste Find Hour Food
1 0.059899 0 0.020189 0 0
2 0 0.019732 0 0.0097772 0
3 0 0 0 0.021937 0.048091
4 0.014544 0 0.02224 0 0
5 0 0 0 0.022076 0
6 0 0 0 0 0
7 0 0 0 0 0.010203
8 0 0 0 0.024645 0.022793
9 0 0 0 0 0
10 0 0 0.03587 0 0
Then, we use models to extract keywords from all TFIDF corpus, evaluate the importance of each word, and find the keywords. By extracting TFIDF from Sina Weibo content, we use the keyword LDA topic model to cluster the keyword text and give the topic of the Weibo content through quantitative analysis. After experimental analysis, we get 10 related themes from the results such as food collection, travel, popular TV series, soul chicken soup, beauty life, car brand service, animal world, life encyclopedia, game situation, cartoon home, and so on. Subsequently, we judge the weight based on these keywords and recommend it to different users, and thereby take gourmet collection as an example to verify the feasibility of the user recommendation system.

Fig. 4. The coverage condition of the value of $ω$.


Parameter Settings
According to similar research proposed by Liu and Yang [17], we have predicted the optimal values of $β_0, β_1, λ_0$ and $λ_1$ as 0.1, 0.1, 0.5, 0.1. Meanwhile, according to similar research proposed by Zhuang et al. [20], we have predicted the optimal values of $γ_{mr$} as $\frac{1}{6}. Before calculating the recommendation score, we first need to calculate the weight $ω$. In relation to this, to determine the optimal value of parameter $ω$, we perform the iterative calculations. Based on this, we take the coverage rate as the objective function, for which the larger the coverage is, the better the performance will be. In the iterative calculation, we set a discrete range on a scale of 0 to 1 and the interval between each parameter to 0.1.
As shown in Fig. 4, when $ω < 0.7$, the coverage rate increases with an increase of $ω$, and reaches a peak at $ω=0.7$. When $ω > 0.7$, the coverage rate gradually decreases with the increase of $ω$ until $ω=1$. The chart below shows that the coverage rate is optimal when $ω=0.7$, thus we choose $ω=0.7$ for our calculation.

Result Analysis
User influence analysis based on topic-rank model
The Topic-rank model is used to get the ranking of users’ influence on food collection topics. Through this, we find that the top users are all Weibo VIP users, with the top 12 user information shown in Table 3. Among them, we can see that the number of fans is gradually decreasing according to the username, and the corresponding topic influence is basically related to the number of fans. For example, the number of followers of the username “Little Prince Click” reaches a total number of 1.1 million, and his topic influence is 495.9317687. Then, the number of fans with the username “Call me King Pomelo” is 996,000, and his topic influence is 470.1998826, while “Fan Ma small kitchen” has a total number of 302,000 fans, and his topic influence is 416.0112037. In other words, topic influence increases as the number of fans decreases. Additionally, we can also find that there is no obvious relationship between the number of user-related microblogs and the number of fans and topic influence. In other words, the number of fans is large, and the number of microblogs posted by users with a high topic influence can be more or less. For example, the number of fans of user “Little Prince Click” and user “Learn to cook with me” is close to the topic influence, while the number of Sina Weibo of user “Little Prince Click” is 1,435, but the number of Sina Weibo of user “Learn to cook with me” is 9,353. Therefore, the number of Sina Weibo in the latter is far greater than that in the former. Meanwhile, the number of Sina Weibo accounts for user “New entertainment sauce” and user “Fan Ma small kitchen” approach more than 4,000. However, the former has a total number of 1.05 million followers and related topic influence is 453.7786144, while the latter only has a total number of 302,000 followers and related topic influence is 416.0112037. Accordingly, although the two have similar numbers of Sina Weibo, the number of fans and topic influence are quite different.

Table 3. User influence on the topic of food collection
User name Related microblogs  Fans(t) Topic influence
Little Prince Click 1,435 1,100 495.9317687
Learn to cook with me 9,353 1,200 493.1696467
Universe is invincible food sky ball  3,108 194 481.8455426
Is the gas 1,481 1,060 478.9801616
The meat search 466 1,610 471.3904158
Call me King Pomelo 1,723 996 470.1998826
Little nervous bug 1,122 875 468.643223
The tip of the tongue bot 14,962 677 466.4002235
Big red hey 1,794 2,080 462.976748
The Queen’s Sweet House 25,421 475 456.6121148
New entertainment sauce 4,511 1,050 453.7786144
Fan Ma small kitchen 4,601 302 416.0112037
Subsequently, we analyze the reasons for this phenomenon. Due to reasons such as the time and place of their topic release, some users may cause Sina Weibo to become popular with its high quality despite the small number of its topics. Then, this attracts a lot of fans and the topic's influence is relatively high as well. Despite the large number of related microblogs, some people do not meet the expectations and needs of netizens, so their attention level is low and their influence is small. Additionally, due to the long-term persistence of publishing Sina Weibo for some people, the quality of their Sina Weibo has changed from quantitative to qualitative, resulting in gradually accumulating many fans as well as gradually increasing their topical influence.

Recommendation analysis based on topic-rank recommendation model
We have randomly selected three users from the network nodes of food collection and used the proposed topic-rank recommendation model to calculate the network node recommendation scores of these three users (not including the users themselves). The higher the score is, the greater the correlation will be. When users conduct microblogging activities, the recommendation system can select and recommend users with high relevance. Table 4 shows the top five situations where each randomly selected user has the highest recommendation score. Accordingly, this recommendation ensures not only that the same types of themes are known by different people, but also that different types of themes have the opportunity to be promoted to different people, thereby increasing the diversity and activity of the platform.

Table 4. Recommend scores results
User’s name Related user Recommend scores
_ Pear vortex _An LL_MOMOKO_ 8.9
A cute freak 8.9
I am not sour, I am sweet 7.8
Nectarine balls 7.7
Concubine eat not fat _ 7.2
Very milky, very salty, very nice Maiden Alice 9
Linyi Cuisine Probe Shop Bo 8.7
Milk Girl, 8.5
Remember the solar eclipse 8.4
Li Ziran 8.1
MissPluto-Charon Cha Qiuqiu 8
Amanda's little kitchen 7.8
Food to stick 7.8
Barbeque sauce 7.7
Maiden Alice 7.7
Comparative analysis of topic-rank model and personalized page-rank model
To evaluate the quality of topic-rank and personalized page-rank more effectively, we have chosen to use the coverage rate to judge. Then, to test the strength of user influence, we have conducted a comparative test and compared the calculation results of the topic-rank model and the personalized page-rank model on the food collection topic covered by the user’s influence, with the result presented in Fig. 5. In the 0–50 stage, the topic-rank model and personalized page-rank model show a monotonously increasing state. However, the coverage rate of the topic-rank model is significantly higher than that of the personalized Page-rank model. After 100–120, the influence ranges of the topic-rank model and personalized Page-rank model begin to stabilize, and the influence ranges gradually approach each other. The intersection of the two models at 130 indicates that the two models achieve the same effect at this point in time. Additionally, the experiment has found that the Topic-rank model has a greater coverage rate than the personalized page-rank model and a higher number of users, although the top 100 users of the two models have a topical influence of more than 90%.

Fig. 5. Comparison of user influence.


Comparative analysis of topic-rank recommendation model with some recommendation models
To evaluate the performance of different recommendation models, we compare the proposed topic-rank recommendation model with the friend information service recommendation model [22] based on community division and user similarity, as well as the microblog recommendation model [23] that combines user clustering and the Bandits model. Table 5 below shows the results (the recommended number is 10, 20, and 30).
We set the recommended number to 10, 20, and 30 and evaluate the three models, respectively. Table 5 shows that when the number of recommendations is 10, the precision and recall of the model are 2.2% and 0.6%, which are higher than those of the friend information service recommendation model and microblog recommendation model, respectively. As the number of recommendations increases, the precision and recall results of the three models gradually approach together. The F-score metric shows that the model’s effect is 2.4% and 0.6%, which are higher than the friend information service recommendation model and microblog recommendation model, respectively. For the RMSE metric, the model is 44% and 4%, which are lower than the friend information service recommendation model and microblog recommendation model, respectively. Therefore, it can be examined that when the number of recommendations is relatively small, the precision and recall results of the proposed topic-rank recommendation model have outstanding advantages.
Therefore, the proposed model is better than the friend information service recommendation model and microblog recommendation model. The reason for this is that the proposed recommendation model considers the combination of user influence and preference rather than user similarity alone.

Table 5. Performance comparison of the three models in the Sina Weibo dataset
Model Precision Recall F-score RMSE
L=10 Zhang and Kai [22] 0.012551414 0.044275507 0.019558343 0.570864499
He et al. [23] 0.028513294 0.054841279 0.037519369 0.168369206
Topic-rank recommendation 0.034612044 0.062055076 0.04443813 0.1277696
L=20 Zhang and Kai [22] 0.011016445 0.04291194 0.01753203 0.674698452
He et al. [23] 0.02013688 0.048128011 0.028393746 0.288822064
Topic-rank recommendation 0.02461475 0.051786164 0.033368802 0.213158706
L=30 Zhang and Kai [22] 0.007762016 0.038899835 0.01294167 0.73105
He et al. [23] 0.010485052 0.042826037 0.016845772 0.371263
Topic-rank recommendation 0.014525385 0.046109792 0.022091548 0.248934


Applicability and Limitations
This paper has used topic analysis and user preference analysis to make topic recommendations for users. In essence, the recommendation system simulates a certain behavior of individuals, then analyzes and processes the data that needs to be determined with the recommendation model, and thereby recommends the processed results of users with related needs.
The applicability of the model proposed in this paper is as follows:
1) Topics on social networking platforms like Sina Weibo, WeChat, and Red Booklet also incorporate factors such as the kind of likes, favorites, comments, reposts, etc. The text is also processed, and the recommended objects are selected for recommendation. Accordingly, the topic recommendation system proposed in this paper can be used on various social networking platforms to provide references to major businesses.
2) Knowledge sharing platforms such as Zhihu, Douban, and Brief also warrant a large amount of text. Therefore, the recommendation system proposed in this paper can be well-applied to such platforms and provide certain reference significance of them.
The proposed recommendation model suffers from two limitations as follows:
1) In the topic-rank recommended system, we have analyzed how to deal with different topic texts, but in the same vein, it also reflects a shortcoming as follows: we never involve the processing of pictures, audio, video, etc. Nowadays, social media actively utilizes text, as well as pictures, audio, and video, which account for a large proportion. Therefore, the proposed recommendation system still has certain limitations on achieving the personalized recommendations of users.
2) The Topic-rank recommended system is based on the scenario of social networking platforms, and the final orientation is the user's behavioral intention in the Sina Weibo scenario, (in other words, whether the user likes the topic in the recommended item). Therefore, in relation to e-commerce, we need to redefine the main problem of measuring user behavior intentions. On social network platforms, users are meant to be provided with content recommendations, whereas the e-commerce platform intends to provide users with product recommendations. Due to such differences in recommended items, the theme influence is practically not prominent in the e-commerce platform recommendation system, and the weight of influence of user preference will also change accordingly, while also involving the influence of other factors.


Conclusion

As a form of information dissemination medium, major social networking platforms are widely used among people in general as an important means for them to obtain and transmit information. For example, the ways of using Sina Weibo to speed up the transmission of information and allowing users to quickly and accurately obtain the required information in the complex information wave has become a widely examined issue [2428]. Recommendation quality is an important indicator to measure the satisfaction among Sina Weibo users [2933]. Based on this, we analyze the current research status of user influence at home and abroad, then compare and explore the role and importance of users on influence. Compared with traditional technology, the Topic-rank model provides a specific user influence algorithm. In order to adapt to the current online society, using topics combined with influence to make recommendations can bring greater efficiency. Therefore, this paper has taken food collection as the research topic and refines the user influence of Sina Weibo into the user influence of the topic-relevant Sina Weibo social network, and finally obtained the user topic influence recommendation model. Foremost, we use TFIDF and LDA topic model data crawlers to cluster texts and sort out tens of thousands of Sina Weibo content into the 10 topics with the highest word frequency, which makes this research more convincing. Compared with the personalized page-rank model, the Topic-rank model used in this paper eliminates zombie fans and navy, and considers the influence of specific topics on users. The proposed model eliminates the interference caused by the low topic attention level and the larger coverage rate. The LDA model is used to extract and calculate microblog content that is highly relevant to the topic. In addition to this, it also calculates the direct and indirect influence of users to make the results more accurate. The topic divides the direct impact into three credits, namely user activity, user quality, and user Sina Weibo for calculation. Indirect influence adopts an improved SIR model, which specifically influences fans through strong and weak connections, and finally calculates the user's influence on each topic. Ultimately, we have found that the proposed recommendation model can provide Sina Weibo users more suitable contents and hope to provide certain references to online platforms, enterprises, and governments.
For follow-up work, we need further consideration of the influence difference in the degree of user preference. For example, we can distinguish the difference in influence between explicit preference and potential preference by setting different weights. Furthermore, since user preference usually changes dynamically and static user preference is difficult to meet the actual needs of users, we will introduce the time factor of user preference to build a unified framework for obtaining user dynamic preferences and conducting in-depth research.


Author’s Contributions

Conceptualization, FB, CX. Methodology, WX, FB, YF. Software, WX, YF. Validation, WX, YF, FB. Formal analysis, FB. Investigation, YF. Data curation, WX. Writing of the original draft, WX, YF, FB. Writing of the review and editing, WX, YF, FB, CX. Visualization, WX. Supervision, FB. Project administration, FB. Funding acquisition, FB, CX. All authors have read and agreed to the published version of the manuscript.


Funding

This research is funded by Natural Science Foundation of Zhejiang Province (No. LQ20G010002, LY20G010001), the National Science Foundation of China (No. 71702164), the project of China (Hangzhou) Cross-border E-commerce College (No. 2021KXYJ06), the Philosophy & Social Science Foundation of Zhejiang Province (No. 21NDJC083YB), and the Contemporary Business & Trade Research Center of Zhejiang Gongshang University (No. XT202103, XT202105).


Competing Interests

The authors declare that they have no competing interests.


Author Biography

Author
Name : Fuguang Bao
Affiliation : Zhejiang Gongshang University
Biography : Fuguang Bao is a Researcher of Contemporary Business and Trade Research Center, Zhejiang Gongshang University. He was a post-doctoral with the Department of Information System, Xidian University, from 1 January 2018 to 1 January 2020. His current research interests include Consumer Behavior, Electronic Commerce, and Personalized Recommendation.

Author
Name : Wenqian Xu
Affiliation : Zhejiang Gongshang University,cherry_kkdzz@163.com
Biography : Wenqian Xu is a graduate student in the School of Management Engineering and Electronic Commerce, Zhejiang Gongshang University. Her current research interests include public opinion transmission and personalized recommendation.

Author
Name : Yao Feng
Affiliation : Zhejiang Gongshang University,Isabella_fy2000@163.com
Biography : Yao Feng is a graduate student in the School of Management Engineering and Electronic Commerce at Zhejiang Gongshang University. Her current research interests include analysis of online public opinion and sentiment tendencies of netizens, big data predictive analysis, e-commerce and personalized recommendations.

Author
Name : Chonghuan Xu
Affiliation : Zhejiang Gongshang University
Biography : Chonghuan Xu is Associate Professor at the School of Business Administration, Zhejiang Gongshang University. He was a Visiting Fellow with the Department of Information System, City University of Hong Kong, from 1 March 2019 to 31 August 2019. His current research interests include Consumer Behavior, Marketing Communication, Electronic Commerce, and Personalized Recommendation. His work has been published in Information processing & management, Knowledge-based systems, Electronic Commerce Research and Applications, International Journal of Bio-Inspired Computation, Multimedia tools and applications, Psychology Research and Behavior Management and Frontiers in Public Health.


References

[1] Q. Y. Zhou and W. Z. Yang, “Hot topics and enlightenment of user model in recommendation system: analysis of knowledge graph based on core literatures in recent ten years,” Information Science, vol. 38, no. 9, pp. 166-173, 2020.
[2] F. Bao, Y. Wu, Z. Li, Y. Li, L. Liu, and G. Chen, “Effect improved for high-dimensional and unbalanced data anomaly detection model based on KNN-smote-LSTM,” Complexity, vol. 2020, article no. 9084704, 2020. https://doi.org/10.1155/2020/9084704
[3] Z. P. Wang and R. Xia, “Micro-blog content recommendation method based on social tag mining,” Information Science, vol. 39, no. 5, pp. 91-96, 2021.
[4] H. Li, X. Ma, J. Shi, Z. Zhong, and H. Cai, “Microblog recommendation by trust and social relationship,” Journal of Chinese Information Processing, vol. 31, no. 2, pp. 146-153, 2017.
[5] Y. He, Y. Feng, S. Zhao, and Y. Ma, “Recommending contents based on Zhihu Q&A community: case study of logistics topics,” Data Analysis and Knowledge Discovery, vol. 2, no. 9, pp. 42-49, 2018.
[6] C. S. Dai, Y. B. Wang, and Q. Wang, “Analysis of news hotspot evolution model based on online topic model,” Software Guide, vol. 19, no. 1, pp. 84-88, 2020.
[7] Z. M. Zhong, Y. Guan, Y. Hu, and C. H. Li, “Mining user interests on microblog based on profile and content,” Journal of Software, vol. 28, no. 2, pp. 278-291, 2017.
[8] A. P. Zeng, D. Liu, and H. M. Chen, “An improved apriori algorithm based on similarity,” Advanced Materials Research, vol. 532-533, pp. 1825-1829, 2012.
[9] M. Xin and L. Wu, “Using multi-features to partition users for friends recommendation in location based social network,” Information Processing & Management, vol. 57, no. 1, article no. 102125, 2020. https://doi.org/10.1016/j.ipm.2019.102125
[10] L. Guo, J. Ma, and Z. Chen, “Trust strength aware social recommendation method,” Journal of Computer Research and Development, vol. 50, no. 9, pp. 1805-1813, 2013.
[11] X. Xiao, Y. Qian, D. Li, and L. Tan, “Clustering recommendation algorithm based on user interest and social trust,” Journal of Computer Applications, vol. 36, no. 5, pp. 1273-1278, 2016.
[12] C. H. Chen, “Improved TFIDF in big news retrieval: an empirical study,” Pattern Recognition Letters, vol. 93, pp. 113-122, 2017.
[13] P. Qin, W. Xu, and J. Guo, “A novel negative sampling based on TFIDF for learning word representation,” Neurocomputing, vol. 177, pp. 257-265, 2016.
[14] H. Park, T. Park, and Y. S. Lee, “Partially collapsed Gibbs sampling for latent Dirichlet allocation,” Expert Systems with Applications, vol. 131, pp. 208-218, 2019.
[15] X. Wang, X. Yang, X. Wang, M. Xia, and J. Wang, “Evaluating the competitiveness of enterprise’s technology based on LDA topic model,” Technology Analysis & Strategic Management, vol. 32, no. 2, pp. 208-222, 2020.
[16] D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent Dirichlet allocation,” Journal of Machine Learning Research, vol. 3, pp. 993-1022, 2003.
[17] L. Liu and C. Yang “A new assessment algorithm on influence of users in microblog community,” Computer Applications and Software, vol. 34, no. 7, pp. 212-216, 2017.
[18] Y. Xing, X. Wang, L. Zhang, and S. Li, “Research on the influence of network nodes in new media environment based on information entropy: a case study of WeChat,” Library and Information Service, vol. 62, no. 5, pp. 76-86, 2018.
[19] C. Ju, Q. Gu, Y. Fang, and F. Bao, “Research on user influence model integrating personality traits under strong connection,” Sustainability, vol. 12, no. 6, article no. 2217, 2020. https://doi.org/10.3390/su12062217
[20] K. Zhuang, H. Shen, and H. Zhang, “User spread influence measurement in microblog,” Multimedia Tools and Applications, vol. 76, no. 3, pp. 3169-3185, 2017.
[21] C. Ju, F. Bao, C. Xu, and X. Fu, “A novel method of interestingness measures for association rules mining based on profit,” Discrete Dynamics in Nature and Society, vol. 2015, article no. 868634, 2015.
[22] J. Zhang and X. Cai, “Research on friend information service recommendation based on community division and user similarity,” Information Studies: Theory & Application, vol. 42, no. 4, pp. 151-157, 2019.
[23] Y. He, J. Xu, and B. Zhang, “Recommendation model lntegrating user clustering and bandits algorithm,” Journal of Chinese Computer Systems, 2021 [Online]. Available: https://kns.cnki.net/kcms/detail/21.1106.tp.20210923.1448.008.html.
[24] K. Xiang, C. Xu, and J. Wang, “Understanding the relationship between tourists’ consumption behavior and their consumption substitution willingness under unusual environment,” Psychology Research and Behavior Management, vol. 14, pp. 483-500, 2021.
[25] J. Wang, C. Xu, and W. Liu, “Understanding the adoption of mobile social payment? From the cognitive behavioral perspective. International Journal of Mobile Communications, 2022. https://doi.org/10.1504/IJMC.2022.10036150
[26] Z. Tang, H. Hu, C. Xu, and K. Zhao, “Exploring an efficient remote biomedical signal monitoring framework for personal health in the covid-19 pandemic,” International Journal of Environmental Research and Public Health, vol. 18, no. 17, article no. 9037, 2021. https://doi.org/10.3390/ijerph18179037
[27] Z. Tang, H. Hu, and C. Xu, “A federated learning method for network intrusion detection,” Concurrency and Computation: Practice and Experience, 2022. https://doi.org/10.1002/cpe.6812
[28] F. Bao, L. Mao, Y. Zhu, C. Xiao, and C. Xu, “An improved evaluation methodology for mining association rules,” Axioms, vol. 11, no. 1, article no. 17, 2022. https://doi.org/10.3390/axioms11010017
[29] C. Xu, “A novel recommendation method based on social network using matrix factorization technique,” Information Processing & Management, vol. 54, no. 3, pp. 463-474, 2018.
[30] C. Xu, A. S. Ding, and K. Zhao, “A novel POI recommendation method based on trust relationship and spatial-temporal factors,” Electronic Commerce Research and Applications, vol. 48, article no. 101060, 2021. https://doi.org/10.1016/j.elerap.2021.101060
[31] C. Xu, “A big-data oriented recommendation method based on multi-objective optimization,” Knowledge-Based Systems, vol. 177, pp. 11-21, 2019.
[32] C. Xu, A. S. Ding, and S. S. Liao, “A privacy-preserving recommendation method based on multi-objective optimisation for mobile users,” International Journal of Bio-Inspired Computation, vol. 16, no. 1, pp. 23-32, 2020.
[33] C. Xu, D. Liu, and X. Mei, “Exploring an efficient POI recommendation model based on user characteristics and spatial-temporal factors,” Mathematics, vol. 9, no. 21, article no. 2673, 2021. https://doi.org/10.3390/math9212673

About this article
Cite this article

Fuguang Bao1,2,3, Wenqian Xu2, Yao Feng2, and Chonghuan Xu1,3,4,5,*, A Topic-Rank Recommendation Model Based on Microblog Topic Relevance & User Preference Analysis, Article number: 12:10 (2022) Cite this article 2 Accesses

Download citation
  • Recived27 July 2021
  • Accepted28 January 2022
  • Published28 February 2022
Share this article

Anyone you share the following link with will be able to read this content:

Provided by the Springer Nature SharedIt content-sharing initiative

Keywords