Understanding User Topic Preferences across Multiple Social Networks

Current work on mining users’ topic preferences usually uses the data generated by individual data, which include explicitly labeled profile information, such as age, date of birth, gender, interests, and other data. The information also includes users’ comments and rating data of various content in social networks[9]. However, due to the service limitations of online social networks or personal privacy issues, users’ detailed data cannot be easily obtained. Most studies often need to integrate users’ implicit feedback data, such as data on users’ publicly posted content, forwarded content information, etc. Hong et al.[10] view users’ posting data in the Twitter network as documents and use the LDA topic model to model users’ topic model. Considering that users’ posts in social networks are relatively short, Zhao et al.[2] assume each user’s post only has one topic and then restructured a probability generating model based on the LDA model to mine users’ topic preferences. Cao et al. [11] combined users’ social, interest, and behavioral footprints in the Twitter network and used the matrix decomposition technique to obtain users’ topic preference features. Chen et al.[12] divided users’ attributes into individual interests and shared interests based on users’ item ratings. Related works have further used multi-modal data to enrich the user’s feature content. You et al.[13] analyzed the images in social networks. They train the image data to extract the hidden features. Then, based on the hidden features, a label propagation algorithm is constructed to achieve images classification. At last, they obtain a user interest model based on image information. Farnadi et al.[14] used deep learning technique to fuse multi-modal data from social networks. The fused data were classified based on the Big Five personality traits of users from the MyPersonality project, and then mining and analysis of user characteristics was carried out.

Mislove et al.[15] pointed out that users with similar attributes are more likely to establish friendships and can use the characteristics of users’ friends to infer users’ attributes. Chen et al.[16] used relationships and structure characteristics of social networks to construct a user-label attribute prediction model. Xu et al.[17] modeled users based on three factors, including breaking news, posts from social friends, and user’s intrinsic interest.

Due to the sparsity of users’ data, related works have also attempted to integrate data from similar users. Hu et al. [18] developed a topic model to analyze the change of users’ topic preference from a community view; Feng et al. [19] used user groups to alleviate the sparsity problem of short texts on social media and to detect emotions and relevant topics based on groups; Wang et al.[20] integrated users’ topic and network topologies to catch the topic correlations inside a community.

Users have kinds of topic content in social networks, and each user has unique topic preferences than others. To model this situation, we view users’ topic preference distribution as a multinomial distribution over topics $\theta_{u}$. Each topic $z$ can be generated from the distribution.

Users create posts based on their topic preferences. In our MSNT model, the word generation process over the topic is divided into two types of processes.

Global topics words generation: In multiple social network situations, the model first uses a multinomial distribution $\varphi_{z}^{p}$ to generate word $w$ over topic $z$ based on the user’s all content data. This can find co-occurring words from the global perspective on all social networks. We can discover users’ topics based on complete multiple social network data to reflect the intrinsic preferences of users.

Local topics words generation: The specific social network may also influence a user to use different words to make posts. The model uses a multinomial distribution $\varphi_{z}^{s}$ to generate word $w$ over topic $z$ only on user content data in a specific social network. We can find topics to reflect the specific content characteristic based on the single social network.

In order to unify users’ global topics and local topics, this paper describes a switch variable $x$ to control the selection of words for topic content. The probability distribution that $x$ takes the 0 or 1 is denoted as a binomial distribution $\sigma_{zs}$. The binomial distribution can represent the latent proportion distribution of global topic words and local topic words. When $x=0$, the users select global topic word $w$ under topic $z$, and the distribution of this process is denoted as multinomial distribution $\varphi_{z}^{p}$. When $x=1$, the users generate local topic word $w$ with topic $z$ in social network $s$, and distribution of this process is denoted as multinomial distribution $\varphi_{z}^{s}$. The semantic content between the global topic and local topic is associated with each other by switch variables. In addition, we assume that a background topic captures the background words used across the multiple social networks to improve the ability to mining other topics. Therefore, the switch variable $x$ first takes the value tag from a binomial distribution $\sigma_{B}$. If $x$ get $2$, word $w$ is generated from a background topic word multinomial distribution $\varphi_{B}$. If $x$ is not $2$, the switch variable $x$ then takes value from the binomial distribution $\sigma_{zs}$. The word $w$ is from the global topic or local topic depending on the variable $x$ value of $0$ or $1$.

This paper also assumes each post $c$ has only one topic $z$ by referring to the work on social network topic mining[23, 8], because the length of posts in social networks is much smaller than the length of normal documents, and the user will only focus on one topic when creating a post.

The service differences in social networks drive users to choose different social networks. Users may choose different social networks to post content based on topics. Users’ social network choice is influenced by potential relevance between users’ topic preferences and social network topics. In the model, the probability distribution that user $u$ chooses social network $s$ with topic $z$ is described as a multinomial distribution $\rho_{uz}$.

We summarize the above generative process as follows.

1. 1.

   Sample the background topic vocabulary Multinomial distribution $\varphi_{B}$ from a prior Dirichlet distribution: $\varphi_{B}\sim Dir(\beta_{B})$; Sampe the Binomial distribution of whether user’s content word is background word from a prior beta distribution: $\sigma_{B}\sim Dir(\tau_{B})$;

2. 2.

   For each topic $z$,

   1. (a)

      Sample the global topic vocabulary Multinomial distribution $\varphi_{z}^{p}$ from a Dirichlet prior distribution. $\varphi_{z}^{p}\sim Dir(\beta^{p})$;

   2. (b)

      For each social network $s$,

      1. i.

         Sample the local topic vocabulary Multinomial distribution in social networks $s$ from a prior Dirichlet distribution: $\varphi_{z}^{s}\sim Dir(\beta^{s})$; Sample the Binomial distribution of whether user’s content word is global or local topic word from a prior beta distribution: $\sigma_{zs}\sim Dir(\tau_{zs})$;

      2. ii.

         Sample the user social network preference Multinomial distribution from a prior Dirichlet distribution: $\rho_{uz}\sim Dir(\lambda_{uz})$;

3. 3.

   For each user $u$,

   1. (a)

      Sample the user topic preference Multinomial distribution from a prior Dirichlet distribution: $\theta_{u}\sim Dir(\alpha)$.

   2. (b)

      For each user $u$’s post,

      1. i.

         Sample a topic assignment $z$ from Multinomial distribution: $z\sim Mult(\theta_{u})$,

      2. ii.

         Sample a social network indicator $s$ from Multinomial distribution: $s\sim Mult(\rho_{uz})$;

      3. iii.

         For each word $w$,

         1. A.

            Sample a $x$ value from binomial distribution: $x\sim Bern(\sigma_{B})$,

         2. B.

            If $x=2$, sample a word $w$ from Multinomial distribution: $w\sim Mult(\varphi_{B})$;

         3. C.

            If $x\neq 2$, sample a $x$ value from binomial distribution: $x\sim Bern(\sigma_{zs})$. If $x=0$, sample a word $w$ from Multinomial distribution: $w\sim Mult(\varphi_{z}^{p})$. If $x=1$, sample a word $w$ from Multinomial distribution: $w\sim Mult(\varphi_{z}^{s})$.

Viewing users as a document set, the model in this paper can be regarded as a document topic discovery model. Therefore, this paper uses the Perplexity[25], Likelihood[25] and word Pointwise Mutual Information Score (PMI-score)[26] metrics to evaluate the model performance.

Specifically, the perplexity is a widely used evaluation metric in the field of natural language processing, and the higher quality of the model is corresponding to the lower perplexity value. Its specific form is shown in Eq.(9).

where $P(w_{d}|M)$ denotes the probability that document $w_{d}$ is output by model $M$ and $N_{d}$ denotes the length of document $w_{d}$. Here, since this paper considers each post of user as a document and marks it as one topic. Then $P(w_{d}|M)=\sum_{k\in Z}\theta^{k}_{u}\prod\limits_{w\in W_{d}}P(w_{d}|M)$.

The likelihood calculates the probability of generating the sentence in the test corpus. The likelihood value is higher, and it can show better quality of the model. Its specific form is shown in Eq.(10).

Generally, the multiplication value of probability is transformed into summation by log operation to prevent the multiplication value from being too small.

The PMI-score measures the topic quality based on the co-occurrence word pairs in one topic. The words within the topic are correlated, and the higher mutual information score indicates a higher model quality. Its specific form is shown in Eq.(11).

where, $K$ denotes the number of topic. $T$ denotes the number of words in topic. The word mutual information calculates the score of the $Top-T$ words with the highest probability of occurrence in the topic. $PMI_{k}(w_{i},w_{j})=log\frac{p(w_{i},w_{j})}{p(w_{i})p(w_{j})}$. $p(w_{i},w_{j})$ is probability of co-occurring word pair $w_{i}$ and $w_{j}$. $p(w)$ is probability of occurrence of word $w_{i}$.

To demonstrate the effectiveness of the proposed MSNT model, we choose different representative baseline methods as follows.

LDA[25], which is a classical topic discovery model. During topic discovery, the model integrates multiple social network data directly without distinguishing the source of the data, and one document consists of all contents generated by the user.

TwitterLDA[23], which takes into account the short-text features in social networks, links one post to only one topic in the topic discovery process, and the model integrates data of multiple social networks without distinguishing the sources of data.

MTM[24], the model can mine topic-related opinion words from different corpus. During the topic discovery process, the model can distinguish the source corpus of the data. In the experiments, various social networks are corresponding to the different corpus. The topic discovery part of the model is also equivalent to the ccLDA [35] topic discovery model.

MutilLDA[8], which has similar features to TwitterLDA, links one post to only one topic in the topic discovery process. The model considers users’ choice preferences of social networks across multiple social networks.

MSNT, which is the topic discovery model across multiple social networks in this paper.

First, we show perplexity experiment results. The lower perplexity value is better. It can be observed that perplexity values of all models tends to decrease as the number of topic increases shown in Fig.3.

Starting from the number of topics, about 80, the decrease of the perplexity value gradually tends to be smooth. In particular, the perplexity values of TwitterLDA, MultiLDA, and the MSNT are significantly lower than MTM and LDA. The reason is that TwitterLDA, MultiLDA, and MSNT are specific to short texts of social networks, and the background topics in these models reduce the influence of noisy words in social networks. The perplexity experiment results of MSNT in this paper are better than TwitterLDA and MultiLDA, because MSNT is designed to apply to the scenario of multiple social networks compared with TwitterLDA. For MultiLDA, MSNT involves users’ preference of choosing social networks and mining users’ topic preference from a global view on fusing multiple social networks data. It can also discover differences in local topic content across multiple social networks. Overall, MSNT obtains better performance than other works due to modeling global topics and local topics being more reasonable in multiple social network scenarios, i.e., consistency and complementarity co-exist.

The likelihood value can describe the generation probability of a sentence in the test set based on the model. Its rising and decreasing trends are similar to perplexity. From experiment results shown in Fig.4, we can also see that the increasing trend of likelihood gradually becomes smooth as the number of topics increases. Compared with other works, MSNT significantly gets better experiment results.

For word PMI-score, which is used to measure the lexical relevance in one topic, the higher mutual information score indicates a higher model quality. Experiment results are shown in Fig.5. It can be seen that the number of topic is from 20 to 100, and the top 50 words are taken in one topic to calculate the word PMI-score under different topic numbers.

The figure shows that TwitterLDA, MultiLDA, and MSTN perform better than MTM and LDA because these models are designed based on social networks with short texts. The MSTN in this paper achieves better results than TwitterLDA and MultiLDA because MSTN considers local topics that are more likely to have some feature topic words in a specific social network. These words are easier to form co-occurring word pairs in a specific social network, thus obtaining better results about word PMI-score.

Perplexity is utilized to determine the topic number in related work[25]. When the number of topics reaches 80, the perplexity progressively becomes smooth, as mentioned in the previous section. Therefore, we use the experiment results when the number of topics is 80 to show the general topic content of each network. The number of topics is also the same in work[8].

The general topic words are shown in in Tab.IV. We can find that the general topic contents that users like to discuss on Twitter are daily stuff, media events, etc. Twitter users use more oral words. There are larger verbs used by users compared with other social media platforms. Due to users’ daily images on Instagram, classified as a photo-based social network, the objects in photos are typically people or scenes, and the subject is inclined to landscape, fashion, food, etc. Tumblr resembles a typical blog in appearance. Tumblr features more text material than Twitter, as well as more nouns and fewer vocal verbs. On the Tumblr network, users create content by combining words with other multimedia elements, such as photographs, to describe topics. The topic content shown in Tab.IV is about vehicles and business marketing, respectively.

Then we compare global and local topic content in each social network and understand users’ latent motivations for using social networks. We discover that, for one type of topic content, global topic content has an explicit relationship with each social network’s local topic content. Different social network service providers can share data and give more comprehensive services to users by utilizing global topics and local topics. As shown in Tab.V, the topicA content is food, and the local topic has a large number of words that are explicitly embodied in the global topic. Global and local topics are more relevant. When a user prefers this type of topic, their preference tends to be consistent across networks.

The other type of topic content, shown in Tab.VI, is that local topic content is the implicit correlation to the global topic. The notable words in the local topic content are not the same as keywords in the global topic content. TopicB is a topic about residential design, local topic content in Instagram reflects “plant” decorations, and local topic content in Tumblr focuses on the room design itself. The local topic reflects the influence of social networks on users’ topic content. For latent semantic content, some things described in global topic content can be examined and understood through local topic content. It is feasible to deliver customized demand services to users in each social network based on this type of topic.

We can quantitatively explore the similarities and differences between global topics and local topics across multiple social networks.

Based on the probability distribution of word over topic $p(w|z)$, i.e., $\varphi_{z}^{p}$ or $\varphi_{z}^{s}$ in this paper, we can calculate the similarity between two word distributions over topic across social networks. This paper uses Jensen-Shannon Divergence (JSD) to measure the similarity, which takes values in the range of $[0,1]$, and two probability distributions are more similar to each other with respect to the lower JSD score.

We first calculate JSD scores between each identical local topic across social networks. Then we can obtain proportion distribution based on JSD scores. The JSD score proportion distribution of identical local topics across social networks is shown in Fig.6. Overall, the JSD values are concentrated in the middle of distribution in most cases, reflecting the consistency and complementarity are co-exist across multiple social networks as described in this paper.

Similarly, we can also obtain the JSD similarity between the local topic and the global topic in each social network, i.e., the JSD score between $\varphi_{z}^{s}$ and $\varphi_{z}^{p}$. The JSD score proportion distribution between local topic and global topic across social networks is shown in Fig.7. It can be seen that most of the distributed JSD values are concentrated in the interval less than 0.3, which indicates the consistency of topic content between global topic and local topics.