We propose and apply a simplified nowcasting model to understand the correlations between social attention and topic trends of scientific publications.

First, topics are generated from the obesity corpus by using the latent Dirichlet allocation (LDA) algorithm and time series of keyword search trends in Google Trends are obtained. We then establish the structural time series model using data from January 2004 to December 2012, and evaluate the model using data from January 2013. We employ a state-space model to separate different non-regression components in an observational time series (i.e. the tendency and the seasonality) and apply the “spike and slab prior” and stepwise regression to analyze the correlations between the regression component and the social media attention. The two parts are combined using Markov-chain Monte Carlo sampling techniques to obtain our results.

The results of our study show that (1) the number of publications on child obesity increases at a lower rate than that of diabetes publications; (2) the number of publication on a given topic may exhibit a relationship with the season or time of year; and (3) there exists a correlation between the number of publications on a given topic and its social media attention, i.e. the search frequency related to that topic as identified by Google Trends. We found that our model is also able to predict the number of publications related to a given topic.

First, we study a correlation rather than causality between topics’ trends and social media. As a result, the relationships might not be robust, so we cannot predict the future in the long run. Second, we cannot identify the reasons or conditions that are driving obesity topics to present such tendencies and seasonal patterns, so we might need to do “field” study in the future. Third, we need to improve the efficiency of our model by finding more efficient variable selection models, because the stepwise regression method is time consuming, especially for a large number of variables.

This paper analyzes publication topic trends from three perspectives: tendency, seasonality, and correlation with social media attention, providing a new perspective for identifying and understanding topical themes in academic publications.

To the best of our knowledge, we are the first to apply the state-space model to examine the relationships between healthcare-related publications and social media to investigate the relationships between a topic’s evolvement and people’s search behavior in social media. This paper thus provides a new viewpoint in the correlation analysis area, and demonstrates the value of considering social media attention in the analysis of publication topic trends.

#### Keywords

- Social media
- Publication topic trends
- Correlation
- State-space model
- Variable selection
- Nowcasting

We live in a data-driven society where social media platforms allow users to generate content to share, communicate, and discuss their opinions with each other on various events without being at the same time and in the same place. The effect of social media on our lives has never been greater on personal as well as political and social concerns, where its impacts can be seen from local and national to global arenas. An emergence of societal and political topics thus corresponds with the traffic of social media discussions. For example, news and comments about natural disasters and political uprisings travel at breakneck speed in Twitter or Google. Social media are becoming an important channel that can also greatly influence our perception of a certain topic. In economics, social media can have a significant effect on a firm’s reputation, sales, and even survival (Kietzmann et al., 2011). Altmetrics (Priem et al., 2010) use social media to estimate the early impact of publications or researchers. Dong and Bollen (2015) applied Google search-engine query data to detect consumer confidence indexes. Scott and Varian (2014) predicted weekly initial claims for unemployment and monthly retail sales using Google Trends and Google Correlate data. These evidences indicate that a range of patterns can be discovered by analyzing people’s behaviors and their topically relevant online activities in social media.

In a similar vein, the proposed work seeks to understand whether there is the correlation between the major social media Google Trend and scientific topics in academic publications. Correlation analysis identifies the degree of relationship or dependency between two types of variables. For example, the Pearson product-moment correlation coefficient (PMCC) and rank-correlation coefficient are widely used in linear relationship analysis (Kendall, 1962; Rodgers & Nicewander, 1988). PMCC and rank-correlation coefficient are not robust correlations, however, especially when there are outliers in samples; consequently, the result might be different from the truth (Wilcox, 2005). And when other things significantly influence the objective index, it becomes difficult to find a correct correlation. We therefore propose a new method to find correlations between scientific topic trends in academic publications and the social media attention they garner by means of a regression analysis between these social media and scientific topic trends. We first divide the scientific topic trends in academic publications into three parts: tendency, seasonality, and correlations with social media attention that are named the regression component. We then analyze the correlation between the regression component and the social media attention.

We approach the possibility of these correlations with the concept of “seasonality” that describes the observable patterns of topical tendency over time. Seasonality helps us understand a topic’s cyclic changes by accounting for fluctuation in its own dynamics. A number of seasonality findings have been presented in the economics and engineering literature. For example, Scott and Varian (2014) proposed a nowcasting (a contraction of “now” and “forecasting”) model commonly used in economics and meteorology to separate tendency, seasonality, and regression effects from economic phenomena. Analyzing topic seasonality is critical to understanding how a topic evolves in a characteristic pattern, especially for research related to topic trends and prediction. In this paper, we quantify a seasonality factor as well as overall dynamic tendency and relationships between topic evolutions and potential indicators. To this end, we choose the seasonable topic “obesity” with a particular focus on two commonly representative sub-topics, “child obesity” and “diabetes,” both of which are major concerns in current health initiatives due to their increasing prevalence.

On a global scale, obesity has more than doubled since 1980 where two million children under age five were overweight or obese in 2013 (WHO, 2015). In America the large majority of American adults are obese, making it a nationwide epidemic. The condition has been implicated as a leading factor in deadly diseases such as heart disease, stroke, and diabetes. Obesity has become a focus of attention for broader audiences, including scientific researchers, social media users, medical experts, and patients themselves. Obesity-related topics, including diet, lifestyle, and child obesity, are frequently discussed in both academic publications and non-professional social media.

A recent study indicated that obese youth are likely to have a risk of cardiovascular disease (e.g. high cholesterol or high blood pressure), where among boys from age 5 to 17, 70% are at risk for cardiovascular disease (Freedman et al., 2007). Obese adolescents are more likely to have pre-diabetes conditions, where their blood glucose levels indicate a high risk of diabetes (Li et al., 2009; Centers for Disease Control and Prevention, 2011). Another study found that children with obesity as young as age two are likely to have obesity later in life (Freedman et al., 2005). Obesity at an early age may often lead to weak bones and joints, sleep apnea, and numerous social and psychological problems, such as stigmatization and poor selfesteem (Daniels et al., 2005; Dietz, 2004). It is also likely to trigger diseases later in life, including heart disease, Type 2 diabetes, stroke, osteoarthritis, and several types of cancer. These include organ cancers such as breast, colon, endometrium, esophagus, kidney, pancreas, gall bladder, thyroid, ovary, cervix, and prostate, as well as multiple myeloma and Hodgkin’s lymphoma (Kushi et al., 2006). Extensive research efforts and accomplishments related to obesity have thus been conducted and published by academic scholars, which leads to a large number of digital texts. Detailed textual analyses on issues such as topic extraction, topic evolution, and topic trending are important for gaining a comprehensive understanding of obesity for better healthcare initiatives and health policy planning.

We propose a simplified nowcasting model that combines stepwise regression and Bayesian sampling to describe the relationships between obesity topics and social media. The experimental results show that the proposed state-space model can capture the impact of dynamic tendency and seasonality, and the impact of public attention from social media (e.g. Google Trends). To the best of our knowledge, we are the first to apply the state-space model to examine the relationships between healthcare-related publications and social media to investigate the relationships between a topic’s evolvement and people’s search behavior in social media. The proposed study investigates the nature of a topic’s evolvement in three components (i.e. dynamic tendency, seasonality, and correlation with social media) and the ways in which it evolves (i.e. quantification of the above three components).

This paper is organized as follows: the data and methodology section describes the proposed model. The result section explains the experiment results for the topics of “child obesity” and “diabetes.” The discussion section evaluates the performance of the model. Finally, the paper summarizes and suggests future work.

This paper proposes to study topics involved in obesity publications, seasonality patterns behind these topics, and the relationships between the topics and social media attention. We use academic papers and online search queries to derive topics and social attention based on two datasets: (1) academic papers about obesity in PubMed during a specific time period, and (2) Google Trends data for a specific number of search queries over the same time period.

For study data we downloaded obesity-related papers from PubMed for the period of January 2004 to January 2013. We use data from January 2004 to December 2012 for modeling and the January 2013 data for evaluation. Search strategies are based on the following terms (including plurals and variants) as determined by checking the Unified Medical Language System (UMLS) and consulting medical domain experts: OBESITY, OBESE, ADIPOSITY, OVERWEIGHT, EXCESS FOOD INTAKE, FEAR OF BECOMING FAT, LEPTIN, and BARIATRIC. In total, 98,063 articles (henceforth the “obesity paper dataset”) are collected.

Google Trends

The data are normalized by the overall search volumes. It is scaled so that the maximum time series number equals 100, i.e. values on

Google Trends data express some trends of search queries, i.e. the change in search volume of queries over time. Meanwhile, the change in the number of obesity-related papers over time could also be seen as a type of trend. One of our tasks herein is to find a relationship between these two trends. The search query’s context can be set to several categories, such as Arts & Entertainment, Finance, Games, and Health. In this paper, we choose seven categories where people might talk about obesity: Beauty & Fitness, Food & Drink, Health, Hobbies & Leisure, Jobs & Education, People & Society, and Sports. Categories such as Shopping or Pets & Animals are not included. PubMed data and Google Trends time-series data can be matched. Since Google Trends data can be provided weekly and PubMed data are released monthly, we convert all weekly data to monthly by taking a four-week moving average. For every selected topic discussed above, we obtain Google Trends time-series data from January 2004 to January 2013.

The overall framework of the methodology is shown in Figure 2, including generating topics from the obesity corpus using the latent Dirichlet allocation (LDA) algorithm (Blei, Ng, & Jordan, 2003), obtaining time series of keyword search trends in Google Trends, training the structural time series model using data from January 2004 to December 2012, and evaluating the model using data from January 2013.

In this paper, we employ a state-space model to separate different non-regression components in an observational time series (i.e. the tendency component and the seasonality component) and apply the “spike and slab prior” and stepwise regression to analyze the correlations between the regression component and the social media attention. We combine the two parts using Markov-chain Monte Carlo (MCMC) sampling techniques to make the model run continuously, step by step:

Data preparation: Using LDA, we obtain the two topics (consisting of a number of keywords) from the obesity database (PubMed articles), and determine the probability that a particular document

State-space model: In the section Model Training, the first three sub-sections (i.e. spike and slab prior, prior of

Variables selection and regression: We then use the spike and slab prior for the first step of variables selection, and use stepwise regression to finish this work. Last, we can get a correlation between the regression component and the social media attention variables, i.e. a new

Repetition of Step 2 and Step 3: Markov-chain Monte Carlo (sub-section 5) is used to repeat Step 2 and Step 3 to obtain convergent results.

Model evaluation: Forecasting (sub-section 6) could be used to evaluate our model to some degree.

The state-space model is one of the most popular methods used to solve the time-series problem. Commonly used for dynamic analysis, the model provides a unified methodology for treating a wide range of problems in time-series analysis (Durbin & Koopman, 2001). State-space time-series analysis began with Kalman (1960) and has been widely applied in engineering, economics, and social sciences.

State-space methods have yielded valuable results in recent years. For example, Rueda and Rodríguez (2010), in a study estimating and forecasting fertility rates, introduced multivariate state-space models that are dynamic alternatives to logistic representations for fixed time points. Costa and Alpuim (2010) contributed to the problem of state-space model parameter estimation by proposing estimators for the mean, the autoregressive parameters, and the noise variances. Al-Anaswah and Wilfling (2011) used a state-space model with Markov-switching to detect speculative bubbles in stock-price data, and found that in the stock markets considered, their identification procedure correctly detects most speculative periods that have been classified as such by economic historians. Unnikrishnan (2012) made a prediction of magnetic sub-storms using a state-space model, generating outputs for storm events that reasonably reproduce the observed values, which demonstrates its prediction capability. Dong et al. (2014) focused on developing flexible and explicitly multivariate state-space models for network flow rate and mean-speed predictions. Using two-minute measurements from an urban freeway network, they provided practical guidance for selecting the most appropriate models for congested and non-congested conditions. Ghosh et al. (2014) introduced Bayesian inference in nonparametric dynamic state-space models, and illustrated their methods with simulated datasets, using the Markov-chain Monte Carlo (MCMC) approach for studying posterior distributions of interest.

The unique contribution of the state-space model is its modularity. It can capture the relationship between a dependent variable and each individual independent variable. In this paper, we apply a state-space model to separate the tendency, seasonality, and regression component from the publication trends under a scientific topic. Each of these three components can be modeled separately (Draper & Smith, 1998). Tendency of a topic reflect quantitative changes (increase or decrease) in publication volume over time, which may result from the existence of more authors and the development of technology over time. Seasonality is a time-series-based pattern that is predictable and can be repeated at different periods within a year. Seasonality tells us the periodic change in publication numbers of a topic over a year, which may result from the publication cycle of related journals and conferences. The regression component is something that correlates with the public’s social attention that is used to determine the popularity of a topic based on the frequency of query keywords. Due to its modularity and capability in separating and evaluating regression components individually, we apply a state-space model to study impacts of the tendency, seasonality, and regression component on the number of publications.

Let _{t}

There are two equations in this model, where the first one is known as the observation equation because it links the observed variable with the unobserved latent state variable, and the second one is known as a transition equation because it defines an evolvement process for the unobserved variable. Where _{t}_{t}_{t}_{t}_{t}η_{t}_{t}_{t}

Yet because we want the model to capture effects of the three aspects of tendency, seasonality, and regression component, the model could be rewritten as:

where _{t}_{t}_{t}_{t}β

Following the methods discussed in Scott and Varian (2014), we assign the regression effects _{t}β_{t}_{t}_{t}

where

The training of the state-space model involves variable selection and sampling techniques. Here we use the spike and slab prior and stepwise regression (Draper & Smith, 1998) for variable selection, and the Koopman smoother and MCMC methods (Durbin & Koopman, 2001) for simulation and sampling.

Since there are a large number of variables in regression effects, and while the observation data are typically brief, we use the spike and slab prior on the regression coefficients to indicate sparsity. The spike and slab prior can make some of the coefficients of regression variables zeros. Let _{k}_{k}_{k}_{k}_{γ}_{k}

It is very common to see a Bernoulli distribution (0 – 1 distribution) as the spike and slab prior. From _{K×1} as such:

But the initial value of _{k}_{k}

This assumes that the conditional prior ^{2} after obtaining a sample _{1:n} from independent Gauss distribution ^{2}), the posterior of ^{2} will be expressed in the following form:

Initial values of

where _{t}^{2}_{1:n}, compute the value of

According to Zellner’s ^{−1} as follows: Let _{1}, _{1}_{n}^{T}_{t}

where ^{T}X^{T}X_{1:n}). For the symmetric matrix ^{−1}, let ^{−1} related to _{k}

Then the conditional prior of _{γ}

where _{γ}_{γ}_{γ}

Let

The initial values of

There are many methods available for solving the state-space model, such as the Kalman filter (Andrew, 1989; Kalman, 1960) and Bayesian computation methods (De Jong & Shephard, 1995; McCausland, Miller, & Pelletier, 2011). As _{1:n}) under the assumption that _{1} and _{1} are known, where _{1} ~ _{1}, _{1}). It then modifies the initial conditions as P_{1} has infinite variance.

Using _{1:n}, we can get the error terms _{1:n}, _{1:n}, _{1:n} from Equation (1). The corresponding post prior of

We can select the most important independent variables from

Denote _{t}_{t}β_{t}^{*} = ^{*} is

Then, based on the _{γ}_{γ}_{k}

Estimate a _{γ}^{*} and _{γ}

Get a _{γ}

Within Equation (1), _{t}_{t}

where

Simulate

Simulate

Estimate

Simulate

Repeated cycling through the four steps above generates a sequence of draws ^{(1)}, ^{(2)}, ^{(3)}, … (say, 6,000 times). These ^{(i)} (

This model also has prediction ability. For every _{n+1} using Equation (1). Then we can obtain _{n+1} after estimating _{n+1}_{n+1} is known. The forecasting work can be done using the mean:

In this paper we investigate impacts of the three components defined in our state model on the number of publications from PubMed data: tendency, seasonality, and influence of the regression component correlated with Google Trends data. Each PubMed article related to obesity consists of fields of PubMed ID, title, abstract, date, authors, and others. LDA is applied to obtain latent topics from this textual information (Blei, Griffiths, & Jordan, 2010). The number of topics (

The number of publications (^{t}) related to a specific topic (^{th} month can be calculated from the following two steps: (1) topic distribution _{i}^{th} topic in that document; and (2) ^{t} can be obtained by summing the

In the results of LDA topics modeling, two topics are highly related to diabetes. We thus combine them into one topic based on the additive law of probability, i.e.

Another output from LDA is the topic-term distribution. For each topic, every word has a corresponding probability, indicating its likelihood belonging to that topic. For topics “child obesity” and “diabetes,” we choose 50 and 100 keywords with the highest probabilities and use them as search-query keywords in Google Trends. Each word is searched under seven different categories defined by Google, to obtain seven time-series-based trends. If the number of queries using one of these words is too small to have an explicit trend, the word under that category is dropped. We eventually find 344 and 616 search volume trends for topics “child obesity” and “diabetes,” respectively.

We define two time-series-based measures of the publication topic trends (i.e. tendency and seasonality) here for the topics “child obesity” and “diabetes.” The tendency of a topic, the first component of the publication topic trends, is the change in the number of publications for that topic over time. It represents the overall popularity of that topic at different time periods. Changes in these tendencies could result from many aspects, such as an increase in the scholars in this field, related technology under rapid development, and more research findings emerging over time. Figure 4(a) and 4(b) show the tendencies for the two chosen topics. The increasing trends indicate that the number of publications related to these two topics generally increased from 2004 to 2014, except for some fluctuations between 2004 and 2006. It also indicates that a growing amount of scientific research is being directed to these two topics, as the scholars in this field are publishing more research.

It is still difficult to detect the rate of change by viewing only these tendencies, however. We use “growth rate” to capture the changing rate. Figure 4(c) and 4(d) show growth rates for the two topic tendencies, where the average growth rate for “child obesity” is 0.158 and 0.183 for “diabetes.” This means that the number of publications on “child obesity” increases more slowly than the number for “diabetes.” While research on “child obesity” is very popular, research on “diabetes” has become more popular in recent years.

Tendency and growth rate can help us explain the overall characteristics of topic evolution, such as their increase (or decrease) or speed of increase (or decrease) in popularity over different time periods. However, they do not make it easy to answer the following questions: what are major factors leading to these fluctuations, and are there any patterns behind these changing curves? To resolve these issues, considering the publication of related periodicals and the convening of related meetings are determined to some degree. The truth would result in a common phenomenon: the publication numbers on the two chosen topics increase in some months and decrease in other months, where the same is true every year, i.e. some of these changes are repeated periodically. This phenomenon, defined as seasonality in this paper, is seen in many domains. For example, sales of down jackets increase in fall and winter and decrease in spring and summer, data that have remained consistent from year to year for obvious reasons. To capture the seasonal effect on topics among publications (which are not obvious) we consider seasonality as a separate component in our model. As shown in Figure 5, there are more publications on “child obesity” published in June, September, and December than in January, May, and November. Similarly, more publications on “diabetes” are published in April, June, and December than in January, May, July, and November. The number of publications on a topic in a time period might indicate the degree of interest in that topic for a given period.

The extent of interest in obesity and diabetes can be observed in two ways: the number of scientific publications in related journals and conferences, and the frequency of related query terms in Google Trends used by the public. In our model, we have a regression component that describes the correlation between Google Trends data and the number of publications. The regression components for “child obesity” are mostly positive, while negative or close to zero for “diabetes,” as shown in Figure 6. A positive

In this paper, we propose a state-space model to capture individual factors of tendency, seasonality, and regression component (correlated with social media attention in Google Trends) in the number of publications for a topic, to help us comprehensively understand the topic, including what the topic and its evolution is, how it evolves, and its relationships with social media. We choose two commonly representative sub-topics of “child obesity” and “diabetes” as cases to demonstrate our findings. We use stepwise regression for variable selection, combined with MCMC to train our model. The experimental results show that (1) the number of publications on “child obesity” increases at a slower rate than that of “diabetes” publications, which indicates that the research on “diabetes” is becoming more popular in recent years; (2) different topics exhibit different seasonal patterns in terms of the number of publications on such topics. There are more publications on “child obesity” published in June, September, and December than in January, May, and November. Similarly, more publications on “diabetes” are published in April, June, and December than in January, May, July, and November; and (3) there exists a relationship between the number of publications on a given topic and the search frequency of terms related to that topic on Google Trends.

In spite of the promise of this novel approach to capturing factors of topics, there are a number of shortcomings that should be addressed in future work. For example, we study a correlation rather than causality between topics’ trends and social media. As a result, the relationships might not be robust, so we cannot predict the future in the long run. Also, we cannot identify the reasons or conditions that are driving obesity topics to present such tendencies and seasonal patterns. This limitation may be caused by government policy, climate change, or even new technology. To find the causal relationships, we need to do “field” study in the future. In addition, we need to improve the efficiency of our model by finding more efficient variable selection models, such as Bayesian inference, because the stepwise regression method is time consuming, especially for a large number of variables. Furthermore, we will try to model topics using related indicators that express social attention in other forms of social or traditional media, such as blog data or newspaper data, so we can make our model more accurate by triangulating or cross-validating results. We also want to incorporate sub-model techniques that could transform into our algorithm higher dimensions of big data into a certain number of relatively smaller data subsets, as described by Zhou et al. (2013), which might make our model more efficient.