Attendance Prediction Using Social Media


While planning for an event, every planner is disturbed by the number to expect and what do with the unexpected deviation from the expectation. They are required to put together the resources they have and those that they would need so as to ensure that an event takes place with minimum hitches. Perhaps this work of event planning would be a little much easier, if the planners were to be notified on the number of people attending. This way, controlling crowds in events would be easier as only the invited would be allowed to the event. However, in cases of popular events, this method is hard to use as more people including those who are not invited would opt to attend and this makes it harder for crowd control Social Media Attendance Prediction Sample Essay. In the modern day, the crowd control task is made simpler by social media, as most people post what they are up to on their social media platforms.

Given the cloud nature of the internet, it is therefore easier for people to predict how popular an event would be by monitoring user`s social media posts. Complimentarily, databases make it easier to know what number of people to expect in an event. This study, therefore, aims at ruling out the best pre-processing techniques that, when combined, would yield more reliable results as far as attendance prediction using social media is concerned. In this study, a combination of pre-processing techniques is employed in the analysation of tweeter tweets and, thus, giving us the possibility of people to attend an event in Australia Social Media Attendance Prediction Sample Essay.

Background Information

Social Media refers to any online platform that allows users to save their identities and most importantly allows the users to share information with friends, co-workers, other registered users and family members. Today, the use of social media platforms is becoming more popular (U.S.C. bereau, 2016). There are a number of social media platforms, each of which provides different kind of user needs today. For instance, Linkedin provides professional networks where employers can get their potential employees. Also, Social Media data provides vast information on human being’s everyday actions, feelings and thoughts. Therefore, the pattern in which users behave on social media directly matches with the day to day events in the real world Social Media Attendance Prediction Sample Essay. Therefore, according to researchers, social media use has proven reliability as far as future prediction is concerned (Philips, Dowling, Shaffer, Hodas & Volkava, 2017). Konkel (2013), confirms the reliability of social media data by arguing that information on occurrence of earthquake voyages faster on social media than the earthquake itself does on the earth`s crust.

Similarly, there have been more works done on future prediction using data from social media. For instance, EMBERS is a current system that is used in forecasting civil unrest and riots using social media data. This system uses both Social media data and non-social media data to predict when, where and why there is a possibility of riot and protest occurrence (Ramakrishnan, Butler, Muthiah, Self, Khandpur, Saraf, Wang, Cadena, Vullikanti & Korkmaz, 2014). Lately, the upcoming of Evnt based social media has been a topic of discussion to. Meetup and plancast are some of the event based social media that are well known. They offer platforms where events can be posted and the users get to choose which one they would like to attend (She, Tong & Chen, 2015). It is therefore more important to note that social media data can be useful in many reasons and has proven to be reliable for event forecasting. Therefore social media can be used as to forecast an event Social Media Attendance Prediction Sample Essay.

Zhang and Lv (2015) argue that the problem of understanding people’s attendance of an event has been a subject of active research and has therefore attracted valuable insights on human behaviour analysis and event related recommendations. Many studies explain how data on social media can be used to predict event attendance. This study however has its focus on the use of social media to forecast the attendance of an event. In this way the data from social media will assist in predicting attendance so as to manage crowds. This study, as opposed to the others looks at combination of different pre-processing techniques to filter data and make it useful in the future predictions.

Related Works

Consumption of goods and services can be predicted by observation of the customer`s fun and passion. Football play attendance for instance can be influenced by the passion the attendees have on football. On the contrary, some attendees may attend a football play due to fun (Wakefield, 2019). Social media has the data that establishes passion and fun of their users. From previous studies, it is evident that event attendance prediction can be done from posts on social Medias by the users Social Media Attendance Prediction Sample Essay. However, it is not possible to predict the possibilities of one to attend even after they post that they would. Therefore by monitoring of passion of the users, it is easier for one to tell that a user who posted that they would attend would really attend the event. Monteiro, Ounis, Mcdonald and Perego (2017) further argue that large population events today are mostly replicated in social media e.g. Facebook, twitter and Instagram. This study exploited the use of machine learning language to interpret social media data so as to come up with prediction of major events attendance.

Kalyanam, Quezada, Poblete and Lankriet (2016) argue that On-line social networks are becoming primary source of breaking news. It is believed that real life events` information is published in large quantities by the social networks daily. Some of these events can end up impacting the social networks in a very strong way. Kwak, Lee, Park and Moon (2010) confirm the same by outlining that millions of people today receive their breaking news through social media. However, according to Hu, Fu, Fang and Xu (2017), popularity of information on social media experiences a rising and falling evolution. Understanding the evolution of social media information can assist a lot especially in the control of rumours and false news. Due to this, it is easier for the information analyst to recognise what a user is up to and what he/she like doing making it easier for them to plan for future events.

As the common social media platforms are well known for their connection characteristics, the new upcoming Event Based Social networks that offer connection on both offline and online social interaction platforms offer a prodigious opening for understanding behaviours in the planetary that is cyber-physical in nature. Different factors on the event based social networks can assist a lot in the prediction of attendance of an event. These factors include but are limited to Content preference, temporal and spatial contexts and social influence (Du, Mei, Yu and Wnag, 2014). Furthermore Zhang, Zhao and Cao (2015) argue that predicting possibility of a user to attend a future event is by looking at the nature of the previous events attended. The study by Zhang, Zhao and Cao (2015) further outlines the three users` past activities features that includes semantic, temporal and spatial as features that can be of great use in predicting the possibility of one to attend an event. Therefore the data from social media plays a vibrant role in bringing up the events attended previously by the user and so makes it easier to predict the possibility of the user to attend the future event Social Media Attendance Prediction Sample Essay.

Boecking, Hall and Schneider (2015) in a study that sought to examine the past present and therefore determine the future of occurrence of events due to politics, relied on domestic events in Egypt through analysing twitter data and studying underlying communication patterns. By examining twitter data, it is easier for one to establish future events and how they can affect the people. This confirms studies that claim that future events can be predicted by observing the past data on social media.


More studies have been done on event based social networks and the studies suggest that the social networks have worked to their expectations. The event based social networks were originally meant to predict the attendance of some people in some events and due to that the users get to post what events they will be attending and therefore easier to confirm attendance of a celebrity (Wu, Dong, Shi, Swami, Chawla, 2018). However, the issue on crowd management is still not catered for by these social media as in cases of popular events, the need to look at the data of all popular social media arises. Therefore there is the need of a more reliable technique for machine language that will allow users interpret social media data into data that can give users the outcome as per what the users post. On the contrary, Facebook, currently the largest social media has made organizing of events much simpler. Event planners can send invitations to people and therefore planning becomes much simpler. However, the issue of guests not showing is still a problem Social Media Attendance Prediction Sample Essay. Therefore there is the need to design a criteria to determine a user`s likelihood to attend an event (Michalco & Navrat, 2012).

studies conducted future predictions from data obtained from social media show that social Medias are very useful for this purpose as they have data that can assist event planners predict the crowd that is likely to attend their events. This study however focuses on twitter tweets to propose a method that can be used to extract the textual features from tweets and then predict the number of attendees likely to attend an event. This study focus on establishing the best combination of pre-processing techniques that would give us the best results as far as attendance prediction is concerned.


Data in the real world is said to be inconsistent and full of errors. In data mining, pre-processing is the progression of transforming data that is raw in nature into an understandable format. Pre-processing steps are vibrant for data quality improvement. The pre-processing steps make a feature extraction process more effective and reliable (Farook, Govindaraju and Perrone, 2005). In data, the pre-processing steps filter the data so as to achieve the final refined words that makes sense. This study first sought to establish a combination of pre-processing techniques that would contain the most informative value in interpreting the text. In machine learning, NLP Bi-LSTM setup refers to the Natural language processing technique that makes use of the backwords as well as the forward propagation of signals (Uppal, 2019). This study therefore combines the pre-processing techniques with the NLP Bi-LSTM setup (used for extracting event attendance from twitter tweets) to use twitter tweets to extract event attendance.

The study formed 2 sets of techniques namely: the basic techniques used in both cases and the set of techniques that are used to determine the most reliable set of techniques in terms of event attendance accuracy

Basic Pre-Processing for all cases

The pre-processing of texts was done on all the tweets extracted. The following elaborates pre-processing steps Social Media Attendance Prediction Sample Essay.


This step served the purpose of investigating the words in a text. This step breaks down a sentence into words, phrases and symbols along with other significant elements of the sentence called the tokens.


“I haven’t slept for fourteen hours, but I think I can go another eight without it”

For the given sentence the tokens are as follows:

“I” “haven’t” “slept” “for” “fourteen” “hours” “but” “I” “think” “I” “can” “go” “another” “eight” “without” “it”


Large texts and documents have a large variance of capitalisations that work together to form sentences. Since documents are a combination of a vast amount of sentences, the capitalisation in these creates a problem when undergoing classification. This problem can be solved by turning all words into lower case; however this in its own cause’s problems as it makes interpretation of abbreviations difficult.


“US” (United States) is converted to “us” (Pronoun).

Noise Removal

Most Tweet data consists of URL’s and White-spaces. Use of these are important when co-relating the tweets with the network but in our case the classification is dependent on the text of the tweet by the user and thus the URL’s are considered as noise and hence removed.

Word Embedding

Word embedding is the most common representation of a text or document words. In this technique each word is mapped into a N dimension vector of real numbers. This way, the words gives out their contexts. There are many word embedding techniques that had been proposed Social Media Attendance Prediction Sample Essay. However the study concentrated on GloVe with in one of the most commonly used technique.


This is a technique for the feature extraction in a sentence, used for word representation. The technique uses a set of n words that occur as “the given order” in the text set. This cannot be confused for a representation of the text itself, but can be a feature that is used to represent the text. N-Gram can be used as uni-gram, bi-grams or tri-grams. However, 2 gram and 3 gram are used more commonly as they can extract more information as compared to 1-gram. An example of N-Gram is:

“I haven’t slept for fourteen hours, but I think I can go another eight without it”

For 2-gram the tokens formed are:

“I haven’t”, “haven’t slept”, “slept for”, “for fourteen”, “fourteen hours”, “hours but”, “but I”, “I think”, “think I”, “I can”, “can go”, “go another”, “another eight”, “eight without”, “without it”.

For 3-gram the tokens formed are:

“I haven’t slept”, “haven’t slept for”, “slept for fourteen”, “for fourteen hours”, “fourteen hours but”, “hours but I”,  “I think I”, “but I think”, “I can go”, “think I can”, “can go another”, “go another eight”, “another eight without”, “eight without it”.

Weighted words

Weighting is a process used to assess the significance of each word in document. It entails the assigning of numerical values to terms to show their significance in the document. The study used Term Frequency and Inverse Document Frequency (TF-IDF) method in this process. TF-IDF method reduces the effect of obliquely common words in the body and assigns higher weight to words that occur in a higher or lower frequency within the document. It is used to overcome the difficulty faced with common terms in a document, however it can’t deal with the similarity between words in a document since there are autonomously existing as an index Social Media Attendance Prediction Sample Essay.

Pre-processing techniques used to check accuracy

Stop Words

Texts and documents usually contain words that in some cases hold no significance, and thus better not be used as an input to the classification algorithms. However, these words are of no effect to the texts.


Sometimes one word may occur in a text or document in a variable form (change in tense, or singular, plural form), and the meaning of the word maybe the same. The process of uniting these words into the same feature space is by use of stemming. The stemming process changes the words to get variable forms by using different language processes. For example “sleeping” is converted into “sleep”.


This is a process to replace suffixes or remove them completely to get the basic word in a given text or document. The purpose of this is normally to remove endings so that to return to the base form of a word, also known as the lemma.

Example :

“Studies” – “study”

Number and Punctuation

It is common to find numbers and punctuation marks in a text or document. These are important for understanding the document when being read by humans however the ML techniques used for NLP are not suited for their interpretation and therefore they are removed Social Media Attendance Prediction Sample Essay.


Mentions are usually used in texts that are sourced from social media and micro-blogging sites. The use of @ before a name is used to enlarge the reach of a tweet. Therefore it can assist in showing how broad a tweet is being discussed. This cannot assist in any technique but can in cases where an analyst is viewing tweets as a user.

For example,

“The points raised by @babababloo are those that society keep at a down low.”


A hashtag is a word or phrase headed by a hash (#), and is used in to recognise a keyword or subject of attention and enables the search for it. This means that at the time of processing, the tweets in the datasets have already been selected based on a hashtag.


#LetsGo, #ComicCon, #Farewell2k18, #NewYear2k19.


The study used data sets that contained posts from twitter that were about Oz-comicon held in 2016 and 2017. Tweets were collected using the terms `Oz-comicon` and `comicon` as the identifiers to locate tweets related to comicon. The collected tweets were then split into training and testing set. Training set was generated by random sampling of 2400 different tweets obtained from the dataset. A binary label is then allocated to each tweet. The human assessment was based on the text that showed that the user who posted the tweet had an interest to attend the event. Any other topic out of the comicon event attendance possibility was definitely labelled Negative Social Media Attendance Prediction Sample Essay.

The effectiveness of our classifier is analysed with those of the baselines, in terms of its capability to predict classification accurately and tackle the issues of imbalance in the data. This study also sought to establish some of the suitable machine learning algorithms for classification. The following algorithms were chosen for the purpose of the study:

(a)Attendance classifier

(b)Naive Bayes

(c) SGD classifier

(d) Liner SVM

(e) Decision Tree Repressor

(f) Random Forest

(g) Decision Tree Classifier

Naïve Bayes as one of the algorithms selected is based on the assumption that the fact that a feature in a class is present is not related to the presence of any other feature in the class. Decision tree Repressor on the other hand is made by developing a tree from breaking down datasets into smaller subsets (Singh, Dutta & Aharwal, 2015). Linear SVM entails the designing of a vector machine that is supported in linear way from cutting plane algorithm. SGD stands for Stochastic Gradient Descent which entails the discriminative learning of linear classifiers. Mythili and Shanavas (2014) regard random forest as the tool the assembles different learning methods to come up with a common machine learning tool and decision tree classifier as a flow chart representing both internal and leaf nodes of a process.

We consider the best classification algorithm, after the comparison, to verify the most optimum combination of text pre-processing and feature in attendance classification task.

Table 1 below shows the list of cases and the combination of different techniques that were tested. From the table, 1 shows that the technique worked while x sows that the technique did not work to expectations Social Media Attendance Prediction Sample Essay.

  Lemmatization Stemming Number and Punctuation Hashtags Mentions
Case – 0 X X X X X
Case – 1 X X X X 1
Case – 2 X X X 1 X
Case – 3 X X X 1 1
Case – 4 X X 1 X X
Case – 5 X X 1 X 1
Case – 6 X X 1 1 X
Case – 7 X X 1 1 1
Case – 8 X 1 X X X
Case – 9 X 1 X X 1
Case – 10 X 1 X 1 X
Case – 11 X 1 x 1 1
Case – 12 X 1 1 X X
Case – 13 X 1 1 X 1
Case – 14 X 1 1 1 X
Case – 15 X 1 1 1 1
Case – 16 1 X X X X
Case – 17 1 X X X 1
Case – 18 1 X X 1 X
Case – 19 1 X X 1 1
Case – 20 1 X 1 X X
Case – 21 1 X 1 X 1
Case – 22 1 X 1 1 X
Case – 23 1 X 1 1 1
Case – 24 1 1 X X X
Case – 25 1 1 X X 1
Case – 26 1 1 X 1 X
Case – 27 1 1 X 1 1
Case – 28 1 1 1 X X
Case – 29 1 1 1 X 1
Case – 30 1 1 1 1 X
Case – 31 1 1 1 1 1




Table 2 below shows predictions as per the study.

  Yes No
Predicted Yes Strong Positive Weak positive
Predicted No Weak Negative Strong Negative


In our model, Strong Positive (SP) implies a correct classification of a user attendance using the textual features of the tweet. Weak Positive (WP) is tweets by users that are classified as attending the event but are not labelled as such. Weak Negative (WN) are tweets by users that are classified as not attending the event but are labelled as attending the event but are not classified as such by the algorithm. Lastly, Strong Negative (SN) is correctly classified for users that are not attending the event Social Media Attendance Prediction Sample Essay.

Accuracy is the metric that is used mostly for evaluation of the generalization performance of the classifier on a test set. It is defined as follows:

When there is great difference in  the prior probability of classes, other measures of assessment should be employed as in this case, the accuracy is said to inappropriate. Metrics remain sensitive to the bias between classes since they fail to consider costs for wrong classifications. The following are Recall (or Strong Positive Rate) and Specificity (or Strong Negative Rate) defined from the accuracy rates on Zneg and Zpos:

From these metrics, an expression of Accuracy can be given:

  1. Positive (Pos) = Strong Positive + Weak Negative
  2. Negative (Neg) = Strong Negative + Weak Positive


Hence, for imbalanced classes, Pos << Neg, displays the fact that the Recall does not affect the value of the accuracy metrics. When a balanced metric between specificity and recall is required, then a different metrics should be employed. Arithmetic mean could be an option in this case Social Media Attendance Prediction Sample Essay.

Accuracy = ½ (Recall + Specificity)

However, the geometric mean is the suitable method in such case as it combines both the specificity and recall. This is because the lower values are compensated for by higher ones in the above measure. The Geometric mean is as follows:



Research Question 1:

Research question 1 is addressed by comparison of how effective these classifiers are as compared to those of the baselines. The following algorithms will be used in the testing of the classifiers:

(a)Attendance classifier

(b)Naive Bayes

(c) SGD classifier

(d) Liner SVM

(e) Decision Tree regressor

(f) Random Forest

(g) Decision Tree Classifier

Table 3 shows the tasks Comparison with Baseline

Attendance classifier 90.78% 91.63% 91.21%
Naive 76.60% 92.78% 84.30%
SGD Classifier 80.14% 92.02% 85.87%
Linear SVM 80.14% 94.11% 86.84%
Decision Tree regressor 73.05% 84.98% 78.79%
Random Forest Classifier 65.25% 94.49% 78.52%
Decision Tree Classi_er 74.11% 87.07% 80.33%


Table 3 shows how the classifiers in this study attained high (approximately 91% accuracy) percentage of effectiveness in classifying users that are likely to participate in an event than the baselines Social Media Attendance Prediction Sample Essay.  The model in this study along with our selected performance metric helps to determine that this model can tackle the issues faced by the imbalance in the data. As seen in the graph in Fig 1 all the baseline models incline their performance towards the class with highest number of sample, in other words they tend to over-t in their training. Whereas our algorithm tends to keep a balance in the learning phase and doesn’t over train itself on a single class, best recall and accuracy is achieved by the algorithm proposed.

Research Question 2:

Research question 2 looks at the contributions of combination of text pre-processing and features in text classification task. In answering this part, we consider only the Attendance classifier classification method as it proved itself to be the best model when compared with the baselines in the previous section. The proposed model proved itself credible enough to tackle the problems faced by imbalance of the data along with its primary task of classifying tweets in accordance with the intention of attendance Social Media Attendance Prediction Sample Essay.

Table 4 Result for selected cases

Table 4 shows that using different combinations of techniques can vary the accuracy by 9%. In the preliminary studies and observation 2 findings were concluded; first, when people express their intention of attending an event in a tweet, they tend to use 5 words on average; due the dearth of attendance data in textual form and further removing the stop words, reduces the amount of interpret-able data available for the algorithm to train itself. So after this conclusion Stop words weren’t removed from any combination of techniques. Secondly using the word embedding (GloVe) with almost every combination of the techniques

From table 1, the final list of techniques used to check which combination works best include; Lemmatization, Stemming, Removing Hashtags, Removing Mentions and removing Number and Punctuation. The list of cases was reduced from 128-cases to just 32-cases.

15-fold validation was carried out for every case, out of these folds the best, worst and the average of all the folds are show in Table 4. Table 5 below shows the results.


This study proposes an approach that can be used to predict user attendance of an event from their twitter posts. This approach is mainly done by classification of non-Geo-tagged contents on the posts by the users.

The study entailed training classifiers using tweets that were related to the event (Comicon festivals in Australia). Evaluation of precision and accuracy of classifiers in comparison to multiple baselines and highlighting of the most edifying group of features was also done. From the results, this approach performed much better as compared to the other baselines unveiling 86% to 90% ac-curacy in the classification of event crowd probability. Therefore this study contributes to the other studies on the use of data obtained from social media to predict future events as in it brings the new concept of combining the pre-processing techniques to analyse data so as to come up with the predictions Social Media Attendance Prediction Sample Essay.


Boecking, B., Hall, M., & Schneider, J. (2015) Policy & Internet. Event Prediction With Learning Algorithms—A Study of Events Surrounding the Egyptian Revolution of 2011 on the Basis of Micro Blog Data V7 (2) P159 – 184

Du, R., Mei, T., Yu, Z & Wnag, Z. (2014) Predicting activity attendance in event-based social networks: Content, context and social influence Retrieved from

Farooq, F., Govindaraju, V., & Perrone, M. (2005) “Pre-processing methods for handwritten Arabic documents,” Eighth International Conference on Document Analysis and Recognition (ICDAR’05), Seoul, South Korea, V 1. P267-271

Hu, Y., Hu, C., Fu, S., Fang, M., & Xu, W. (2017) Predicting Key Events in the Popularity Evolution of Online Information Retrieved from

Kalyanam, J., Quezada, M., Poblete, B. & Lankriet, G. (2016) Prediction and Characterization of High-Activity Events in Social Media Triggered by Real-World News Retrieved from

Konkel, F. (2013)Tweets give usgs early warning on earthquakes. The Business of Federal

Kwak, H., Lee, C., Park, H., Moon, S. (2010) What is Twitter, a Social Network or a News Media? In: Proceedings of the 19th International Conference on World Wide Web. WWW’10. New York, NY, USA: ACM;. p. 591–600. Retrieved from:

Michalco, J., & Navrat, P. (2012) Arrangement of Face-to-Face Meetings Using Social Media. Studies in Informatics and Control, V21 (4)

Monteiro, V., Ounis, I., Mcdonald, C., & Perego, R. (2017)  The 2017 IEEE/ACM International Conference. Exploring social media for event attendance. Retrieved from

Mythili, M.S., and Shanavas, M. (2014) An Analysis of students’ performance using classification algorithms Retrieved from’_performance_using_classification_algorithms


Philips,L., Dowling,C., Shaffer, K., Hodas, N., & Volkava, S. (2017) Using Social Media to Predict the Future: A Systematic Literature Review Retrieved from

Ramakrishnan, N., Butler, P., Muthiah, S., Self, N., Khandpur, R., Saraf, P., Wang, W., Cadena, J., Vullikanti, A & Korkmaz, G. (2014) ’beating the news’ with embers: forecasting civil unrest using open source indicators. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1799–1808. ACM, 2014.

She, J., Tong, Y., & Chen, L. (2015) Utility-aware social event-participant planning. In SIGMOD, pages 1629– 1643. ACM

Singh, M., Dutta, A., & Aharwal, R.P. (2015) An e-Journal of RBIMS Analysis: Classification Data Mining Process in Primary Education System V1 (1)

  1. S. C. Bureau (2016) Population estimates

Uppal, A (2019) Sentence classification using Bi-LSTM. entence classification using Bidirectional-LSTM model and comparison with other baseline models Retrieved from

Wakefied, K (2019) Journal of Sports Management. Using Fan Passion to Predict Attendance, Media Consumption, and Social Media Behaviours V30(3) p229-247 Retrieved from

Wu, X., Dong, Y., Shi, B., Swami, A & Chawla, N.V.  (2018) Who will Attend This Event Together? Event Attendance Prediction via Deep LSTM Networks

Zhang, S., & Lv, Q. (2015) Event Organization 101: Understanding Latent Factors of Event Popularity. Computer Science Department University of Colorado Boulder, Boulder, CO 80309 USA

Zhang, X., Zhao, J., & Cao, G.  (2015) Who Will Attend? – Predicting Event Attendance in Event-Based Social Network. Department of Computer Science and Engineering The Pennsylvania State University, University Park, PA, 16802 Social Media Attendance Prediction Sample Essay.

Sociability and Life Expectancy Example Paper



× How can I help you?