Pfitzer, Garas, and Shweitzer [3] classified twitter posts as serving distinctly one of two functions. A tweet is either information creation or the pure distribution of information where a user reposts another’s original idea or thought (retweet). It was found that emotional divergence has an impact on the probability of a piece of infor- mation being retweeted; tweets with higher emotional divergence have a higher probability of being retweeted. Previous work in sentiment analysis on Twitter has revealed that there is a distinct correlation of the swing of the collective mood due to the ability to retweet another user’s post.

In a study by Garas, Garcia, Skowron, and Shweitzer [4] the communication patterns used across online cha- trooms were analyzed for patterns of both real time information exchange and overall emotion. Models of online interaction were used to determine and analyze emotional triggers in conversation, as well as a comparison of behavior both online and in the real world. It was discovered that people did not change their pattern of expres- sion significantly between previously studied forms of communication. Additionally, instant chatrooms do not show the same heated emotional patterns that are often found in alternate forms of online communication; in- stantaneous reply chats are more balanced in emotion about various topics.

The language of tweets is unique due to the 140-character limit imposed upon individual posts, causing users to often utilize shorthand notation as well as emoticons in sentiment expression [5] . Pak and Paroubek [6] have done groundbreaking work in the field of short text analysis while classifying tweets based on the correlation between emoticons and parts of speech. The emoticons were used to determine the overall sentiment of the tweet [7] , as the character limit makes it more likely for there to be only a single sentiment. The remainder of the tweet was split into distinct parts of speech using the Tree Tagger algorithm, showing which part of speech has the greatest impact on a post’s overall sentiment [6] [8] .

Most closely related to the research done in this paper is the work of Aston, Liddle, and Hu [9] on Twitter sentiment classification in data stream. The Perceptron algorithm and its voted version with static feature selec- tion were used to predict sentiment in a data stream environment for large amounts of Twitter data. Their streaming prediction achieved similar accuracy to that of the batch prediction of [2] . The words in the top fea- tures were uncovered.

Similarly we aim to apply the Modified Balanced Winnow algorithm to social network sentiment classifica- tion in data stream and to implement online feature selection in conjunction with this in order to account for changing data over time.

Traditional studies in sentiment analysis have primarily used batch mode learning to repeatedly traverse rela- tively small datasets. However, in most real world applications of algorithms designed to analyze sentiment from OSNs, the datasets are much larger and constantly changing due to the nature of online social networking. The combination of the sheer size and dynamic nature of OSNs makes batch learning an impractical solution [10] .

In the case of OSN sentiment classification using a streaming algorithm is a more efficient solution. Analyz- ing data in a stream allows posts to be processed in real time as they appear on social networks. Running algo- rithms in a data stream environment means that it is only viable to perform a single pass over the data. This me- thod results in decreased accuracy, but means that we gain a considerable improvement in runtime. For these reasons as well as the ability to handle small changes in data streams over time [10] , we chose stream algorithms to predict sentiment, using Massive Online Analysis (MOA) and Weka frameworks [11] .

For the purpose of this research we utilized three public datasets: Sanders Corpus [12] , STS-Gold [13] , and SentiStrength [5] .

Sanders Corpus consists of 5513 manually classified tweets. As a large number of these have become unavaila- ble since the creation of the dataset and because certain tweets were irrelevant to our study of sentiment analysis the total number used was reduced to 3320. We were interested in two subsets of the reduced Sanders Corpus, specifically analyzing tweets for positive or negative sentiment. In the positive/negative subset all tweets with a neutral label were removed from the dataset, then the remaining tweets are analyzed for positive or negative sentiment overall.

This dataset contains 2034 tweets, which are hand labeled according to their sentiment, either positive or nega- tive. They are assigned sentiment values of either 0 or 4 based on how negative (0) or positive (4) the tweet is. This dataset annotates the tweets and entities (target subject) separately, allowing for finer sentiment of tweets containing annotated entities.

This dataset is composed of 4242 posts from 5 different social media sites (MySpace, YouTube, Digg, BBC, Runners World) which have been manually labeled according to their sentiment, with a score being given for both positive and negative sentiment appearing in the post. Each post is assigned two scores, one from −1 to −5 representing how negative the post is with −5 being the most negative, and another from 1 to 5 representing how positive the post is with 5 being the most positive.

Because we are representing the posts as n-grams, each one has 95ⁿ features. As a result of the large number of features for each post, classification takes a great deal of time, which is problematic in a data stream environment. For this reason it is advantageous to reduce the number of features that must be considered when performing classification. In addition to time consumption, many features of each post do not play a significant role in the classification of the sentiment. Thus we performed online feature selection to discover the top features of each gram representation [14] . Unlike static feature selection methods, which find the top features once in a batch learning environment [9] , we performed feature selection dynamically, with a newly calculated set of top features at each time interval of 100 sequential posts. At first we removed all but the top features when performing clas- sification but found that this resulted in a considerable reduction in accuracy. We discovered that in order to maintain a high level of accuracy while reducing runtime it was necessary to keep a percentage of the less sig- nificant features in the classification of posts.

The Winnow algorithm is a mistake driven machine learning technique that uses a promotion parameter α in or- der to train on pre-labeled instances [14] . The instances are labeled with boolean values and classified as either 0 or 1. Each instance is input to the learner, classified, and finally compared to the actual class label. The process of the Winnow algorithm is outlined briefly in Winnow Algorithm below.

Balanced Winnow extends upon the Winnow algorithm by adding a demotion parameter β. It utilizes a promo- tion parameter α > 1, a demotion parameter 0 < β < 1, and a threshold θ. For each instance trained on, if the in- stance is predicted correctly then the training step is skipped, if it is predicted as 1 with an actual class 0 the weight matrix is multiplied by the demotion parameter, and if it is predicted as 0 with actual class 1 the weight matrix is multiplied by the promotion parameter.

We used MBW to classify posts. Like Balanced Winnow, MBW requires a promotion parameter α and demotion parameter β. It is separated from Balanced Winnow by utilizing larger margins and a minor change in the update rules for the weight matrices [14] . MBW is outlined in detail in the figure Modified Balanced Winnow. We in- itially trained the MBW classifier over the first 100 instances. For each following instance we first classified it using MBW; if the classification was correct we continued to the next instance and updated the correct count. If it was incorrect we updated the incorrect count and updated the weight matrix.

MBW requires four different user defined parameters (promotion, demotion, feature selection, good feature se- lection), which may lead to a decrease in accuracy when chosen randomly. Therefore we performed extensive checks over these parameters to reveal possible good ranges for each. It was determined that the larger the gram size used, the more accurate our prediction became, since larger gram sizes reveal more of a word than smaller gram sizes. Also, when using feature selection and good feature selection, we noticed that accuracy was greatest with the more features and good features used, as shown in Table 1. We divided our datasets into sequential segments of size 100 to reveal the importance of features over time. We then ran MBW on each segment of 100 as a stream.

Our division created 22 segments (or timestamps) for Sanders. Figure 1 illustrates the feature selection of the top 20 features over the 22 timestamps of Sanders 5 gram representation. As the figure illustrates, the top fea- tures oscillate in importance in the beginning, but as the data is continually fed in, the few top features begin to stabilize and retain their higher position. Features that are subsets of the words “Google”, “Apple”, and “Andro- id” are some of these features that retain a higher importance over time. For all gram sizes of Sanders, high ac- curacy predictions fell within the range of promotion and demotions values of 0.9 and 1.1.

A high accuracy is seen when α and β values increase together from around {1, 1} onward. By incorporating dynamic feature selection in our MBW we achieved an accuracy of 73.3% while [9] achieved an accuracy of 77% using 5 grams with manual feature selection. In terms of a data stream, it is important to perform dynamic feature selection due to the changing importance of features with new incoming data.

The work of [9] performed feature selection on a static batch and then performed sentiment analysis on this reduced feature set. This approach has downfalls as unknown data coming in over time will change the impor- tance of features.

The SentiStrength dataset was divided into 112 sets of size 100 each. We achieved an accuracy of 73.6% using promotion and demotion values of 0.7 and 1.4 respectively. Features such as “love”, “you”, “happy”, “thank”, and “good” contain the highest importance through this dataset, shown in Figure 2 of 5 grams representation. The STS_Gold dataset was divided into 42 sets, with the highest accuracy of 87.5% for promotion and demotion values of 2 and 1 respectively.

The promotion (α) and demotion (β) table of SentiStrength 3 grams representation revealed that values of α and β which optimize the accuracy of prediction seem to follow a logarithmic pattern from α in the range 1 - 3 and β in the range 0.2 - 0.9. In the STS_Gold dataset high accuracy is seen when α is greater than 1 but less than 3 and β is less than α.

In terms of accuracy with feature selection, it is best to use all features. In Sanders and STS_Gold, the differ- ence in accuracy from using all features to using 10% of the features is around 10%, while with SentiStrength, the difference is only around 2% to 3%.