Hi all, this is a Data Science Project ! - Twitter Analytics with Machine Learning and Deep Learning methodologies.
- The purpose of this project is to apply Machine Learning and Deep Learning methodologies in social media analytics such as Twitter in order to predict the game results of NBA. Thus, more business opportunities may be detected and utilized not only for NBA but also for other relevant sport leagues.
- Inferential Statistics
- Web Scraping
- Data Wrangling
- Data Visualization
- Machine Learning
- Deep Learning
- Predictive Modeling
-
Supervised Learning:
- Naive Bayes Classifiers (NB)
- Support Vector Machines (SVM)
-
Unsupervised Learning:
- K-means Clustering (K-means)
- Latent Dirichlet Allocation (LDA)
-
Deeping Learning:
- Convolutional Neural Network (CNN)
- Python, Jupyter Notebook
- Numpy, Pandas, Sklearn, Nltk
- Matplotlib, Scipy, Seaborn, Keras
-
Motivation:
- Since we are NBA sport fans and twitter lovers, we decide to work on a social media analytics project combining the skills and technologies from we have learned to see how it works. Therefore, we apply Machine Learning and Deep Learning methodologies to predict the NBA game results from Twitter. Meanwhile, in order to obtain good performance of each methodology, accuracies are also important and necessary for every individual user.
-
Data and Scope:
-
There are total 30 teams in the NBA league and can be divided by 15 teams respectively in Eastern and Western conferences. Here, we randomly pick 8 teams as sample data and scrape the data from tweets.
-
Since we cannot obtain the data for more than past 30 days from the standard twitter API, we move on to find another way to gain the data.
-
We apply twitter search engine to receive the tweets related to “NBA teams” in 2016 and 2017. The basic idea is to request directly from tweets and then return them as JSON files.
-
We collect tweets every day for the entire 2017 NBA season. Since the competition started from 2016-10-25 to 2017-04-12, we get more than 2,000 tweets for a day and totally more than 2,500,000 tweets!
-
-
Initial Text Analysis:
-
At first, we do not know which methodology has the best performance for this project. Therefore, we consider and decide to try three possible ways (Highest Frequency Words, Injury and Recovery Factors and Sentiment Analysis) to test with our sample dataset as our first step.
-
Highest Frequency Words:
- Separate the entire dataset into three subset datasets based on the time period and they are "3 days", "7 days" and "1 month" individually.
- Remove the useless contents such as 'twitter', 'http', 'com', 'pic', 'ift', 'tt', 'https' and so on.
- After pointing out the top 20 highest frequency words in each time period, we take 5 most meaningful words for the demonstration.
- However, the results are not good enough to make the game predictions accurately.
-
Injury and Recovery Factors:
- In this section, we are more interested in finding some influential factors which could affect the results of the games.
- Therefore, we wonder if there are injured players or recovered players in the team because they may have huge influences on the game results if they are key players for that team.
- Here are some commom key words for tweets related to injury and recovery words:
- Injury words (Negative): ['hurt','injury','injured','broken','tear','missed','ill', 'illness']
- Recovery words (Positive): ['recover','recovery','return','health','healthy','heal', 'back', 'rehab']
- After that we count these words and determine if Injury words > Recovery words then which means it is a good expectation for that team.
- Otherwise, if Recovery words > Injury words then which means it is a bad expectation for the team.
-
Sentiment Analysis:
- First of all, we count the amount of positive/negative words from all tweets. Once the amount of positive words greater than negative words we treat it as a good result.
- In addition, if good results are more than bad results 24 hours before the competition, we predict the result of the game is to win.
- Furthermore, we apply TF-IDF model to filter tweets and discover this accuracy is much higher than the previous ones.
-
-
After processing three different kinds of methodologies, we find out that Sentiment Analysis has the best performance at the end; therefore, we determine to implement Sentiment Analysis with mentioned algorithms.
-
-
In this section, we gain the accuracy of combining algorithms with and without Sentiment Analysis:
-
Supervised Learning Algorithms:
(1) Accuracy of Naive Bayes Classifiers (NB):
Non_Sentiment Analysis Sentiment Analysis around 50% to 60% around 50% (2) Accuracy of Support Vector Machines (SVM):
Non_Sentiment Analysis Sentiment Analysis around 50% to 60% around 50% -
Unsupervised Learning Algorithms:
(3) Accuracy of K-means Clustering (K-means):
Non_Sentiment Analysis around 50% (4) Accuracy of Latent Dirichlet Allocation (LDA):
Non_Sentiment Analysis around 50% -
Deep Learning Algorithm:
(5) Accuracy of Convolutional Neural Network (CNN):
Non_Sentiment Analysis around 60% - P.S. We do not apply Sentiment Ananlysis for Unsupervised Learning and Deep Learning Algorithms because Sentiment Ananlysis is not appropriate for Unsupervised Learning and also no need for Deep Learning Alogrithms.
-
- After comparing with five different algorithms, we conclude that supervised learning algorithms (Naive Bayes and SVM) have better performances than unsupervised learning algorithms (K-means and LDA). The possible reason is because labels utilized in supervised learning alogritms may strengthen the features of tweets; therefore, it can improve the accuracy of predictions in NBA game results. However, CNN still has the highest accuracy performance over than any other methodologies. That is to say, CNN is the most practical algorithm in this twitter project and we should keep modifying the parameters set inside the model in order to gain a better performance.