UFC Fight Analysis

Project Overview

Selected topic

The topic selected by the team was UFC Fight Analysis of all UFC fights from 2013.

Reason topic was selected

The team selected to analyze UFC fights from 2013 because the team members had prior interest in UFC fighting, and were intrigued by the data contained in the dataset.

Description of the source of data

• The original data of UFC fights from 2013 was obtained from Kaggle.
• The scraped data of all UFC fights was obtained from UFC Stats.

Questions the team hopes to answer with the data

The questions we hope to answer with the data include:

• Can our machine learning model predict the winner (target) based on the features?
• Can our machine learning model predict the winby based on features?
• Is there a relationship between fighter age and winner outcome?
• Is there a relationship between fighter height and winner outcome?
• Is there a relationship between fighter weight and winner outcome?

Communication Protocols

The team attends a standing meeting daily from 6-7pm EST on Discord to discuss progress made on the project, and other project-related matters. The team also maintains constant communication as-needed via Discord chat. The team maintains meeting notes, scheduling, and organization in Notion.

Data Exploration

The team explored various sites for the most interesting and feasible dataset, and finally settled on UFC Fight Data from Kaggle. After exploring and cleaning the data, the team discovered several issues within the dataset, including mismatched values to some rows. The team decided the best course of action would be to scrape the data directly from the Kaggle dataset's source, which was UFC Stats. The team developed a scraper to scrape data from the UFC Stats website into a new CSV file to explore, clean, and preprocess for analysis.

Created charts

The team created various charts to gain a better understanding of the data, such as the comparison between Red and Blue Winners, and Box & Whiskey Plots to identify outliers in the data.

• Winner (Red vs. Blue)

  

• Box & Whisker Plots

  • Age

    

  • Height

    

  • Weight

    

Created buckets

The team bucketed the Age, Weight, and Height data, then created charts of the bucketed groups to gain a better visualization of the fighters' stats.

• R_Age Bucket

  

• B_Age Bucket

  

• R_Height Bucket

  

• B_Height Bucket

  

Database Integration

The team created a database in pgAdmin, which contained the following 4 tables:

1. ufc_table - The table containing all scraped data.
2. fighter_stats - The table containing all fighter stats.
3. fight_stats - The table containing all fight stats.
4. joined_table - The table joining [table1] and [table2].
5. fighter_agg_stats - The table aggregating fighter, irrespective of Red or Blue corner.

The database tables were populated from within the UFC_Final_Project.ipynb Python file. The team used to_sql to overwrite the table with updated scraped data each time time the file is run. Then the team used the psycopg2, sqlalchemy, and io libraries to populate the tables in the pgAdmin database with data from the correspoding Pandas DataFrames.

Database Schema:

• Given the number of features we are dealing with the above image is not able to capture the all of the table descriptions. If you are interested, you can download a .txt file for the full schema here.

Machine Learning Model

Preliminary Data Preprocessing

During the preliminary data preprocessing phase of the project, the team performed the following actions to clean and transform the data as a preprocessing step for the machine learning model.

1. Imported scraped data into a Pandas DataFrame

2. Dropped duplicate rows (fights)

   • To ensure the scraped data did not contain duplicate rows (fights), duplicate rows were dropped using drop_duplicates based on the columns Event_Date, B_Name, and R_Name, where Event_Date contained the date of the fight, B_Name contained the name of the Blue fighter, and R_Name contained the name of the Red fighter.

3. Converted Event_Date column values to datetime64 data type

   • The data type of the Event_Date column was converted to datetime64 using the to_datetime function.

4. Dropped rows (fights) that happened before 5/3/2001

   • UFC Fights had little to no rules prior to 5/3/2001. Examples of this were as follows:
     • Some fights had no time limit.
     • Some fights included fighters of different weight classes, putting fighters of the lower weight classes at a disadvantage. In some cases, fighters had 100+ lb. weight discrepencies.
   • However, major rule changes were implemented on 5/3/2001, eliminating these unfair circumstances. As such, rows with an Event_Date before 5/3/2001 were dropped to maintain consistency in the rules set forth in the fights analyzed.

5. Replaced "--", "---" and "No Time Limit" with np.NaN

   • No Time Limit should already not exist due to the date restriction above but if it does, it will be replaced with NaN.
   • "--" and "---" represent NO value. Not zero; Nothing. An example would be the take-down percentage column, where these values are present quite often. This is due to the fact that the fighter didn't even attempt a single take-down. To clarify a little more, if a fighter was to attempt a take-down but failed to land that take-down, they would then have a take-down percentage of 0%.

6. R_Draws and B_Draws were split to create a No_Contest for each corner color

   • Some of the Draws column values contained "(x NC)", where "x" represents the amount of no contests.
   • The "x" value was extracted and put into its own No_Contest column.

7. Rearranged columns

   • With the new No_Contest columns created, the DataFrames columns are rearranged to for origination.
   • R_No_Contest column moved to the position after R_Draws.
   • B_No_Contest column moved to the position after B_Draws.

8. Used .loc on the Weight_Class column in order to keep the standardized weight classes

   • Standardized weight classes: Heavyweight, Light Heavyweight, Middleweight, Welterweight, Lightweight, Featherweight, Bantamweight, Flyweight, Strawweight, Women's Strawweight, Women's Flyweight, Women's Bantamweight, Women's Featherweight.

9. R_Height and B_Height bucketed using quartile (4 buckets created).

   • R_Height_Bucket and B_Height_Bucket columns created.

10. R_Age and B_Age bucketed using quartile (4 buckets created).

    • R_Age_Bucket and B_Age_Bucket columns created.

11. Gender column created based on Weight_Class column value containing "Women's" or not.

    • If the fighter is a women, the Gender column will contain a value of 0.
    • If the fighter is a man, the Gender column will contain a value of 1 .

12. Converted columns to best inferred possible dtypes using .covert_dtypes supporting pd.NA.

    • These inferred data types may not be correct and in our situation, a lot were incorrect.
    • Columns with the data type of "string" or "object" were inspected to figure out why they were inferred this way.
      • No issues were found in any of the columns so they were converted to the correct data type (Categorical OR Numerical).

13. Set Categories converted to category datatype using astype

Categorical Data:

View Categorical Columns

'Weight_Class', 'Win_By', 'B_Name', 'B_Stance', 'R_Name', 'R_Stance', 'R_Age_Bucket', 'B_Age_Bucket', 'R_Height_Bucket', 'B_Height_Bucket', 'Gender'

Numerical Data:

View Numerical Columns

'Max_Rounds', 'Ending_Round', 'B_Age', 'B_Height', 'B_Weight', 'B_Reach', 'B_Wins', 'B_Losses', 'B_Draws', 'B_No_Contest', 'B_Career_Significant_Strikes_Landed_PM', 'B_Career_Striking_Accuracy', 'B_Career_Significant_Strike_Defence', 'B_Career_Takedown_Average', 'B_Career_Takedown_Accuracy', 'B_Career_Takedown_Defence', 'B_Career_Submission_Average', 'B_Knockdowns', 'B_Significant_Strikes_Landed', 'B_Significant_Strikes_Attempted', 'B_Significant_Strike_Perc', 'B_Significant_Strikes_Distance_Landed', 'B_Significant_Strikes_Distance_Attempted', 'B_Significant_Strikes_Clinch_Landed', 'B_Significant_Strikes_Clinch_Attempted', 'B_Significant_Strikes_Ground_Landed', 'B_Significant_Strikes_Ground_Attempted', 'B_Head_Significant_Strikes_Attempted', 'B_Head_Significant_Strikes_Landed', 'B_Body_Significant_Strikes_Attempted', 'B_Body_Significant_Strikes_Landed', 'B_Leg_Significant_Strikes_Attempted', 'B_Leg_Significant_Strikes_Landed', 'B_Total_Strikes_Attempted', 'B_Total_Strikes_Landed', 'B_Takedowns_Attempted', 'B_Takedowns_Landed', 'B_Takedown_Perc', 'B_Submission_Attempts', 'B_Grappling_Reversals', 'B_Round_One_Knockdowns', 'B_Round_One_Significant_Strikes_Landed', 'B_Round_One_Significant_Strikes_Attempted', 'B_Round_One_Significant_Strike_Perc', 'B_Round_One_Significant_Strikes_Distance_Landed', 'B_Round_One_Significant_Strikes_Distance_Attempted', 'B_Round_One_Significant_Strikes_Clinch_Landed', 'B_Round_One_Significant_Strikes_Clinch_Attempted', 'B_Round_One_Significant_Strikes_Ground_Landed', 'B_Round_One_Significant_Strikes_Ground_Attempted', 'B_Round_One_Head_Significant_Strikes_Attempted', 'B_Round_One_Head_Significant_Strikes_Landed', 'B_Round_One_Body_Significant_Strikes_Attempted', 'B_Round_One_Body_Significant_Strikes_Landed', 'B_Round_One_Leg_Significant_Strikes_Attempted', 'B_Round_One_Leg_Significant_Strikes_Landed', 'B_Round_One_Total_Strikes_Attempted', 'B_Round_One_Total_Strikes_Landed', 'B_Round_One_Takedowns_Attempted', 'B_Round_One_Takedowns_Landed', 'B_Round_One_Takedown_Perc', 'B_Round_One_Submission_Attempts', 'B_Round_One_Grappling_Reversals', 'B_Round_Two_Knockdowns', 'B_Round_Two_Significant_Strikes_Landed', 'B_Round_Two_Significant_Strikes_Attempted', 'B_Round_Two_Significant_Strike_Perc', 'B_Round_Two_Significant_Strikes_Distance_Landed', 'B_Round_Two_Significant_Strikes_Distance_Attempted', 'B_Round_Two_Significant_Strikes_Clinch_Landed', 'B_Round_Two_Significant_Strikes_Clinch_Attempted', 'B_Round_Two_Significant_Strikes_Ground_Landed', 'B_Round_Two_Significant_Strikes_Ground_Attempted', 'B_Round_Two_Head_Significant_Strikes_Attempted', 'B_Round_Two_Head_Significant_Strikes_Landed', 'B_Round_Two_Body_Significant_Strikes_Attempted', 'B_Round_Two_Body_Significant_Strikes_Landed', 'B_Round_Two_Leg_Significant_Strikes_Attempted', 'B_Round_Two_Leg_Significant_Strikes_Landed', 'B_Round_Two_Total_Strikes_Attempted', 'B_Round_Two_Total_Strikes_Landed', 'B_Round_Two_Takedowns_Attempted', 'B_Round_Two_Takedowns_Landed', 'B_Round_Two_Takedown_Perc', 'B_Round_Two_Submission_Attempts', 'B_Round_Two_Grappling_Reversals', 'B_Round_Three_Knockdowns', 'B_Round_Three_Significant_Strikes_Landed', 'B_Round_Three_Significant_Strikes_Attempted', 'B_Round_Three_Significant_Strike_Perc', 'B_Round_Three_Significant_Strikes_Distance_Landed', 'B_Round_Three_Significant_Strikes_Distance_Attempted', 'B_Round_Three_Significant_Strikes_Clinch_Landed', 'B_Round_Three_Significant_Strikes_Clinch_Attempted', 'B_Round_Three_Significant_Strikes_Ground_Landed', 'B_Round_Three_Significant_Strikes_Ground_Attempted', 'B_Round_Three_Head_Significant_Strikes_Attempted', 'B_Round_Three_Head_Significant_Strikes_Landed', 'B_Round_Three_Body_Significant_Strikes_Attempted', 'B_Round_Three_Body_Significant_Strikes_Landed', 'B_Round_Three_Leg_Significant_Strikes_Attempted', 'B_Round_Three_Leg_Significant_Strikes_Landed', 'B_Round_Three_Total_Strikes_Attempted', 'B_Round_Three_Total_Strikes_Landed', 'B_Round_Three_Takedowns_Attempted', 'B_Round_Three_Takedowns_Landed', 'B_Round_Three_Takedown_Perc', 'B_Round_Three_Submission_Attempts', 'B_Round_Three_Grappling_Reversals', 'B_Round_Four_Knockdowns', 'B_Round_Four_Significant_Strikes_Landed', 'B_Round_Four_Significant_Strikes_Attempted', 'B_Round_Four_Significant_Strike_Perc', 'B_Round_Four_Significant_Strikes_Distance_Landed', 'B_Round_Four_Significant_Strikes_Distance_Attempted', 'B_Round_Four_Significant_Strikes_Clinch_Landed', 'B_Round_Four_Significant_Strikes_Clinch_Attempted', 'B_Round_Four_Significant_Strikes_Ground_Landed', 'B_Round_Four_Significant_Strikes_Ground_Attempted', 'B_Round_Four_Head_Significant_Strikes_Attempted', 'B_Round_Four_Head_Significant_Strikes_Landed', 'B_Round_Four_Body_Significant_Strikes_Attempted', 'B_Round_Four_Body_Significant_Strikes_Landed', 'B_Round_Four_Leg_Significant_Strikes_Attempted', 'B_Round_Four_Leg_Significant_Strikes_Landed', 'B_Round_Four_Total_Strikes_Attempted', 'B_Round_Four_Total_Strikes_Landed', 'B_Round_Four_Takedowns_Attempted', 'B_Round_Four_Takedowns_Landed', 'B_Round_Four_Takedown_Perc', 'B_Round_Four_Submission_Attempts', 'B_Round_Four_Grappling_Reversals', 'B_Round_Five_Knockdowns', 'B_Round_Five_Significant_Strikes_Landed', 'B_Round_Five_Significant_Strikes_Attempted', 'B_Round_Five_Significant_Strike_Perc', 'B_Round_Five_Significant_Strikes_Distance_Landed', 'B_Round_Five_Significant_Strikes_Distance_Attempted', 'B_Round_Five_Significant_Strikes_Clinch_Landed', 'B_Round_Five_Significant_Strikes_Clinch_Attempted', 'B_Round_Five_Significant_Strikes_Ground_Landed', 'B_Round_Five_Significant_Strikes_Ground_Attempted', 'B_Round_Five_Head_Significant_Strikes_Attempted', 'B_Round_Five_Head_Significant_Strikes_Landed', 'B_Round_Five_Body_Significant_Strikes_Attempted', 'B_Round_Five_Body_Significant_Strikes_Landed', 'B_Round_Five_Leg_Significant_Strikes_Attempted', 'B_Round_Five_Leg_Significant_Strikes_Landed', 'B_Round_Five_Total_Strikes_Attempted', 'B_Round_Five_Total_Strikes_Landed', 'B_Round_Five_Takedowns_Attempted', 'B_Round_Five_Takedowns_Landed', 'B_Round_Five_Takedown_Perc', 'B_Round_Five_Submission_Attempts', 'B_Round_Five_Grappling_Reversals', 'R_Age', 'R_Height', 'R_Weight', 'R_Reach', 'R_Wins', 'R_Losses', 'R_Draws', 'R_No_Contest', 'R_Career_Significant_Strikes_Landed_PM', 'R_Career_Striking_Accuracy', 'R_Career_Significant_Strike_Defence', 'R_Career_Takedown_Average', 'R_Career_Takedown_Accuracy', 'R_Career_Takedown_Defence', 'R_Career_Submission_Average', 'R_Knockdowns', 'R_Significant_Strikes_Landed', 'R_Significant_Strikes_Attempted', 'R_Significant_Strike_Perc', 'R_Significant_Strikes_Distance_Landed', 'R_Significant_Strikes_Distance_Attempted', 'R_Significant_Strikes_Clinch_Landed', 'R_Significant_Strikes_Clinch_Attempted', 'R_Significant_Strikes_Ground_Landed', 'R_Significant_Strikes_Ground_Attempted', 'R_Head_Significant_Strikes_Attempted', 'R_Head_Significant_Strikes_Landed', 'R_Body_Significant_Strikes_Attempted', 'R_Body_Significant_Strikes_Landed', 'R_Leg_Significant_Strikes_Attempted', 'R_Leg_Significant_Strikes_Landed', 'R_Total_Strikes_Attempted', 'R_Total_Strikes_Landed', 'R_Takedowns_Attempted', 'R_Takedowns_Landed', 'R_Takedown_Perc', 'R_Submission_Attempts', 'R_Grappling_Reversals', 'R_Round_One_Knockdowns', 'R_Round_One_Significant_Strikes_Landed', 'R_Round_One_Significant_Strikes_Attempted', 'R_Round_One_Significant_Strike_Perc', 'R_Round_One_Significant_Strikes_Distance_Attempted', 'R_Round_One_Significant_Strikes_Distance_Landed', 'R_Round_One_Significant_Strikes_Clinch_Attempted', 'R_Round_One_Significant_Strikes_Clinch_Landed', 'R_Round_One_Significant_Strikes_Ground_Attempted', 'R_Round_One_Significant_Strikes_Ground_Landed', 'R_Round_One_Head_Significant_Strikes_Attempted', 'R_Round_One_Head_Significant_Strikes_Landed', 'R_Round_One_Body_Significant_Strikes_Attempted', 'R_Round_One_Body_Significant_Strikes_Landed', 'R_Round_One_Leg_Significant_Strikes_Attempted', 'R_Round_One_Leg_Significant_Strikes_Landed', 'R_Round_One_Total_Strikes_Attempted', 'R_Round_One_Total_Strikes_Landed', 'R_Round_One_Takedowns_Attempted', 'R_Round_One_Takedowns_Landed', 'R_Round_One_Takedown_Perc', 'R_Round_One_Submission_Attempts', 'R_Round_One_Grappling_Reversals', 'R_Round_Two_Knockdowns', 'R_Round_Two_Significant_Strikes_Landed', 'R_Round_Two_Significant_Strikes_Attempted', 'R_Round_Two_Significant_Strike_Perc', 'R_Round_Two_Significant_Strikes_Distance_Attempted', 'R_Round_Two_Significant_Strikes_Distance_Landed', 'R_Round_Two_Significant_Strikes_Clinch_Attempted', 'R_Round_Two_Significant_Strikes_Clinch_Landed', 'R_Round_Two_Significant_Strikes_Ground_Attempted', 'R_Round_Two_Significant_Strikes_Ground_Landed', 'R_Round_Two_Head_Significant_Strikes_Attempted', 'R_Round_Two_Head_Significant_Strikes_Landed', 'R_Round_Two_Body_Significant_Strikes_Attempted', 'R_Round_Two_Body_Significant_Strikes_Landed', 'R_Round_Two_Leg_Significant_Strikes_Attempted', 'R_Round_Two_Leg_Significant_Strikes_Landed', 'R_Round_Two_Total_Strikes_Attempted', 'R_Round_Two_Total_Strikes_Landed', 'R_Round_Two_Takedowns_Attempted', 'R_Round_Two_Takedowns_Landed', 'R_Round_Two_Takedown_Perc', 'R_Round_Two_Submission_Attempts', 'R_Round_Two_Grappling_Reversals', 'R_Round_Three_Knockdowns', 'R_Round_Three_Significant_Strikes_Landed', 'R_Round_Three_Significant_Strikes_Attempted', 'R_Round_Three_Significant_Strike_Perc', 'R_Round_Three_Significant_Strikes_Distance_Attempted', 'R_Round_Three_Significant_Strikes_Distance_Landed', 'R_Round_Three_Significant_Strikes_Clinch_Attempted', 'R_Round_Three_Significant_Strikes_Clinch_Landed', 'R_Round_Three_Significant_Strikes_Ground_Attempted', 'R_Round_Three_Significant_Strikes_Ground_Landed', 'R_Round_Three_Head_Significant_Strikes_Attempted', 'R_Round_Three_Head_Significant_Strikes_Landed', 'R_Round_Three_Body_Significant_Strikes_Attempted', 'R_Round_Three_Body_Significant_Strikes_Landed', 'R_Round_Three_Leg_Significant_Strikes_Attempted', 'R_Round_Three_Leg_Significant_Strikes_Landed', 'R_Round_Three_Total_Strikes_Attempted', 'R_Round_Three_Total_Strikes_Landed', 'R_Round_Three_Takedowns_Attempted', 'R_Round_Three_Takedowns_Landed', 'R_Round_Three_Takedown_Perc', 'R_Round_Three_Submission_Attempts', 'R_Round_Three_Grappling_Reversals', 'R_Round_Four_Knockdowns', 'R_Round_Four_Significant_Strikes_Landed', 'R_Round_Four_Significant_Strikes_Attempted', 'R_Round_Four_Significant_Strike_Perc', 'R_Round_Four_Significant_Strikes_Distance_Attempted', 'R_Round_Four_Significant_Strikes_Distance_Landed', 'R_Round_Four_Significant_Strikes_Clinch_Attempted', 'R_Round_Four_Significant_Strikes_Clinch_Landed', 'R_Round_Four_Significant_Strikes_Ground_Attempted', 'R_Round_Four_Significant_Strikes_Ground_Landed', 'R_Round_Four_Head_Significant_Strikes_Attempted', 'R_Round_Four_Head_Significant_Strikes_Landed', 'R_Round_Four_Body_Significant_Strikes_Attempted', 'R_Round_Four_Body_Significant_Strikes_Landed', 'R_Round_Four_Leg_Significant_Strikes_Attempted', 'R_Round_Four_Leg_Significant_Strikes_Landed', 'R_Round_Four_Total_Strikes_Attempted', 'R_Round_Four_Total_Strikes_Landed', 'R_Round_Four_Takedowns_Attempted', 'R_Round_Four_Takedowns_Landed', 'R_Round_Four_Takedown_Perc', 'R_Round_Four_Submission_Attempts', 'R_Round_Four_Grappling_Reversals', 'R_Round_Five_Knockdowns', 'R_Round_Five_Significant_Strikes_Landed', 'R_Round_Five_Significant_Strikes_Attempted', 'R_Round_Five_Significant_Strike_Perc', 'R_Round_Five_Significant_Strikes_Distance_Attempted', 'R_Round_Five_Significant_Strikes_Distance_Landed', 'R_Round_Five_Significant_Strikes_Clinch_Attempted', 'R_Round_Five_Significant_Strikes_Clinch_Landed', 'R_Round_Five_Significant_Strikes_Ground_Attempted', 'R_Round_Five_Significant_Strikes_Ground_Landed', 'R_Round_Five_Head_Significant_Strikes_Attempted', 'R_Round_Five_Head_Significant_Strikes_Landed', 'R_Round_Five_Body_Significant_Strikes_Attempted', 'R_Round_Five_Body_Significant_Strikes_Landed', 'R_Round_Five_Leg_Significant_Strikes_Attempted', 'R_Round_Five_Leg_Significant_Strikes_Landed', 'R_Round_Five_Total_Strikes_Attempted', 'R_Round_Five_Total_Strikes_Landed', 'R_Round_Five_Takedowns_Attempted', 'R_Round_Five_Takedowns_Landed', 'R_Round_Five_Takedown_Perc', 'R_Round_Five_Submission_Attempts', 'R_Round_Five_Grappling_Reversals'

Feature Engineering

Weight Classes

The UFC have different weight classes for each fight and was used to introduce new categorical features to our dataset.

Weight Class	Minimum Weight	Maximum Weight
Heavyweight	93	120
Light Heavyweight	83.9	93
Middleweight	77.1	83.9
Welterweight	70.3	77.1
Lightweight	65.8	70.3
Featherweight	61.2	65.8
Bantamweight	56.7	61.2
Flyweight	52.2	56.7
Strawweight*	0	52.2

Numerical & Categorical Transformations

Numerical value(s) transformation:

1. Replace missing values using the null values along each column, and adding a indicator for replacement of null Values

• SimpleImputer(strategy="constant", add_indicator=True)

2. Standardize features by removing the mean and scaling to unit variance

• StandardScaler()

Categorical value(s) transformation:

1. Encode categorical features as a one-hot numeric array

• OneHotEncoder(handle_unknown="ignore")

Training and Testing Sets

Multiple arrays are created from splitting the train and test subsets randomly. The training dataset contains 80% of the data, whereas the testing dataset contains 20%. Additionally, X represents the features and Y as the target variable.

Machine Learning Model Selection

The team determined that the machine learning model for implementation was the VotingClassifier ensemble with soft voting. The top five classifiers previously tested (based on accuracy score) were selected for inclusion in the voting ensemble. With soft voting, each classifier provides a probability value that a specific data point belongs to a particular target class (blue or red winner). The predictions are then added up, and the target label with the greatest sum of weighted probabilities wins the vote. Using VotingClassifier results in better performance than that of any of the five models used in the ensemble. However, one drawback of using this ensemble is that all the models equally contribute to the prediction, even though some might perform better than others.

Classifier	Balanced Accuracy Score	Precision	Precision_Blue	Precision_Red	Recall	Recall_Blue	Recall_Red	Parameters
VotingClassifier	0.907	0.907	0.901	0.910	0.907	0.818	0.954	*
XGBClassifier	0.899	0.898	0.881	0.907	0.899	0.813	0.943	random_state=0
SVC	0.896	0.896	0.890	0.898	0.896	0.792	0.950	random_state=0
GradientBoostingClassifier	0.896	0.895	0.880	0.903	0.896	0.805	0.943	random_state=0
Neural Net (MLPClassifier)	0.892	0.891	0.863	0.905	0.892	0.810	0.934	random_state=0
RandomForestClassifier	0.876	0.878	0.895	0.869	0.876	0.721	0.956	random_state=0
LogisticRegression	0.873	0.872	0.836	0.891	0.873	0.782	0.920	max_iter=1000, random_state=0
AdaBoostClassifier	0.873	0.872	0.841	0.888	0.873	0.774	0.924	random_state=0
BaggingClassifier	0.872	0.870	0.831	0.891	0.872	0.782	0.918	random_state=0
PassiveAggressiveClassifier	0.855	0.853	0.804	0.879	0.855	0.759	0.905	random_state=0
KNeighborsClassifier	0.852	0.851	0.806	0.874	0.852	0.746	0.907
DecisionTreeClassifier	0.817	0.816	0.744	0.853	0.817	0.708	0.874	random_state=0
RidgeClassifier	0.812	0.810	0.735	0.850	0.812	0.703	0.869	random_state=0

*VotingClassifier() Parameters:

VotingClassifier(
    estimators=[
        ("gbc", GradientBoostingClassifier(random_state=0)),
        ("rf", RandomForestClassifier(random_state=0)),
        ("mlp", MLPClassifier(random_state=0)),
        ("svc", SVC(random_state=0, probability=True)),
        ("xgb", XGBClassifier(random_state=0)),
    ],
    voting="soft")

• With default parameters, XGBClassifier has the highest accuracy score out of all classifiers.
• HyperParameter optimization will be the next goal for selecting the best model.

VotingClassifier Results

The top five models selected by accuracy are passed into a soft VotingClassifier ensemble:

1. XGBClassifier
2. SVC
3. GradientBoostingClassifier
4. Neural Net (MLPClassifier)
5. RandomForestClassifier

The idea behind the VotingClassifier is to combine conceptually different machine learning classifiers and use a majority vote or the average predicted probabilities (soft vote) to predict the class labels. Such a classifier can be useful for a set of equally well performing model in order to balance out their individual weaknesses. - SciKit-learn

Classification Report

	precision	recall	f1-score	support
Blue	0.90	0.82	0.86	390
Red	0.91	0.95	0.93	754
accuracy			0.91	1144
macro avg	0.91	0.89	0.89	1144
weighted avg	0.91	0.91	0.91	1144

Confusion Matrix

Dashboard

Ultimately, we chose to create our dashboard using the Streamlit library, an open-source, free, and Python-based framework for deploying data science projects. We initially discussed coding our dashboard directly with HTML/CSS/JS but ultimately agreed that this seemed too finicky for us. Streamlit allowed us to efficiently code our front-end entirely in its Python framework, freeing up more time to get our pipeline, database, and model to work well together with our interactive elements.

Subject to change, our interactive elements will include:

• Two drop-downs to allow a user to assign the fighters to model to either the red or blue corner.
• The above user inputs will also control the images displayed above our interactive elements.
• An ability for our users to create their own fighter by selecting values for key aggregated features to test a hypothetical fighter either again an established fighter or another hypothetical fighter.

In selecting these elements specifically, we are aiming to center our predictive model and keep the user-experience as streamlined we can.

You can view our deployed dashboard here: http://turtledashboard.ddns.net/

Future State

Looking ahead, we are focused most closely on improving feature selection to better hone our model's predictive capability. We are happy with where our "Upcoming Fights" modeling has ended up and want to parallel this success with a more accurate and robust "Fighter vs. Fighter" function.The next version of our app's "Fighter vs. Fighter" function will serve up the results from our "Upcoming Fights" dataset if a user selects a matchup that already exists in our database. Since we will have the most up-to-date statistics for each fighter and all fights, this method should serve up the most accurate prediction for the user.

Beyond feature selection and sharpening our "Fighter vs. Fighter" modeling, we also want to analyze our dataset to find the best features to aggregate for our "Create Your Own Fighter" function. Currently, our app hosts a framework for this function, and we need to narrow down the ~350 features per fighter into more manageable bins that our users would select from. E.g., instead of having a user select the number and type of significant strikes per round, we would combine significant strikes per round into a percentage, which we would then bin and allow the user to select as a feature of their fighter. Our aim is to retain the predictive ability of our model while maintaining a streamlined "Create Your Own Fighter" process.

Existing Coding & Research "To-Do's:"

Exploration/Transformation:

• Create other category for anything that does not fall in standardized Weight_Class.
• Determine why Max_Rounds being inferred as a object and not Ending_Round.
• Look into whether or not there is a benefit to using .reindex when sorting.
• Figure out if there is a better way to define gender than str.contains.
• Convert time features into more usable datatype.

Resources

• Source Code: UFC_Final_Project.ipynb

• Original Data: data.csv

  • Header Breakdown
    • B - Blue corner
    • R - Red corner
    • B-Prev - Previous wins of the fighter in the blue corner
    • R-Prev - Previous wins of the fighter in the red corner
    • Last_round - The round the fight ENDED
    • Max_round - Total rounds the fight was scheduled for
    • Height - Fighter height (cm)
    • Weight - Fighter weight (kg)
    • winby
      • DEC - Decision: Fight went all rounds and the judges decided the winner.
      • KO/TKO
        • Knockout (KO): Opponent was flatlined, out cold.
        • Technical Knockout (TKO): Opponent was not able to respond and the fight was stopped by the ref.
      • SUB - Submission: Opponent was submitted.
    • winner
      • Red - Fighter in the red corner won the fight.
      • Blue - Fighter in the blue corner won the fight.
      • No contest - No contest decisions in MMA are usually declared when an accidental illegal strike (the rules on which differ from each organization and state) causes the recipient of the blow to be unable to continue, that decision being made by the referee, doctor, the fighter or his corner.

• Scraped Data: scraped_data.csv

  • New columns
    • Height - Fighter height (in.)
    • Weight - Fighter weight (lbs.)
    • Accuracy - Accuracy column values are percentages.
    • Defense - Defense column values are percentages.

• Libraries: Pandas, Matplotlib, Scikit-Learn, Joblib, and XGBoost

• Database: pgAdmin 4 with SQLAlchemy and Psycopg2 libraries.

• Google Slides Presentation: UFC Fight Predictor