TED is devoted to spreading powerful ideas on just about any topic. These datasets contain over 4,000 TED talks including transcripts in many languages Founded in 1984 by Richard Salman as a nonprofit organization that aimed at bringing experts from the fields of Technology, Entertainment, and Design together, TED Conferences have gone on to become the Mecca of ideas from virtually all walks of life. As of 2015, TED and its sister TEDx chapters have published more than 2000 talks for free consumption by the masses and its speaker list boasts of the likes of Al Gore, Jimmy Wales, Shahrukh Khan, and Bill Gates.
Number of records: 4,005
Number of attributes: 19
The dataset contains features like:
- talk_id: Talk identification number provided by TED
- title: Title of the talk
- speaker_1: First speaker in TED's speaker list
- all_speakers: Speakers in the talk
- occupations: Occupations of the speakers
- about_speakers: Blurb about each speaker
- recorded_date: Date the talk was recorded
- published_date: Date the talk was published to TED.com
- event: Event or medium in which the talk was given
- native_lang: Language the talk was given in
- available_lang: All available languages (lang_code) for a talk
- comments: Count of comments
- duration: Duration in seconds
- topics: Related tags or topics for the talk
- related_talks: Related talks (key='talk_id',value='title')
- url: URL of the talk
- description: Description of the talk
- transcript: Full transcript of the talk
- views: Contains Count of views of every talk
- The main objective is to build a predictive model, which could help in predicting the views of the videos uploaded on the TEDx website.
- Understanding of ML algorithms
- IDE- Google colab
-
Importing Libraries
-
Loading the dataset
-
EDA on features
-
Feature Engineering
-
Data Cleaning
-
Feature Selection
-
HyperParameter Tuning and Modeling
-
Evaluation and comparision of models
-
Selecting the best model
-
Conclusion
Started with loading the data so far we have done EDA ,feature engineering , data cleaning, target encoding and one hot encoding of categorical columns, feature selection and then model building. So far we have modelled on
- Lasso Regressor
- Ridge Regressor
- KNearestNeighbors Regressor
- Random Forest Regressor
- XGB Regressor
In all of these models our errors have been in the range of 2,00,000 which is around 10% of the average views. We have been able to correctly predict views 90% of the time.
After hyper parameter tuning, we have prevented overfitting and decreased errors by regularizing and reducing learning rate.
Given that only have 10% errors, our models have performed very well on unseen data due to various factors like feature selection,correct model selection,etc. Out of all these models RandomForestRegressor is the best performer in terms of MAE.
In all the features speaker_wise_avg_views is most important this implies that speakers are directly impacting the views.
