Proposed Flow

 - The Data

- Cleaning the data

- Doc2Vec model

- TF-idf model

- Comparison

- Application

-Project Overview

 Scrape The Data

 Clean data

 Doc2Vec model

 TF-idf model

Compare

Clustering

Metrics to Compare

Silhouette Score

Silhouette Score*

t-SNE Visualization*

t-SNE Visualization

Cosine Similarity

Mean Similarity*

Mean Similarity

f1 score*

Classification

f1 score

Clustering

Silhouette Score*

t-SNE Visualization*

Cosine Similarity

Mean Similarity*

f1 score*

Classification

The Data

-The data consists of subtitles  from 860 maths lectures scraped from youtube videos

Scraping the data

1. Find playlists on youtube

2. Youtube API call for  video ids from playlists

3. Store video ids

4. Use video id list with youtube -dl to download the subtitles

Get video ids

Download CCs, extract information

5. Convert files to csv

6. Store text from all csv files in one DataFrame

Cleaning the Data

Before

After

Inspecting the Data

Sources

Subjects

 Scrape The Data

 Clean data

 Doc2Vec model

 TF-idf model

Compare

Clustering

Metrics to Compare

Silhouette Score

Silhouette Score*

t-SNE Visualization*

t-SNE Visualization

Cosine Similarity

Mean Similarity*

Mean Similarity

f1 score*

Classification

f1 score

Clustering

Silhouette Score*

t-SNE Visualization*

Cosine Similarity

Mean Similarity*

f1 score*

Classification

Train a Doc2Vec Model

Clustering the Doc2Vec Vectors

KMeans

Agglomerative

Highest Silhouette Score: .2998

Number of Clusters: 25

Highest Silhouette Score: .2996

Number of Clusters: 36

Visualizing Clusters with t-SNE decomposition

K Means

Agglomerative Clustering

True Labels

 Scrape The Data

 Clean data

 Doc2Vec model

 TF-idf model

Compare

Clustering

Metrics to Compare

Silhouette Score

Silhouette Score*

t-SNE Visualization*

t-SNE Visualization

Cosine Similarity

Mean Similarity*

Mean Similarity

f1 score*

Classification

f1 score

Clustering

Silhouette Score*

t-SNE Visualization*

Cosine Similarity

Mean Similarity*

f1 score*

Classification

KMeans

.2998

Cosine Similarity Matrix

Lecture 1

Lecture 2

Lecture 3

Lecture m

x1

x2

x3

x100

Lecture 1

Lecture 1

Lecture 2

Lecture 3

Lecture 3

Lecture 2

Calculate Mean Similarity

Look up the most similar

Look at the *most* similar

Average Mean Similarity:  .1445

 Scrape The Data

 Clean data

 Doc2Vec model

 TF-idf model

Compare

Clustering

Metrics to Compare

Silhouette Score

Silhouette Score*

t-SNE Visualization*

t-SNE Visualization

Cosine Similarity

Mean Similarity*

Mean Similarity

f1 score*

Classification

f1 score

Clustering

Silhouette Score*

t-SNE Visualization*

Cosine Similarity

Mean Similarity*

f1 score*

Classification

KMeans

.2998

.1445

Document Classification with Doc2Vec

Creating testing data from the model

Use 'infer_vector' to generate test set vectors.

Test set vectors

Classification with  Doc2Vec

SVC = .71

Random Forest = .7

Gradient Boosting = .51

Logistic Regression = .95

Classification with  Doc2Vec

Logistic Regression

Highest f1 score: .94

 Scrape The Data

 Clean data

 Doc2Vec model

 TF-idf model

Compare

Clustering

Metrics to Compare

Silhouette Score

Silhouette Score*

t-SNE Visualization*

t-SNE Visualization

Cosine Similarity

Mean Similarity*

Mean Similarity

f1 score*

Classification

f1 score

Clustering

Silhouette Score*

t-SNE Visualization*

Cosine Similarity

Mean Similarity*

f1 score*

Classification

KMeans

.2998

.1445

.94

Tf-idf and Clustering

Try different values for min_df

for each tf-idf iteration, cluster vectors from 8-50 clusters

K Means

Agglomerative

Highest Silhouette Score: .1801

Highest Silhouette Score: .1297

Actual Labels

Cosine Similarity

Look up related items

Look at *most* similar

Average Mean Similarity: .065

 Scrape The Data

 Clean data

 Doc2Vec model

 TF-idf model

Compare

Clustering

Metrics to Compare

Silhouette Score

Silhouette Score*

t-SNE Visualization*

t-SNE Visualization

Cosine Similarity

Mean Similarity*

Mean Similarity

f1 score*

Classification

f1 score

Clustering

Silhouette Score*

t-SNE Visualization*

Cosine Similarity

Mean Similarity*

f1 score*

Classification

KMeans

.2998

.1445

.94

KMeans

.1800

.065

Classification modeling with tf-idf

Random Forest = .92

SVC = .95

Logistic Regression: .95

Gradient Boosting = .90

Classification with  Tf-idf Vectors

Logistic Regression

Highest f1 score: .95

 Scrape The Data

 Clean data

 Doc2Vec model

 TF-idf model

Compare

Clustering

Metrics to Compare

Silhouette Score

Silhouette Score*

t-SNE Visualization*

t-SNE Visualization

Cosine Similarity

Mean Similarity*

Mean Similarity

f1 score*

Classification

f1 score

Clustering

Silhouette Score*

t-SNE Visualization*

Cosine Similarity

Mean Similarity*

f1 score*

Classification

KMeans

.2998

.1445

.94

KMeans

.1800

.065

.95

Clustering

Doc2Vec

Tf-idf

Cosine Similarity

 .2998

 .1801

.065

.1445

Mean Similarity

Mean Similarity

Classification

 .94

 .95

Compare

Silhouette Score

Avg f1

Mean Similarity

Use Cases

Cosine Similarity is useful for  identifying items are similar to each other.

Clustering can be used on unlabeled data. This can be useful for determining resource allocation for new tasks.

Example:

"We have thousands of unsorted emails in our info@ourco.com inbox. We want to see how we can most efficiently categorize these emails"

Example:

"We want to deliver a premium experience for our users; we want a way to suggest additional products/items that our users are likely to consume."

Text Classification is useful subject identification and document routing.

Example:

"We want a way to efficiently assign helpdesk tickets to the person who is best fit to answer"

Conclusions

-Doc2Vec is better for clustering and establishing similarity

-Overall the tf-idf model outperformed the Doc2Vec model in the classification task.

Next Steps

-Collect more data

-Use a pre-trained word embedding as part of a deeper neural network to classify the texts

-Use LDA, NMF and other advanced NLP techniques to improve scores.