**Topic Modeling – Latent Semantic Analysis (LSA) and Singular Value Decomposition (SVD):**

- Singular Value Decomposition is a Linear Algebraic concept used in may areas such as machine learning (principal component analysis, Latent Semantic Analysis, Recommender Systems and word embedding), data mining and bioinformatics
- The technique decomposes given matrix into there matrices, let’s look at programmer’s perspective of the technique.

[codesyntax lang=”python”]

import numpy as np import pandas as pd

[/codesyntax]

#### Let’s take a small dataset and create Document-Term matrix.

- Note: Need to build Term-Document matrix For Query Retrieval in IR task.

[codesyntax lang=”python”]

documents = ["I enjoy Machine Learning", "I like NLP", "I like deep learning"] from sklearn.feature_extraction.text import CountVectorizer cv = CountVectorizer(ngram_range=(1, 1)) X = cv.fit_transform(documents).toarray()

[/codesyntax]

#### X is the Document-Term Matrix, “DOCUMENTS” – are the “rows” and “TOKENS” – are the “columns” of the matrix

Output:array([[0, 1, 1, 0, 1, 0], [0, 0, 0, 1, 0, 1], [1, 0, 1, 1, 0, 0]], dtype=int64)

[codesyntax lang=”python”]

cv.get_feature_names()

[/codesyntax]

Output:['deep', 'enjoy', 'learning', 'like', 'machine', 'nlp']

[codesyntax lang=”python”]

cv.vocabulary_

[/codesyntax]

Output:{'enjoy': 1, 'machine': 4, 'learning': 2, 'like': 3, 'nlp': 5, 'deep': 0}

##### Notation

- “*” – for Matrix multiplication
- V.T – to denote Transpose of matrix V

#### Now SVD will decompose rectangular matrix X into there matrices. Namely U, S and V.T (V transpose)

- U – the left Eigen vector (orthogonal) matrix, the vectors extract “Document-to-Document similarity”.
- V – the right Eigen vector (orthogonal) matrix, the vectors extract “Strength of Two Different Words Occurring in one Document” (may or may-not be co-occurrence)
- S – the Eigen Values (diagonal) matrix, explains variance ratio.

#### Before we understand how U, V, S extract relation strengh or explain variance ratio. Let’s understand on how to decompose X.

- To decomposing any matrix, we should use Eigen Vectors, Eigen Values decomposition.
- A matrix is
**eligible**to apply Eigen decomposition only when it is a**square and symmetric matrix*. - We can get a SQUARE, SYMMETRIC matrix by multiplying “X.T with X” or “X with X.T”, we can decompose these (np.dot(X.T,X) and np.dot(X,X.T) and get U, S, V matrixes.

#### Why to multiply “X.T with X” or “X with X.T”?, Why to decompose “X.T with X” or “X with X.T”?

**Suppose that X is decomposed as U S V.T****i.e., X = U * S * V.T**

- Now let’s see what is X.T multiplied by X
- X.T * X = (U * S * V.T).T * (U * S * V.T)
- = (V * S.T * U.T) * (U * S * V)
- = V * S.T * (U.T * U) * S * V)
- = V * S.T * (I) * S * V # as U is an orthogonal matrix U.T * U = I
- = V * (S.T * S) * V
- = V * S^2 * V # as S is diagnoal matrix S.T * S is S-square

- We got V * S^2 * V, when we multiply X.T with X, where X is in decomposed form (U * S * V.T),
**Hence, when we decompose X.T * X, we get values of matrix V and S^2** - Now let’s see what is X multiplied by X.T
- X * X.T = (U * S * V.T) * (U * S * V.T).T
- = (U * S * V) * (V.T * S.T * U)
- = (U * S * (V * V.T) * S.T * U)
- = U * S * (I) * S.T * V # as U is an orthogonal matrix U.T * U = I
- = U * (S.T * S) * U
- = U * S^2 * U # as S is diagnoal matrix S * S.T is S-square

- We got U * S^2 * U, when we multiply X.T with X, where X is in decomposed form (U * S * V.T),
**Hence, when we decompose X * X.T, we get values of matrix U and S^2** **Bottom line: To find U, S and V, we need to find Eigen Decoposition of X.T * X and X * X.T**

#### X.T * X is word similarity matrix, the (i, j) position in X.T * X will quantify the overlap of Word-i and Word-j. Number of Documents in which both words i and j occur. (This inference is based out of Document-Term matrix, if it is Tf-IDF, then the inference would be different)

[codesyntax lang=”python”]

np.dot(X.T, X)

[/codesyntax]

Output:array([[1, 0, 1, 1, 0, 0], [0, 1, 1, 0, 1, 0], [1, 1, 2, 1, 1, 0], [1, 0, 1, 2, 0, 1], [0, 1, 1, 0, 1, 0], [0, 0, 0, 1, 0, 1]], dtype=int64)

#### X * X.T is document similarity matrix, the (i, j) position in X * X.T will quantify the similarity of Document-i and Document-j. This analogy suits best with Tf-IDF scores.

[codesyntax lang=”python”]

np.dot(X, X.T)

[/codesyntax]

Output:array([[3, 0, 1], [0, 2, 1], [1, 1, 3]], dtype=int64)

#### The rows in sigular vector matrix V.T are the topics and values in the row denote strenght of words in the topic.

[codesyntax lang=”python”]

U, s, Vh = np.linalg.svd(X) terms = cv.vocabulary_ for i, component in enumerate(Vh): terms_components = zip(terms, component) sorted_terms = sorted(terms_components, key=lambda x:x[1], reverse=True)[:3] # take features for topic print("topic : ", i) for term_socres in sorted_terms: print(10*" ", term_socres[0]) print(50*'*')

[/codesyntax]

Output:topic : 0 deep nlp machine ************************************************** topic : 1 nlp machine learning ************************************************** topic : 2 enjoy learning like ************************************************** topic : 3 nlp enjoy like ************************************************** topic : 4 nlp learning like ************************************************** topic : 5 deep enjoy nlp **************************************************

**The above result may not make sense as the dataset is too small. If we take decent amount data, the results will be promising**

#### Cons:

- SVD is a linear decomposition, hence cannot extract non-linear patterns
- SVD computation is very expensive, cannot handle huge volumens of data (Randomized SVD is one of the solutions)

–By

L Venkata Rama Raju