ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

MODULE 6

Sentiment Analysis,
Opinion Mining &
Spam Detection
Advanced Data Mining (ADM) | IT Engineering — Semester VI

Exam-Focused Study Guide & Notes

Feature Details

Topics Covered Sentiment Analysis, Opinion Mining, Spam Detection

Exam Questions 9 Questions with Model Answers

Marks Covered 5-Mark & 10-Mark Format

Special Sections Exam Tips, Common Mistakes, Quick Revision, Last-Day Sheet

■ HOW TO
This document covers ALL concepts from ADM Module 6. Read the Quick Revision
USE THIS
Summaries before exams. Highlighted boxes = guaranteed exam material.
GUIDE

VESIT IT Engineering Page 1 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ TABLE OF CONTENTS
1 Sentiment Analysis — Complete Process

2 Document-Level Sentiment Classification

3 Sentence-Level Sentiment Classification

4 Opinion Mining — Definition & Challenges

5 Opinion Words — Dictionary & Corpus-Based Approaches

6 Types of Spam & Spammers

7 Hiding Techniques Used by Spammers

8 Group Spam Detection Algorithm

9 Abnormal Behavior Patterns for Spam Detection

— EXAM QUESTIONS WITH MODEL ANSWERS (Q1–Q9)

— Last-Day Revision Sheet

— Important Keywords List

— Short Notes (5 Marks) & Long Answer Format (10 Marks)

VESIT IT Engineering Page 2 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ SECTION 1: SENTIMENT ANALYSIS

1.1 What is Sentiment Analysis?

Also called Opinion Mining, it is the computational study of people's
opinions, sentiments, attitudes, and emotions expressed in text
■ Sentiment Analysis
about entities such as products, services, topics, individuals, and
events.

Sentiment analysis is a field at the intersection of Natural Language Processing (NLP), Computational
Linguistics, and Text Mining.

1.2 Complete Process of Sentiment Analysis

The complete pipeline has 7 key steps:

Step Name Description

Gather text data from sources like reviews, tweets, forums, news articles
Step 1 Data Collection
using web scraping or APIs.

Clean the text: remove HTML tags, stopwords, punctuation; perform

Step 2 Pre-processing
tokenization, stemming, and lemmatization.

Convert text to numerical features using Bag-of-Words (BoW), TF-IDF, or

Step 3 Feature Extraction
Word Embeddings (Word2Vec).

Sentiment Classify text as Positive, Negative, or Neutral using ML models (Naive Bayes,
Step 4
Classification SVM, LR) or Lexicon-based methods.

Identify the specific aspect/feature of an entity about which sentiment is

Step 5 Aspect Extraction
expressed (e.g., "battery life" of a phone).

Aggregation &
Step 6 Combine results and present via dashboards, charts, or sentiment scores.
Visualization

Use insights for business decisions: product improvement, customer service,

Step 7 Application
market research.

1.2.1 Process Flow Diagram

VESIT IT Engineering Page 3 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

Raw Text Input

Pre-processing
(Tokenize, Remove Stopwords, Stem/Lemmatize)

Feature Extraction
(TF-IDF / BoW / Embeddings)

Sentiment Classification
(ML Model OR Lexicon-based)

Result: Positive / Negative / Neutral

1.3 Levels of Sentiment Analysis

Level Description Example

Classifies entire document as A product review is overall positive or

Document Level
positive/negative/neutral. negative.

Sentence Level Each sentence classified separately. "Battery is great but camera is poor."

Positive about price, Negative about

Aspect/Feature Level Sentiment about specific aspects of an entity.
delivery.

Word/Phrase Level Determines polarity of individual words/phrases. "Excellent", "terrible", "not bad"

1.4 Sentiment Classification Approaches

Approach How it works Examples

Naive Bayes, SVM, Logistic

Machine Learning Supervised (labeled data) — trains classifier on features.
Regression

SentiWordNet, VADER,
Lexicon-based Uses sentiment dictionary (opinion words list).
AFINN

Hybrid Combines ML + Lexicon for better accuracy. Most production systems

For a 10-mark answer on sentiment analysis PROCESS, always include: Data Collection
■ EXAM
→ Pre-processing → Feature Extraction → Classification → Aspect Extraction →
TIP
Aggregation. Draw the flow diagram for extra marks.

■ QUICK REVISION SUMMARY

✔ Sentiment Analysis = Computational study of opinions, emotions in text

✔ 7 Steps: Collect → Preprocess → Feature Extract → Classify → Aspect Extract → Aggregate →

Apply

VESIT IT Engineering Page 4 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

✔ 3 Levels: Document, Sentence, Aspect/Feature

✔ 3 Approaches: Machine Learning, Lexicon-based, Hybrid

✔ Key tasks: Opinion extraction, sentiment classification, aspect detection

VESIT IT Engineering Page 5 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ SECTION 2: DOCUMENT SENTIMENT CLASSIFICATION

2.1 Definition
■ Document The task of classifying an entire document (e.g., review, article) as
Sentiment expressing positive, negative, or neutral sentiment. It assumes the
Classification document focuses on a single entity/topic.

2.2 Supervised Learning Approach

Supervised learning requires labeled training data — documents pre-tagged as positive or negative — to train
a classification model.

Step Phase Details

1 Data Collection Collect labeled reviews/documents (e.g., 5-star = positive, 1-star = negative)

2 Feature Extraction Extract features: word frequencies, TF-IDF weights, n-grams, POS tags

3 Model Training Train classifier (Naive Bayes, SVM, MaxEnt) on labeled feature vectors

4 Evaluation Test model using held-out test set; measure Accuracy, Precision, Recall, F1

5 Prediction Apply model to classify new/unseen documents

2.2.1 Common Features Used

Feature Type What it captures Example/Note

Unigrams Individual words "good", "bad", "excellent"

Bigrams/Trigrams Word pairs/triples "not good", "very bad"

TF-IDF Term importance score Downweights common words

POS Tags Part-of-speech info Adjectives are most sentiment-bearing

Negation Handles "not" + word "not happy" → negative

Frequency Word count vectors Bag-of-Words model

2.2.2 Popular ML Classifiers

Classifier Key Strength Details

Naive Bayes Probabilistic; fast; works well with text P(class|words) using Bayes theorem

Finds optimal hyperplane separating

SVM High accuracy, handles high dimensions
classes

Logistic Regression Linear model for probability of class Interpretable, works well with TF-IDF

Maximum Entropy Flexible probabilistic model Similar to Logistic Regression

Deep Learning Captures semantic meaning and

State-of-the-art but needs large data
(LSTM/BERT) context

2.3 Unsupervised Learning Approach

VESIT IT Engineering Page 6 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

Unsupervised learning does NOT require labeled data. It uses semantic orientation of words and phrases to
determine sentiment polarity.

2.3.1 Turney's Semantic Orientation Method (Key Algorithm)

This is one of the most important unsupervised methods. Semantic Orientation (SO)
■ KEY
measures how positive or negative a phrase is based on its association with known
CONCEPT
positive/negative seed words.

Step Action Detail

1 POS Tagging Tag all words with Part-of-Speech (noun, verb, adjective...)

2 Extract Phrases Extract adjective phrases and adverb+adjective patterns

Compute Pointwise Mutual Information (PMI) of each phrase with "excellent" and
3 Calculate PMI
"poor"

4 Compute SO SO = PMI(phrase, "excellent") - PMI(phrase, "poor")

5 Classify SO > 0 → Positive; SO < 0 → Negative; SO ≈ 0 → Neutral

6 Aggregate Average SO of all phrases → classify whole review

■ PMI (Pointwise PMI(phrase, word) = log■ [P(phrase AND word) / (P(phrase) ×

Mutual Information) P(word))]. High PMI = the two terms co-occur more than by chance.

■ Semantic SO(phrase) = PMI(phrase, "excellent") − PMI(phrase, "poor"). Positive

Orientation (SO) SO → positive phrase; Negative SO → negative phrase.

2.4 Comparison: Supervised vs Unsupervised

Aspect Supervised Unsupervised

Labeled Data Required (Training set) Not Required

Accuracy Generally Higher Lower but domain-independent

Domain Dependency Domain-specific; needs retraining Works across domains

Scalability Expensive to label data Easily scalable

Key Methods Naive Bayes, SVM, LR SO-PMI, Lexicon-based

Best Used When Labeled data is available No labeled data; new domain

In exams, always explain BOTH approaches for a 10-mark question. Include Turney's
■ EXAM
SO-PMI algorithm steps for unsupervised approach. Use the comparison table to end
TIP
your answer.

■■ COMMON Don't confuse "document-level" with "sentence-level" classification.

MISTAKE Document-level treats the whole document as ONE unit.

■ QUICK REVISION SUMMARY

✔ Document classification = classify entire document as +ve/-ve/neutral

✔ Supervised: needs labeled data → Train model → Predict

VESIT IT Engineering Page 7 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

✔ Unsupervised: Turney's SO-PMI — no labeled data needed

✔ SO = PMI(phrase, "excellent") - PMI(phrase, "poor")

✔ SO > 0 → Positive; SO < 0 → Negative

✔ Common classifiers: Naive Bayes, SVM, Logistic Regression

VESIT IT Engineering Page 8 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ SECTION 3: SENTENCE-LEVEL SENTIMENT

CLASSIFICATION
Instead of classifying the entire document, each individual sentence
■ Sentence-Level is classified as subjective or objective, and then as
Classification positive/negative/neutral. Also called Fine-grained sentiment
analysis.

3.1 Two-Stage Process

Sentence-level classification has two main stages:

Stage Name What it does

Determine if the sentence expresses an opinion (subjective) or states a fact

Stage 1 Subjectivity Detection
(objective). Only subjective sentences are analyzed further.

Stage 2 Sentiment Polarity For subjective sentences, classify as Positive, Negative, or Neutral.

3.2 Subjective vs Objective Sentences

Type Description Examples

Subjective Expresses opinion, feeling, belief "The camera quality is amazing!" / "I hate this product."

"The phone weighs 180 grams." / "The product was

Objective States verifiable facts
released in 2023."

3.3 Methods for Sentence-Level Classification

3.3.1 Lexicon-based Method

• Uses a sentiment lexicon (opinion word dictionary) with polarity scores
• Each word in the sentence is matched with the lexicon
• Sum of polarity scores → overall sentence sentiment
• Example: "The food was delicious but service was terrible."
– "delicious" = +2, "terrible" = -2 → Net = 0 → Neutral

3.3.2 Machine Learning Method

• Treat each sentence as a document and train a classifier
• Features: n-grams, dependency parse features, SentiWordNet scores
• Challenge: Sentences are short — sparse features

3.3.3 Handling Special Cases in Sentence-level

Case Example Handling Strategy

Negation "not good" → negative Flip polarity for words after "not/never/no"

Sarcasm/Irony "Oh great, another bug!" → negative Hard to detect; needs context or deep learning

Conditional "If this works, it would be great" →

Conditional = not expressing current opinion
Sentences Neutral

VESIT IT Engineering Page 9 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

Comparison "A is better than B" → +A, -B Extract comparative sentiment

"Is this product good?" → No

Questions Questions are typically objective
sentiment

3.4 Comparison: Document vs Sentence Level

Aspect Document Level Sentence Level

Granularity Entire document Each sentence

Assumption One topic/entity per doc Multiple topics allowed

Accuracy May miss contradictions More precise

Complexity Simpler More complex

Example "The product is great overall" "Battery great. Camera bad. Price fair."

■ EXAM Mention BOTH stages (subjectivity detection + polarity) for full marks. Include examples
TIP of subjective vs objective sentences.

■ QUICK REVISION SUMMARY

✔ Sentence-level: each sentence classified separately

✔ Stage 1: Subjectivity Detection (subjective vs objective)

✔ Stage 2: Polarity Classification (+ve / -ve / neutral)

✔ Key challenge: Negation, Sarcasm, Conditional sentences

✔ Methods: Lexicon-based (word scores) + ML (train sentence classifier)

VESIT IT Engineering Page 10 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ SECTION 4: OPINION MINING

4.1 Definition & Key Concepts

Also called Sentiment Analysis, it is the computational study of
■ Opinion Mining opinions, sentiments, evaluations, attitudes, and emotions of
people toward entities, events, topics, and their attributes.

An opinion is a quintuple: (entity, aspect, sentiment, opinion holder,

■ Opinion time) where: Entity = thing being reviewed; Aspect = feature of entity;
Sentiment = polarity; Opinion holder = who expressed it; Time = when.

Example Breakdown
Entity iPhone 15

Aspect Battery life

Sentiment Positive

Opinion Holder John (reviewer)

Time March 2024

Full Opinion "John thinks iPhone 15 battery life is excellent (March 2024)"

4.2 Challenges of Opinion Mining

■ EXAM
This is a HIGH-PRIORITY exam topic. Learn ALL challenges with examples.
PRIORITY

1 Language Complexity "Oh wonderful, another error!" — sarcastic but

Natural language is highly complex. Sarcasm, metaphors, looks positive
and idioms can reverse meaning.

2 Domain Dependency "Unpredictable" — negative (car), positive

Opinion words vary across domains. "Unpredictable" is (thriller movie)
negative for a car but positive for a movie plot.

3 Comparative Opinions "Samsung has a better camera than Apple" —

Sentences comparing two entities are hard to classify simply relative opinion
as +ve or -ve.

4 Implicit Opinions "The phone died after 3 hours of use" —

Opinions expressed without explicit sentiment words. negative, but no negative word used
Context required.

5 Spam & Fake Reviews Hotels hiring people to post fake 5-star reviews
Many reviews are fake, spam, or manipulated. Hard to detect
authentic opinions.

6 Cross-lingual Challenges Indian product reviews in Hindi, Tamil, Hinglish

Opinion mining across multiple languages needs multilingual
resources.

7 Aspect Identification "It fits perfectly" — does "it" refer to size?

Extracting the specific aspect being discussed is non-trivial. design? material?

VESIT IT Engineering Page 11 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

8 Negation Handling "not bad", "never disappoints" — positive

Negation words flip the sentiment of a phrase. despite negative words

9 Subjectivity Detection "The screen is 6.1 inches" (fact) vs "The screen

Distinguishing objective facts from subjective opinions is looks great" (opinion)
difficult.

10 Temporal Opinions Product reviews right after launch vs 1 year

Opinions may change over time and must be tracked later
chronologically.

■ EXAM For 10 marks: Write at least 6-7 challenges with examples. Sarcasm, Domain
TIP Dependency, Implicit Opinions, and Spam are the most commonly asked ones.

■ QUICK REVISION SUMMARY

✔ Opinion = (entity, aspect, sentiment, holder, time) — the quintuple

✔ 10 Major Challenges: Language Complexity, Domain Dependency, Comparatives, Implicit, Spam,

Cross-lingual, Aspect ID, Negation, Subjectivity, Temporal

✔ Key: Sarcasm reverses polarity — "Oh great!" can be negative

✔ Domain dependency: same word = different polarity in different domains

✔ Implicit opinions have no direct sentiment words — requires inference

VESIT IT Engineering Page 12 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ SECTION 5: OPINION WORDS — DICTIONARY &

CORPUS-BASED APPROACHES
Also called sentiment words or polarity words, these are words that
■ Opinion Words
convey positive or negative sentiment. Examples: "excellent", "terrible",
(Opinion Lexicon)
"beautiful", "ugly". They are stored in a sentiment lexicon/dictionary.

5.1 Types of Opinion Words

Type Examples

Positive Words "excellent", "amazing", "fantastic", "love", "brilliant"

Negative Words "terrible", "horrible", "hate", "awful", "dreadful"

Negators "not", "never", "no", "hardly" — flip polarity of adjacent words

Intensifiers "very", "extremely", "incredibly" — amplify polarity

Reducers "somewhat", "slightly", "barely" — reduce polarity

5.2 Dictionary-Based Approach

Core Idea: Start with a small set of seed words and expand the opinion lexicon using a thesaurus (like
WordNet) to find synonyms and antonyms.

5.2.1 Steps in Dictionary-Based Approach

Step Action Detail

Start with a manually compiled small list of known positive words (e.g., "good", "nice")
1 Seed Words
and negative words (e.g., "bad", "poor").

Synonym Use a thesaurus (WordNet) to find synonyms of seed words. Synonyms inherit the
2
Expansion same polarity.

Antonym
3 Find antonyms of seed words. Antonyms get the OPPOSITE polarity.
Expansion

Repeat steps 2-3 with newly added words until no new words are found
4 Iterate
(convergence).

5 Manual Pruning Remove words that do not make sense as opinion words. Clean up the lexicon.

5.2.2 Example
Seed (+ve): "good" Synonym → also +ve: "excellent", "wonderful", "great"

Antonym → -ve: "bad" Synonym of "bad" → also -ve: "terrible", "awful", "poor"

5.2.3 Advantages & Disadvantages of Dictionary-Based

■ Advantage Language-independent (works with any thesaurus)

■ Advantage Easy to implement; no training data needed

■ Disadvantage Cannot capture domain-specific opinion words

VESIT IT Engineering Page 13 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ Disadvantage Thesaurus may be incomplete; misses slang and new words

■ Disadvantage Polarity errors can propagate through synonym chains

5.3 Corpus-Based Approach

Core Idea: Use a large corpus of text and statistical/syntactic patterns to find opinion words and their polarity.
Relies on the assumption that words appearing in similar contexts share similar sentiments.

5.3.1 Two Sub-Approaches

A) Statistical / Corpus Co-occurrence Approach:
• Uses Pointwise Mutual Information (PMI) to measure association between candidate words and seed words
• If a word co-occurs frequently with positive seeds → it is positive
• SO(word) = PMI(word, positive_seed) - PMI(word, negative_seed)
B) Syntactic Pattern-based Approach:
• Uses grammatical/syntactic rules to find opinion words
• Key rule: Words connected by "and" typically share the same polarity
• Words connected by "but" typically have opposite polarity
• Example: "The phone is elegant and stylish" → both "elegant" and "stylish" are +ve
• Example: "The food was cheap but tasteless" → "cheap" = different polarity from "tasteless"

5.3.2 Steps in Corpus-Based Approach

Step Action Detail

1 Select Seed Words Choose small set of known positive/negative words

2 Parse Corpus POS-tag and dependency-parse the corpus

3 Extract Patterns Find adjectives connected by conjunctions (and/but/or)

4 Assign Polarity Propagate polarity via conjunction rules (and=same, but=opposite)

5 Calculate SO Use PMI to compute semantic orientation scores

6 Build Lexicon Compile final opinion word list with polarities

5.4 Comparison: Dictionary-Based vs Corpus-Based

Aspect Dictionary-Based Corpus-Based

Data Required Thesaurus (WordNet) Large text corpus

Training Data Not required Not required (unsupervised)

Domain Specificity Domain-independent Can be domain-specific

New Words Fails for slang/new words Can discover new opinion words

Accuracy Moderate Higher for domain-specific

Method Used Synonym/Antonym expansion PMI + Conjunction rules

Key Tool WordNet / Thesaurus Corpus + Statistical methods

VESIT IT Engineering Page 14 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ EXAM For 5 marks: Explain ONE approach with steps. For 10 marks: Explain BOTH with
TIP comparison table. Always mention seed words and expansion method.

■■ COMMON Don't say corpus-based approach needs labeled data — it is UNSUPERVISED.

MISTAKE The key insight is statistical co-occurrence with seed words.

■ QUICK REVISION SUMMARY

✔ Opinion words = words conveying +ve or -ve sentiment

✔ Dictionary-based: seed words → synonym/antonym expansion via WordNet → iterate

✔ Corpus-based: seed words + PMI on large corpus + conjunction rules (and=same, but=opposite)

✔ Dictionary-based: simple but misses domain-specific and new words

✔ Corpus-based: finds domain-specific words but needs large corpus

VESIT IT Engineering Page 15 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ SECTION 6: TYPES OF SPAM & SPAMMERS

Fake or deceptive reviews written with the intention of misleading
consumers or manipulating the reputation of a
■ Review Spam
product/service/business. It is a serious problem for e-commerce and
review platforms like Amazon, Yelp, TripAdvisor.

6.1 Types of Spam Reviews

1 Untruthful Opinion / Fake Reviews "Excellent product, best I ever bought!" —
Positive or negative fake reviews written to promote or written by a paid fake reviewer
demote a target (product, service, business).

2 Brand/Domain Reviews "Company XYZ is amazing!" without

Reviews that only promote the brand as a whole rather than mentioning any product
reviewing any specific product. Off-topic.

3 Non-reviews "How do I return this?" or "Visit [Link]

Text that is NOT a review at all — questions, ads, spam for deals!"
emails, irrelevant posts mistakenly classified as reviews.

4 Incentivized Reviews Amazon sellers offering free products for 5-star

Reviews written by real users but in exchange for money, reviews
free products, or discounts. Biased opinions.

5 Duplicate/Near-Duplicate Reviews Same 5-star review posted across 100

Same or slightly modified reviews posted multiple times by products
the same spammer across platforms.

6 Early Reviews Review posted within 1 hour of a product going

Reviews posted too soon after product launch — before the live
reviewer could have actually used the product.

6.2 Types of Spammers

Individual Spammer A single person who posts fake reviews. A competitor posting negative reviews of a
May use multiple accounts (sockpuppets). rival's product

Group Spammer An organized group working together to A "review farm" of 50 people each posting
collectively spam a target. Far more 10 fake 5-star reviews for a hotel
dangerous.

Crowdsourced Spammer Uses a crowd/marketplace (like Paying $1 per fake review on Fiverr to
Mechanical Turk) to hire people to write hundreds of workers
fake reviews.

Bot-based Spammer Uses automated bots/scripts to generate A script posting 1000 identical 5-star
and post large volumes of fake reviews. reviews with minor wording variations

Incentivized Reviewer Ordinary users biased by incentives (free Amazon Vine program participants writing
products, discounts, cash) to write only positive reviews
positive reviews.

6.3 Characteristics That Identify Spammers

Characteristic Description

VESIT IT Engineering Page 16 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

Posting Frequency Spammers post unusually high number of reviews in short time

Rating Pattern Always give maximum (5★) or minimum (1★) ratings; no moderate ratings

Review Similarity Reviews very similar across products or accounts

New Account Account recently created; very few reviews except for spam targets

Single Product Focus All reviews concentrated on one product or competitor

Short Tenure Account active for very short time then abandoned

No Social Ties No followers, friends, or community connections — isolated accounts

Remember 6 types of spam reviews and 5 types of spammers. Group spammers are
■ EXAM
MOST dangerous as they coordinate attacks. Individual vs Group spammer distinction is
TIP
frequently tested.

■ QUICK REVISION SUMMARY

✔ 6 Types of spam: Fake reviews, Brand reviews, Non-reviews, Incentivized, Duplicate, Early

✔ 5 Types of spammers: Individual, Group, Crowdsourced, Bot-based, Incentivized reviewer

✔ Group spammers = most difficult to detect — coordinated attack

✔ Spammer signatures: extreme ratings, high post frequency, new accounts, similar content

VESIT IT Engineering Page 17 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ SECTION 7: HIDING TECHNIQUES USED BY

SPAMMERS
Spammers use various techniques to evade detection and make their fake reviews appear legitimate. These
techniques differ for individual spammers and group spammers.

7.1 Individual Spammer Hiding Techniques

1 Mixing Genuine and Spam Reviews Posts 40 genuine reviews over months, then 10
Individual spammers write a mix of real, genuine reviews for fake reviews for target
other products along with their spam reviews. This makes
their profile look authentic.

2 Posting from Different Accounts (Sock Puppets) Creates 20 fake Gmail accounts and posts from
Creates multiple fake accounts (sockpuppets) to post spam each
from. Each account has different username, email, and
review history.

3 Timing Variation Posts 1 fake review per week instead of 10 in

Distributes spam reviews over time rather than posting all at one day
once. Mimics natural review behavior.

4 Content Variation "Excellent product!" vs "Amazing item!" vs

Manually rewrites reviews in different words to avoid "Wonderful purchase!"
duplicate detection algorithms. Paraphrases or uses
synonyms.

5 Rating Variation Gives target product 5 stars; gives competitor 2

Instead of always giving 5 stars, occasionally gives 3-4 stars stars
to appear unbiased.

6 Using VPNs / Different IPs Uses Tor or rotating proxies to post from
Uses VPN or different IP addresses to avoid IP-based different locations
detection systems.

7 Delayed Posting Creates account → waits 3 months → starts

Waits before posting all spam reviews; creates account spamming
weeks before using it for spam.

7.2 Group Spammer Hiding Techniques

1 Time Coordination Group leader assigns each member a specific
Group members coordinate to spread their spam reviews time slot to post
over different time windows. They do NOT all post
simultaneously — that would trigger burst detection.

2 Role Assignment Member A: Reputation builder; Member B:

Some members spam the target product; others build Spam poster; Member C: Account creator
reputation by posting genuine reviews elsewhere.
Specialized roles make the group harder to identify.

3 Separate IP Addresses 20 members each using different cities/VPNs

Each group member posts from different IP/device/location to
avoid IP clustering detection.

VESIT IT Engineering Page 18 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

4 Mixed Legitimate Reviews Group posts 500 genuine reviews + 50 targeted

Group members collectively post legitimate reviews for spam reviews
unrelated products across different domains.

5 Review Content Diversification Template: "Great product. [custom sentence].

The group uses a template but each member personalizes Highly recommend." — each member fills
the content to avoid duplicate detection. [custom sentence]

6 Staggered Account Creation One new account created per day over 30 days
Accounts are created at different times — not all at once —
to avoid detection by account creation burst patterns.

7 Attacking via Multiple Platforms Coordinated attack across 4 platforms

Group simultaneously attacks a target on Yelp, Google simultaneously
Reviews, Amazon, TripAdvisor to maximize impact.

7.3 Comparison: Individual vs Group Hiding Techniques

Aspect Individual Spammer Group Spammer

Scale Single person Organized team

Sophistication Moderate High — coordinated strategy

Detection Difficulty Easier to detect Much harder to detect

Main Technique Multiple accounts + genuine mix Role assignment + time coordination

Impact Limited Massive — can change product ranking

■ EXAM List at least 5 individual techniques and 5 group techniques for 10 marks. Always explain
TIP WHY each technique helps evade detection.

■■ COMMON Don't confuse hiding techniques with spam types. Hiding techniques = methods to
MISTAKE AVOID BEING CAUGHT.

■ QUICK REVISION SUMMARY

✔ Individual hiding: Genuine mix, Sock puppets, Timing variation, Content variation, VPN, Delayed
posting

✔ Group hiding: Time coordination, Role assignment, Separate IPs, Mixed genuine reviews, Content
diversification, Staggered accounts

✔ Group techniques are MORE sophisticated and harder to detect

✔ Key insight: Spammers mimic normal user behavior patterns to avoid detection

VESIT IT Engineering Page 19 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ SECTION 8: GROUP SPAM DETECTION ALGORITHM

The process of identifying groups of reviewers who collaborate to
■ Group Spam post fake reviews together. It analyzes reviewer behavior patterns
Detection collectively rather than individually to expose coordinated spam
campaigns.

■ WHY Group spam is much harder to detect than individual spam. The algorithm looks for
DIFFICULT PATTERNS across multiple reviewers working together.

8.1 Key Intuition Behind Group Detection

• Genuine reviewers rarely review the same set of products
• Spammer groups tend to review the SAME products together
• The more two reviewers share reviewed products, the more suspicious they are
• Groups will show abnormally high review activity for the SAME target in the same time period

8.2 Steps in the Group Spam Detection Algorithm

The algorithm works in 4 main phases:

Phase 1: Data Collection & Preprocessing

• Collect all reviews from the platform (Amazon/Yelp/etc.)
• Extract: reviewer ID, product ID, rating, review date, review text
• Build a reviewer-product matrix: rows = reviewers, columns = products reviewed

Phase 2: Candidate Group Generation

• Find all pairs of reviewers who reviewed the SAME products
• Compute a co-review score for each pair based on number of shared products
• Use clustering algorithms (e.g., K-means or hierarchical clustering) to group similar reviewers
• Candidate groups = clusters of reviewers with high mutual co-review scores

Phase 3: Feature Extraction for Groups

• For each candidate group, extract behavioral features:
• Time Burst: % of reviews posted within a short time window
• Rating Deviation: how extreme/one-sided the group's ratings are
• Content Similarity: how similar the group's reviews are textually (cosine similarity)
• Target Concentration: how focused the group is on a single product/business
• Account Age: average account age of group members

Phase 4: Classification & Ranking

• Apply a spam score function to each candidate group
• Rank groups by spam score (high = suspicious)
• Groups above threshold are flagged as SPAM groups
• Individual members of flagged groups are also marked as suspected spammers

8.3 Spam Score Calculation

VESIT IT Engineering Page 20 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

SpamScore(Group G) = w1 × TimeBurst(G) + w2 × RatingDeviation(G)

■ Spam Score
+ w3 × ContentSimilarity(G) + w4 × TargetConcentration(G). Where
(Group)
w1..w4 are weights. Higher score = more likely spam group.

Feature Spam Indicator Genuine Behavior

Most reviews posted in very short time Genuine reviews spread naturally over
Time Burst
window time

Average rating far from platform mean (all 5★

Rating Deviation Genuine users give varied ratings
or all 1★)

Reviews in group textually very similar (high

Content Similarity Genuine reviews are diverse in content
cosine sim)

Group focuses almost entirely on one Genuine users review many different
Target Concentration
product/business products

Group members all have very new/young Established reviewers have account
Account Age
accounts history

Reviewer-Product
High proportion of shared reviewed products Genuine users rarely overlap extensively
Overlap

8.4 Algorithm Summary Flowchart

■ Input: All Reviews from Platform

■ Build Reviewer–Product Matrix

■ Find Co-reviewing Pairs → Generate Candidate Groups (Clustering)

■ Extract Group Features (Time Burst, Rating, Similarity, Concentration)

■ Calculate Spam Score for Each Group

■ Flag Groups Above Threshold as SPAM → Mark Members as Spammers

For the algorithm, always mention FOUR phases: Data Collection, Candidate
■ EXAM Generation, Feature Extraction, and Classification. List all 6 features (Time Burst, Rating
TIP Deviation, Content Similarity, Target Concentration, Account Age, Reviewer-Product
Overlap).

■ QUICK REVISION SUMMARY

✔ 4 Phases: Collect → Candidate Groups → Extract Features → Classify

✔ Key Insight: Groups of spammers co-review the SAME products

VESIT IT Engineering Page 21 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

✔ Features: Time Burst, Rating Deviation, Content Similarity, Target Concentration, Account Age,
Overlap

✔ SpamScore = weighted sum of all features; threshold determines flagging

✔ Individual members of spam groups also get flagged

VESIT IT Engineering Page 22 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ SECTION 9: ABNORMAL BEHAVIOR PATTERNS FOR

SPAM DETECTION
Spam detection relies on identifying abnormal behavioral patterns that distinguish spammers from genuine
reviewers. These patterns are used as features in spam classification systems.

9.1 Reviewer-Level Behavioral Patterns

A. Rating Patterns
• Rating Deviation (RD): Spammers' ratings deviate significantly from other reviewers' ratings for the same
product.
• Formula: RD(reviewer) = Average |reviewer_rating - mean_rating_for_product| across all reviewed products
• Example: If average rating for "iPhone 15" is 3.8★, a spammer rates it 5★ consistently → High RD
• Extreme Rating Bias (ERB): Spammers tend to give only extreme ratings (1★ or 5★), never moderate (3★).
• Detection: Calculate proportion of 1★ and 5★ reviews. Spammers: >85% extreme ratings.

B. Review Timing Patterns

• Reviewing Burst (BR): Spammers post unusually large number of reviews in a very short time.
• Example: A genuine user might post 2-3 reviews/month. Spammer posts 50 reviews in one day.
• Detection: Flag accounts with more than X reviews in fewer than Y days (X and Y set by platform policy).
• Early Time Frame (ETF): Spammers post reviews very early — immediately after product listing — before
genuinely using the product.
• Example: Review posted within 2 hours of a new product going live on Amazon.

C. Content-Based Patterns
• Review Similarity (RSim): Spammers' reviews are highly similar in content (copy-paste or template-based).
• Measurement: Cosine similarity between reviews. Threshold: if similarity >0.7, flag as suspicious.
• Example: "Great product. Highly recommend." appearing across 200 reviews.
• Short/Generic Reviews (SGR): Spam reviews are often very short and non-specific.
• Example: "Excellent!" or "Very bad!" with no detailed explanation.

D. Reviewer Activity Patterns

• Reviewing Activity (RA): Spammers review many products in a short time period. Abnormally high activity.
• First Reviewing Activity (FRA): Spammer starts high-frequency reviewing immediately after account
creation.
• Genuine users: Activity builds gradually over time.

9.2 Product-Level Abnormal Patterns

Pattern Description Example

Product receives sudden spike of reviews (all 5★ New hotel gets 200 5-star reviews in
Rating Burst
or all 1★) in a short time. one weekend

Bimodal distribution — a legitimate product

Rating Distribution Honest product: mostly 3-4 stars.
usually has bell-shaped rating distribution. Spam
Anomaly Spammed: extreme 1★ and 5★
creates U-shaped (many 1★ and many 5★).

VESIT IT Engineering Page 23 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

Reviewer Too many reviews from users who ONLY 80% of reviewers have reviewed only
Concentration reviewed this product. this one product

Geographic Most reviewers come from same geographic All 100 reviewers appear to be from
Concentration location unusually. same small city

9.3 Summary Table — All Abnormal Patterns

Pattern Name Level What it indicates

Rating Deviation (RD) Reviewer Reviewer rates products very differently from others

Extreme Rating Bias (ERB) Reviewer Only 1★ or 5★ ratings; never moderate

Review Burst (BR) Reviewer Too many reviews posted in very short time

Early Time Frame (ETF) Reviewer Reviews posted immediately after product listing

Review Similarity (RSim) Content Reviews are near-identical across products/platforms

Short/Generic Reviews (SGR) Content Reviews are vague, short, and non-specific

Reviewing Activity (RA) Reviewer Abnormally high total volume of reviews

First Reviewing Activity (FRA) Reviewer Burst of activity immediately after account creation

Rating Distribution Anomaly Product U-shaped rating distribution (bimodal)

Reviewer Concentration Product Most reviewers only reviewed THIS product

For "analyze abnormal behavior patterns WITH examples", always: (1) Name the pattern,
■ EXAM
(2) Define it, (3) Give formula if applicable, (4) Give a concrete example. Focus on: RD,
TIP
ERB, Review Burst, ETF, Content Similarity.

■■ COMMON Don't just list pattern names. Examiners want EXPLANATION + EXAMPLE for
MISTAKE each pattern to award full marks.

■ QUICK REVISION SUMMARY

✔ Abnormal patterns categorized at Reviewer-level and Product-level

✔ Reviewer: Rating Deviation, ERB, Review Burst, ETF, Similarity, Short Reviews, RA, FRA

✔ Product: Rating Burst, Bimodal Distribution, Reviewer Concentration, Geographic Concentration

✔ Key measures: Cosine similarity (content), Frequency analysis (timing), Distribution analysis (ratings)

✔ Normal user behavior is the baseline; spammers deviate significantly

VESIT IT Engineering Page 24 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

✍■ EXAM QUESTIONS — MODEL ANSWERS

■ HOW TO All 9 exam questions from the syllabus are answered below in exam-ready format. Use
USE these as templates for your answers. Structure = Introduction + Body + Conclusion.

Q1. Describe the complete process of sentiment analysis in text mining. [10 Marks]
Introduction: Sentiment Analysis (Opinion Mining) is the computational study of opinions, sentiments, and
emotions expressed in text. It is a core application of Natural Language Processing (NLP) and Text Mining.

Complete Process (7 Steps):

Step Phase Details

Step Collect text data from reviews, tweets, blogs, news. Methods: Web scraping, APIs
Data Collection
1 (Twitter API, Amazon API), datasets.

Clean and normalize text: Remove HTML tags, special characters, stopwords.
Step
Text Pre-processing Perform tokenization (split into words), stemming (reducing to root:
2
"running"→"run"), lemmatization (context-aware: "better"→"good").

Step Convert text to numerical form: Bag-of-Words (word frequency), TF-IDF (term
Feature Extraction
3 importance), Word Embeddings (Word2Vec, GloVe for semantic meaning).

Classify using: Machine Learning (Naive Bayes, SVM, Logistic Regression) or

Step Sentiment
Lexicon-based (SentiWordNet) or Hybrid approach. Output:
4 Classification
Positive/Negative/Neutral.

Step Identify which specific feature of an entity the sentiment refers to. E.g., "battery life
Aspect Extraction
5 is terrible" → Aspect = battery life, Sentiment = Negative.

Step Combine sentiment scores across all sentences/reviews to get overall product
Aggregation
6 sentiment.

Step Use insights for: product improvement, brand monitoring, customer service,
Application
7 competitor analysis.

Conclusion: Sentiment analysis transforms unstructured text into actionable insights, helping businesses
understand customer opinions at scale.

Q2. Explain document sentiment classification based on supervised and unsupervised learning.
[10 Marks]
Introduction: Document sentiment classification assigns an overall sentiment polarity
(positive/negative/neutral) to an entire document. Two main approaches exist: Supervised and Unsupervised
learning.

A) Supervised Learning Approach

• Definition: Uses labeled training data to train a classification model.
• Step 1: Collect labeled documents (positive/negative ratings as labels)
• Step 2: Extract features: unigrams, bigrams, TF-IDF, POS tags
• Step 3: Train classifier (Naive Bayes, SVM, Logistic Regression, MaxEnt)
• Step 4: Evaluate: Accuracy, Precision, Recall, F1-score on test set
• Step 5: Predict sentiment of new documents

VESIT IT Engineering Page 25 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

B) Unsupervised Learning Approach (Turney's SO-PMI Method)

• Definition: No labeled data required. Uses Semantic Orientation (SO) based on PMI.
• SO(phrase) = PMI(phrase, "excellent") − PMI(phrase, "poor")
• Extract phrases (adjective-based patterns) from the document
• Calculate SO for each phrase using web/corpus co-occurrence statistics
• Classify document: Average SO > 0 → Positive; < 0 → Negative

Aspect Supervised Unsupervised

Labeled Data Required Not Required

Accuracy Higher Lower

Domain Domain-specific Domain-independent

Methods Naive Bayes, SVM, LR Turney's SO-PMI

Conclusion: Both approaches are complementary. Supervised is preferred when labeled data is available;
Unsupervised is useful for new domains without labeled data.

Q3. Explain sentence level sentiment classification. [5/10 Marks]

Introduction: Unlike document-level classification, sentence-level sentiment classification analyzes each
sentence individually, enabling finer-grained understanding of opinions within a document.

Two Stages:
• Stage 1 — Subjectivity Detection: Classify each sentence as subjective (expresses opinion) or objective
(states fact). Only subjective sentences proceed to Stage 2.
• Stage 2 — Polarity Classification: Classify subjective sentences as Positive, Negative, or Neutral.

Type Example Sentence Classification

Objective "The battery is 4000 mAh" States a verifiable fact → Skip

Subjective +ve "The display is gorgeous!" Opinion → Positive

Subjective -ve "Worst camera I've seen" Opinion → Negative

Key Challenges:
• Negation: "not good" = negative despite positive word
• Sarcasm: "Oh wonderful, another crash!" = negative
• Conditionals: "If this works..." = no current sentiment
Conclusion: Sentence-level classification is more precise than document-level and is essential for
aspect-based sentiment analysis.

Q4. What is opinion mining? List the challenges of opinion mining. [10 Marks]
Introduction: Opinion Mining (Sentiment Analysis) is the computational study of opinions, sentiments, and
emotions expressed in text. The result is structured as a quintuple: (entity, aspect, sentiment, opinion_holder,
time).

• Entity: The object being evaluated (product, person, organization)

• Aspect: Specific feature being discussed (battery, camera, price)
• Sentiment: Polarity (positive/negative/neutral + intensity)
• Opinion Holder: Who expressed the opinion

VESIT IT Engineering Page 26 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

• Time: When the opinion was expressed

Challenges (10 Key Challenges):

Challenge Why it's hard Example

1. Sarcasm/Irony Reverses intended meaning "Oh wonderful, another bug!"

2. Domain Words change polarity across

"Unpredictable" — good for movies, bad for cars
Dependency domains

3. Implicit Opinions No explicit sentiment word used "Phone died in 3 hours" = negative (inferred)

4. Comparative
Relative opinions hard to classify "A is better than B"
Opinions

5. Negation Handling Negation flips polarity "not bad", "never disappoints"

6. Fake/Spam Reviews Opinions may be fabricated Fake 5-star reviews by competitors

7. Cross-lingual Multiple languages/scripts Hinglish, Tamil, code-mixed text

8. Subjectivity
Fact vs opinion is hard "The screen is 6.1 inches" vs "screen looks great"
Detection

9. Aspect Identification Which aspect does opinion refer to? "It fits perfectly" — size? design?

10. Temporal Opinions Opinions change over time Launch day reviews vs 1-year-later reviews

Conclusion: Despite these challenges, opinion mining has wide applications in business intelligence, product
development, and political analysis.

Q5. Explain dictionary-based and corpus-based approaches for collecting opinion words. [10
Marks]
Introduction: Opinion words are the building blocks of sentiment analysis. Two main approaches exist for
building a comprehensive opinion word lexicon.

A) Dictionary-Based Approach
• Idea: Start with seed words → expand using thesaurus (WordNet) synonyms/antonyms.
• Step 1: Select seed words: a few known +ve ("good","excellent") and -ve ("bad","poor") words
• Step 2: Find synonyms of +ve seeds → new +ve words; synonyms of -ve seeds → new -ve words
• Step 3: Find antonyms of +ve seeds → new -ve words (and vice versa)
• Step 4: Iterate: add new words to seed set; repeat until convergence
• Step 5: Manual pruning: remove inappropriate or ambiguous words

■■
Pro: Language-independent. Con: Misses domain-specific and slang words.
TRADE-OFF

B) Corpus-Based Approach
• Idea: Use statistical patterns in large text corpus to find opinion words.
• Method 1 — PMI/SO Approach: SO(word) = PMI(word, positive_seed) − PMI(word, negative_seed). High
positive SO = positive word.
• Method 2 — Conjunction Rules: Words joined by "and" share polarity; words joined by "but" have opposite
polarity.
• Example: "elegant and stylish" → both +ve. "cheap but tasteless" → opposite polarity.

VESIT IT Engineering Page 27 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■■
Pro: Can discover domain-specific words. Con: Needs large corpus.
TRADE-OFF

Aspect Dictionary-Based Corpus-Based

Data Thesaurus (WordNet) Large text corpus

Domain Domain-independent Can be domain-specific

New Words Cannot discover Can discover new opinion words

Method Synonym/Antonym expansion PMI + conjunction rules

Conclusion: Both methods are often combined in practice to build comprehensive, domain-adapted sentiment
lexicons.

Q6. List and explain types of spam and spammers. [5/10 Marks]
Introduction: Review spam refers to fake or deceptive opinions posted online to manipulate product reputation.
It is a major challenge for review platforms like Amazon, Yelp, and Google Reviews.

Types of Spam Reviews:

# Type Explanation

1 Fake/Untruthful Reviews Positive or negative reviews written to promote or harm a target.

2 Brand Reviews Off-topic reviews praising the brand instead of a specific product.

3 Non-reviews Text that is not a review: questions, ads, irrelevant posts.

4 Incentivized Reviews Real users paid or given free products to write positive reviews.

5 Duplicate Reviews Same or near-identical reviews posted multiple times.

6 Early Reviews Reviews posted before reviewer could have used the product.

Types of Spammers:
# Type Characteristics

1 Individual Spammer Single person; uses multiple sock-puppet accounts.

2 Group Spammer Organized team; coordinated attack; hardest to detect.

3 Crowdsourced Spammer Uses crowd marketplace (Fiverr, mTurk) to hire reviewers.

4 Bot-based Spammer Uses automated scripts to post large volumes of fake reviews.

5 Incentivized Reviewer Regular user biased by rewards; not purely malicious.

Conclusion: Group spammers represent the most serious threat as they coordinate attacks at scale and are
much harder to detect than individuals.

Q7. Explain hiding techniques used by spammers (individual/group). [10 Marks]

Introduction: Spammers use sophisticated hiding techniques to disguise their spam activity and avoid detection
by automated systems.

Individual Spammer Hiding Techniques:

VESIT IT Engineering Page 28 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

Technique How it helps evade detection

Mixing Genuine Reviews Posts real reviews alongside fake ones to appear authentic.

Sock Puppets Creates multiple fake accounts with different identities.

Timing Variation Distributes spam over weeks instead of posting all at once.

Content Variation Rewrites/paraphrases spam reviews to avoid duplicate detection.

Rating Variation Occasionally gives moderate ratings to appear unbiased.

VPN/IP Masking Uses VPN or proxies to hide true IP address.

Delayed Posting Creates account weeks before using it for spam.

Group Spammer Hiding Techniques:

Technique How it helps evade detection

Time Coordination Members post at different times to avoid burst detection.

Role Assignment Different roles: some build reputation, some post spam.

Separate IPs Each member posts from different location/IP.

Mixed Genuine Reviews Group also posts real reviews elsewhere to build credibility.

Content Diversification Uses templates with individual variations per member.

Staggered Account Creation Creates accounts at different times to avoid creation burst detection.

Multi-platform Attack Simultaneously attacks target across multiple review platforms.

Conclusion: Group spammers are more organized and use coordinated strategies that make individual
account-level detection insufficient. Graph-based and behavioral analysis approaches are needed to detect
group spam effectively.

Q8. Explain steps in the Group Spam Detection algorithm. [10 Marks]
Introduction: Group spam detection identifies clusters of reviewers who collaborate to post fake reviews. It
analyzes collective behavior patterns rather than individual reviewer behavior.

Key Insight: Legitimate reviewers rarely review the same set of products, but spam groups show unusually high
co-review overlap.

Phase Name Details

Data Collection & Collect all reviews. Extract reviewer ID, product ID, rating, date, content.
Phase 1
Preprocessing Build reviewer–product matrix.

Find reviewer pairs with high co-review overlap. Apply clustering

Phase 2 Candidate Group Generation (K-means/hierarchical) to group similar reviewers. Form candidate spam
groups.

For each group extract: Time Burst (burst of reviews), Rating Deviation
(extreme ratings), Content Similarity (text similarity), Target
Phase 3 Feature Extraction
Concentration (focused on one product), Account Age (all new
accounts), Co-review Overlap.

SpamScore(G) = weighted sum of all features. Groups above threshold

Phase 4 Spam Score & Classification
→ flagged as spam. Members also marked suspicious.

VESIT IT Engineering Page 29 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

Conclusion: The Group Spam Detection algorithm effectively identifies coordinated spam campaigns by
exploiting the co-review overlap and behavioral anomalies that are impossible to hide at the group level.

Q9. Analyze abnormal behavior patterns used for spam detection with examples. [10 Marks]
Introduction: Spam detection systems analyze deviation from normal reviewer behavior. Spammers exhibit
distinct abnormal patterns at the reviewer level and product level that can be used to detect them.

Reviewer-Level Patterns:
Pattern Description Example

Reviewer's ratings significantly differ from Product avg = 3.5★; Spammer rates it 5★
Rating Deviation (RD)
average ratings for same product. every time → High RD.

Extreme Rating Bias Over 85% of reviews are 1★ or 5★; no Spammer gives all 200 reviews either 1★ or
(ERB) moderate ratings. 5★.

Unusually large number of reviews in a Normal user: 2-3 reviews/month. Spammer:

Review Burst (BR)
very short time window. 50 reviews in 1 day.

Reviews posted immediately after product

Early Time Frame Review posted 2 hours after product goes live
listing — impossible to have used the
(ETF) on Amazon.
product.

Reviews are near-identical (high cosine "Great product!" appearing in 200 reviews
Content Similarity
similarity > 0.7). from same reviewer.

Reviews are extremely short, vague, and

Short Generic Reviews "Excellent!" or "Terrible!" with no explanation.
non-specific.

First Reviewing Activity Account created recently → immediately Account created yesterday; 40 reviews
(FRA) posts high volume of reviews. posted today.

Product-Level Patterns:
Pattern Description Example

Product receives sudden spike of reviews in Hotel gets 200 5★ reviews in one
Rating Burst
short time. weekend.

Bimodal Rating Ratings are either very high or very low; no Most ratings are 1★ or 5★ with very few
Distribution moderate ratings. 2★-4★.

Reviewer 80%+ of reviewers reviewed ONLY this one 90 out of 100 reviewers have no other
Concentration product. reviews.

Conclusion: By combining multiple behavioral features, spam detection systems can accurately identify
spammers even when they use hiding techniques. Machine learning classifiers trained on these features
achieve high accuracy in real-world review platforms.

VESIT IT Engineering Page 30 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ LAST-DAY REVISION SHEET — MODULE 6

■ Sentiment Analysis

✔ 7 Steps: Collect→Preprocess→Feature Extract→Classify→Aspect Extract→Aggregate→Apply

✔ 3 Levels: Document, Sentence, Aspect/Feature

✔ 3 Approaches: ML (Supervised), Lexicon-based (Unsupervised), Hybrid

■ Document Classification

✔ Supervised: labeled data → ML model (Naive Bayes, SVM, LR)

✔ Unsupervised: Turney's SO-PMI: SO = PMI(phrase,"excellent") - PMI(phrase,"poor")

✔ SO > 0 = Positive; SO < 0 = Negative

■ Sentence Classification

✔ Stage 1: Subjectivity Detection (Subjective vs Objective)

✔ Stage 2: Polarity (Positive/Negative/Neutral)

✔ Challenges: Negation, Sarcasm, Conditionals

■ Opinion Mining

✔ Quintuple: (entity, aspect, sentiment, holder, time)

✔ 10 Challenges: Sarcasm, Domain, Implicit, Comparative, Negation, Spam, Cross-lingual, Subjectivity,

Aspect ID, Temporal

■ Opinion Words

✔ Dictionary-based: seed words → synonym/antonym via WordNet → iterate

✔ Corpus-based: PMI on corpus + conjunction rules (and=same polarity, but=opposite)

✔ Dictionary: simple but misses new words; Corpus: finds domain-specific words

■ Spam & Spammers

✔ 6 Spam types: Fake, Brand, Non-review, Incentivized, Duplicate, Early

✔ 5 Spammer types: Individual, Group, Crowdsourced, Bot-based, Incentivized reviewer

✔ Group spammers = hardest to detect

■ Hiding Techniques

✔ Individual: Genuine mix, Sock puppets, Timing variation, Content variation, VPN, Delayed posting

✔ Group: Time coordination, Role assignment, Separate IPs, Diverse content, Staggered accounts

■ Group Spam Detection

✔ 4 Phases: Collect → Candidate Groups → Extract Features → Classify

✔ Key features: Time Burst, Rating Deviation, Content Similarity, Target Concentration

✔ SpamScore = weighted sum; groups above threshold = flagged

VESIT IT Engineering Page 31 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ Abnormal Behavior Patterns

✔ Rating Deviation, Extreme Rating Bias, Review Burst, ETF, Content Similarity, Short Reviews

✔ Product-level: Rating Burst, Bimodal Distribution, Reviewer Concentration

✔ Key: Compare to baseline normal behavior; spammers deviate significantly

VESIT IT Engineering Page 32 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ IMPORTANT KEYWORDS LIST

Sentiment Analysis Computational study of opinions, sentiments, emotions in text

Opinion Mining Same as Sentiment Analysis; studying opinions about entities

Opinion Quintuple (entity, aspect, sentiment, holder, time) — formal opinion definition

Semantic Orientation (SO) SO = PMI(phrase, +seed) - PMI(phrase, -seed); +ve = positive word

PMI Pointwise Mutual Information — measures co-occurrence strength

TF-IDF Term Frequency - Inverse Document Frequency; feature weighting method

Subjectivity Property of expressing opinion rather than stating facts

Polarity Direction of sentiment: Positive, Negative, or Neutral

Opinion Word Word conveying sentiment; also: sentiment word, polarity word

Seed Words Initial small set of known +ve/-ve words used to build lexicon

WordNet Lexical database (thesaurus) used in dictionary-based approach

Corpus-based Approach using statistical analysis on large text collections

Naive Bayes Probabilistic classifier commonly used for sentiment classification

SVM Support Vector Machine — finds optimal hyperplane for classification

Review Spam Fake/deceptive reviews to manipulate product reputation

Sock Puppet Fake account created by spammer to post spam

Group Spammer Organized group collaborating to post coordinated fake reviews

Rating Deviation (RD) Reviewer's ratings differ significantly from other reviewers

Review Burst (BR) Abnormally high number of reviews posted in short time

Early Time Frame (ETF) Reviews posted immediately after product listing

Time Burst Group spam feature: most reviews concentrated in short time window

Co-review Overlap Number of shared reviewed products between two reviewers

Spam Score Weighted score computed to rank likelihood of group being spam

Negation Handling Processing "not", "never" etc. that flip word polarity

Aspect-based SA Extracting sentiment about specific features/aspects of entity

Lexicon Dictionary of opinion words with polarity scores

VADER Valence Aware Dictionary for sEntiment Reasoning — popular lexicon

SentiWordNet Sentiment lexicon assigning +ve/-ve/-ve scores to WordNet words

Bimodal Distribution Rating distribution with peaks at both extremes (1★ and 5★)

Conjunctive Rule "and" → same polarity; "but" → opposite polarity for opinion words

VESIT IT Engineering Page 33 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ SHORT NOTES FORMAT (5 MARKS)

✍■ FORMAT For 5-mark questions: Write 1 introduction sentence + 4-5 key points + 1 conclusion
TIP sentence. Use bullet points. Target ~150-200 words.

■ SHORT NOTE: Sentiment Analysis — Overview

• Sentiment Analysis is the computational study of opinions and emotions in text, used in NLP and Text Mining.
• Three levels: Document (whole document), Sentence (each sentence), Aspect (specific features)
• Approaches: Supervised ML (Naive Bayes, SVM), Unsupervised Lexicon-based, and Hybrid
• Applications: Product reviews, social media monitoring, customer feedback analysis
• Key process: Data Collection → Preprocessing → Feature Extraction → Classification → Insight

■ SHORT NOTE: Opinion Words — Collection Approaches

• Opinion words convey sentiment (e.g., "excellent", "terrible").
• Dictionary-based: Start with seed words → expand via WordNet synonyms/antonyms → iterate until
convergence.
• Corpus-based: PMI scoring + conjunction rules ("and"=same polarity, "but"=opposite polarity).
• Dictionary-based is language-independent but misses domain-specific words.
• Corpus-based can discover new domain-specific opinion words from large text collections.

■ SHORT NOTE: Types of Spam & Spammers

• Spam reviews are fake/deceptive opinions written to manipulate product reputation.
• Types of spam: Fake reviews, Brand reviews, Non-reviews, Incentivized, Duplicate, Early reviews.
• Types of spammers: Individual, Group, Crowdsourced, Bot-based, Incentivized reviewer.
• Group spammers are most dangerous: coordinated, high-volume, hard to detect.
• Detection uses behavioral features: rating patterns, posting frequency, account age.

■ SHORT NOTE: Abnormal Behavior Patterns

• Spam detection identifies deviation from normal reviewer behavior patterns.
• Key patterns: Rating Deviation (RD), Extreme Rating Bias, Review Burst, Early Time Frame (ETF).
• Content-based: Review similarity (cosine similarity > 0.7), short/generic reviews.
• Product-level: Rating burst, bimodal distribution, reviewer concentration.
• Detection systems combine multiple features using ML classifiers for accuracy.

VESIT IT Engineering Page 34 © Study Notes — All Concepts Covered

ADM MODULE 6 — Sentiment Analysis, Opinion Mining & Spam Detection Exam Study Guide

■ LONG ANSWER FORMAT (10 MARKS)

✍■ FORMAT For 10-mark questions: Introduction (2-3 lines) + Detailed Explanation with sub-sections
TIP + Diagram/Table + Examples + Conclusion. Target: 400-500 words + 1 diagram/table.

Section What to write Marks Weightage

Introduction Define the key term + state what you will explain + brief overview 1-2 marks

First major aspect with clear heading + 2-3 bullet points +

Main Body: Point 1 2-3 marks
example

Main Body: Point 2 Second major aspect with heading + bullets + example/diagram 2-3 marks

Comparison Table Side-by-side comparison if two approaches are asked 1-2 marks

Diagram/Flowchart Draw process flow, architecture, or classification structure 1-2 marks

Conclusion Summarize key takeaways + mention real-world application 0.5-1 mark

Template for 10-Mark Answer

Q: "Explain [Topic X] with [Approach A] and [Approach B]." → Para 1 (Intro): "Topic X
is... It is used for... Two main approaches are A and B." → Para 2 (Approach A):
■ Heading → Definition → Steps 1-5 → Example → Para 3 (Approach B): Heading →
TEMPLATE Definition → Steps 1-5 → Example → Table: Compare A vs B (5-6 rows) → Diagram:
Process flow → Conclusion: "Both approaches are complementary. A is preferred
when... B is preferred when..."

■ ALL THE BEST FOR YOUR

EXAMS! — ADM MODULE 6
COVERED COMPLETELY

VESIT IT Engineering Page 35 © Study Notes — All Concepts Covered