Source-linked AI summary
Mining User Comment Activity for Detecting Forum Spammers in YouTube
Ashish Sureka
TL;DR
Comment spam harms online communities, while user-level analysis of commenting activity is relatively unexplored and content-based methods are imperfect. This paper mines YouTube users’ recent comment logs for spam indicators and reports effective detection in its empirical analysis.
Problem
The paper investigates whether users’ comment activity can identify YouTube forum spammers, an area described as relatively unexplored compared with imperfect content-based methods.
Source: §1 · pp. 1–2Method
The approach mines recent user comment logs and metadata for timing, spam-hint, and repetition patterns used to score users as potential spammers.
Source: §3 · p. 2, §3.3 · p. 3Results
The authors report that their method was effective on a YouTube sample, with repeated comments, very short intervals, and high spam-hint percentages serving as reliable indicators.
Source: §5 · p. 4Takeaways & Limitations
The authors conclude that these comment-activity patterns are reliable indicators for categorizing YouTube forum spammers.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Research shows that comment spamming (comments which are unsolicited, unrelated, abusive, hateful, commercial advertisements etc) in online discussion forums has become a common phenomenon in Web 2.0 applications and there is a strong need to counter or combat comment spamming. We present a method to automatically detect comment spammer in YouTube (largest and a popular video sharing website) forums. The proposed technique is based on mining comment activity log of a user and extracting patterns (such as time interval between subsequent comments, presence of exactly same comment across multiple unrelated videos) indicating spam behavior. We perform empirical analysis on data crawled from YouTube and demonstrate that the proposed method is effective for the task of comment spammer detection.
1. RESEARCH MOTIVATION AND AIM
Comment spam harms legitimate online communities, while user-level analysis of commenting activity is relatively unexplored compared with content-based detection. This work investigates whether patterns in users’ recent comment activity can identify YouTube forum spammers.
1. RESEARCH MOTIVATION AND AIM
- Comment spam can waste computing resources, lower application reputation, burden moderators, and mislead legitimate users.The paper identifies these as negative impacts of spam across online domains.Source: §1 · p. 1
- Content-based methods have shown encouraging results but are imperfect, and user commenting behavior is relatively unexplored as a way to identify spammers.The paper frames behavioral analysis as user classification rather than post or message classification.Source: §1 · p. 1
- The study investigates usage-based features from a user’s recent comment log and associated metadata to find discriminatory patterns for detecting YouTube forum spammers.Source: §1 · p. 2
2. RESEARCH CONTRIBUTIONS
The paper’s contribution is an empirical study of user comment activity as a basis for detecting YouTube forum spammers. It distinguishes this behavioral focus from related YouTube user-classification work.
2. RESEARCH CONTRIBUTIONS
- The authors present what they describe as the first YouTube study mining users’ recent activity logs for behavioral spam markers.The listed markers include repeated comments across videos, very short comment intervals, and community or moderator spam tags.Source: §2 · p. 2
- Unlike related YouTube user-classification work on reputation or video spamming, this paper examines users’ commenting activity to assess forum-spammer likelihood.Source: §2 · p. 2
- The study evaluates usage-based behavioral features on data crawled from YouTube and reports that they can serve as markers for automatically detecting comment spammers.Source: §2 · p. 2
3. SOLUTION APPROACH
The approach retrieves users associated with spam-hinted comments, mines their recent comment logs, and scores them using four activity indicators. The indicators capture comment timing, spam hints, repetition across videos, and repetition within a user’s log.
3. SOLUTION APPROACH
- The system starts from comments marked hasSpamHint, retrieves the associated users’ recent comment logs and metadata, then computes four indicators to score each user.The paper notes that hasSpamHint is applied to comments rather than users, and that some spam comments remain untagged.Source: §3 · p. 2
3.1 ATDC
- ATDC averages time differences between comments in a user’s recent log; a low value is hypothesized to signal spam.The authors reason that automated posting can produce intervals too short to be manually feasible, and report observing sequences with intervals of a few seconds.Source: §3.1 · p. 3
3.2 PCHF
- PCHF measures the percentage of a user’s comments marked hasSpamHint, with a high value hypothesized as a spam signal.The authors report that manually validated spammers often have high PCHF values.Source: §3.2 · p. 3
3.3 CRAV
- CRAV uses high comment similarity across a high variability of video IDs as an indicator of spam-related content promotion.The described pattern is a user posting exactly the same comment across discussions for different videos.Source: §3.3 · p. 3
3.4 CRR
- CRR averages pairwise exact-match comparisons among a user’s recent comments, capturing repeated messages on the same video.The paper hypothesizes that a high CRR signals spam, whether repetitions occur close together or are spread over time.Source: §3.4 · p. 3
4. EMPIRICAL ANALYSIS
The evaluation finds users with high spam-hint rates and repeated comments, and manual inspection supports the proposed spammer-identification heuristics.
4. EMPIRICAL ANALYSIS
- The activity log contained 13,000 comments from 240 unique users; several users with more than 60 comments had over 70% marked with hasSpamHint.Figures 2–6 plot 119 users with more than five comments.Source: §4 · p. 3
- Table 1 is an illustrative list of comments from users identified as spammers in the experimental dataset.The paper reports that manual inspection of the profiles and comments indicated spam among identified users.Source: Table 1 · p. 3, §4 · p. 4
- Manual inspection confirmed the hypothesis: users A–D posted 35, 118, 36, and 30 comments, respectively; each had CRR above 0.7 and 80% of comments marked as spam by the moderator.Source: §4 · p. 3
- Figure 4 shows both high comment overlap with low video overlap and high comment overlap with high video overlap, including identical comments across multiple videos.High comment overlap with low video overlap indicates similar comments concentrated in a single or small set of videos.Source: §4 · p. 3, Figure 4 · p. 4
- The proposed rule classifies users using PCHF, ATDC, COMOVP, and VIDOVP thresholds, with users required to have more than five comments.Figure 5 identifies users below 3 on log average time difference and above 20 comments as potential spammers.Source: §4 · p. 4
5. CONCLUSIONS
The paper concludes that user comment activity logs provide behavioral indicators for identifying YouTube forum spammers, and reports that its rule-based method was effective on a sample dataset.
5. CONCLUSIONS
- The paper presents a rule-based method that mines users’ comment activity logs to identify YouTube forum spammers.Source: §5 · p. 4
- Repeated identical comments in one or multiple videos, short intervals between comments, and high hasSpamHint percentages are reported as reliable spammer indicators.Source: §5 · p. 4
- Applying the method to a sample dataset revealed that it was effective in identifying spammers.Source: §5 · p. 4