Visual Meme in Social Media: Tracking Real-World News in YouTube Videos

Lexing Xie1, Apostol (Paul) Natsev2, John R Kender3, Matthew Hill2, John R Smith2

1Australian Nation University, 2IBM Research, 3Columbia University

ACM Multimedia 2011, Scottsdale, AZ, Nov 2011 [pdf][slides]
^an short version appeared in ICWSM 2011; ^^an extension of this work appeared in ACM MM 2010

A re-implementation of the meme-detection algorithm, by David Hehir, is here.

Introduction
We propose visual memes, or frequently reposted short video segments, for tracking large-scale video remix in social media. Video remixing is prevalent on social media platforms, it is part of "venacular creativity" (Burgess 2009) where users create "curated selections based on what they liked or thought was important" (Snickars 2010, page 270). Social influence are often characterized from text-based online interactions such as quoting or reweeting (Leskovec 2009). Our tool allows such metric to be developed for visual media.

Here are some example observations:

Our main approaches include: Topic-based monitoring of news content on YouTube, robust visual matching at scale, graph-based influence analysis to observe content importance and user roles.

Event Buzz
YouTube has become a virtual worldwide bazaar for video content of almost every type. With 72 hours of video being added every minute, it is a living marketplace of ideas and a vibrant recorder of current events. We use text queries to pre-filter content, and make the scale of monitoring feasible. This is done via a number of generic, time-insensitive text queries sent to YouTube API. We use the term buzz to refer to all the videos that respond to keyword queries YouTube, although their content may not be directly related to the current event. We use the term meme videos to refer to containing one or more memes. Although being a noisy representation of The volume of retrieved and memes are telling indicators of event evolution in the real world, as seen in example traces on the right.

Meme Tracking

Visual memes are defined as frequently reposted video segments or images. It has been observed that users tend to . News event collections are particularly suited for studying large-scale user curation, since remixing is more prevalent here than on video genres designed for self-expression, such as video blogs. The unit of interaction appears to be video segments, consisting of one or a few contiguous shots.

The Figure on the right shows two example YouTube videos that share multiple different memes. Note that it is impossible to tell from metadata or the YouTube video page that they shared content, and that the appearance of the remixed shots (bottom row) has large variations.

We query YouTube using a set of topic-related keywords to retrieve large amounts of video entries over a few months.
Our tracking algorithm for large-scale video shot matching include two main parts: robust visual matching using color-correlogram, and scalable indexing using approximate nearest neighbors. A high-level flow chart of the algorithm components is found to the right.

Example Memes

Meme clusters from the SwineFlu collection, we can see examples of pig-related song or mock clips (S1 and S3), educational clip (S4).
(S1) (S2) (S3) (S4)

Meme clusters from the Iran2009 collection include potraits of prominent individual, groups during the event, as well as memorable scenes. The total number of visual memes in this collection is over 10,000, only a small sample is shown here.

(I1) (I2) (I3) (I4)
Observations of Meme Influence

Here we visualize author productivity (in number of videos uploaded, on the x-axis) and their influence indices (paper Section 6, on the y-axis) in several scatter plots. For both the Iran3 topic and the SwineFlu topic, we plot the total diffusion influence (summed over all videos by an author) and the normalized diffusion influence (total influence divided by the number of videos the author has).

In the Iran2009 topic (blue dots) we can see two distinct types of contributors.

  • We informally refer to the first type as mavens (Gladwell 2000, marked in red), denoting users who post only a few videos but which tend to be massively remixed and reposted. This particular maven was among the first to post the murder of Neda Soltan, who became the icon of the entire event timeline.
  • We call the second contributor type connectors (Gladwell 2000, circled in green), denoting users who tend to produce a large number of videos, and who have high total influence factor but have low average influence per video. They aggregate notable content and serve the role of bringing this content to a broader audience. (A response metric such as view count or number of comments could further confirm this fact.) We examined the YouTube channel pages for a few authors in this group, and they seem to be voluntary political activists with screennames like "iranlover100", we can also dub them "citizen buzz leaders". Some of their videos are slide shows of iconic images and provide good summaries of the event timeline.
  • Note that traditional news media, such as AljezeeraEnglish, AssociatedPress, and so on (circled in gray), have rather low influence metric for this topic, partially because the Iran government banned foreign journalists and severely limited international media coverage of the event.

    The SwineFlu collection seems different. We can see a number of connectors on the upper right hand side of the total diffusion scatter. But it turns out that they are the traditional media (a few marked in gray), most of which have a large number (>40) of videos with memes. The few mavens in this topic (marked with green text) are less active than in the Iran topic, and notably they all reposted the identical old video containing government health propaganda for the previous outbreak of swine flu in 1976. These observations suggest that it is the traditional new media who seem to have driven most con- tent on this topic, and, while serendipitous discovery of novel content sill exists, it has less diversity.

    Such visualizations can serve as a tool to observe information dissemination patterns in various events, and henceforth characterize influential users. Such tools can identify the key influencers for each event, including both mavens, or early "information specialists", and connectors, who "bring the rest ... together" (Gladwell 2000).

    Dataset

    Selected References