Four mechanisms: social learning, singling out, feedback, and distribution.

This paper examines the relationship between initial user behavior and content production in a social network environment.

Previous studies of participation in online communities have focused on two types of social systems: information commons, where many individuals contribute to the construction of a small number of shared artifacts, and online discussion groups, where individuals exchange messages on a given topic.

Theories about participation can be grouped into three high level categories: what a user sees other users doing (social learning), effects that other users have on the newcomer (feedback), and the general structure of content and exposure achieved through participation (distribution).

Social learning theory indentifies several necessary steps in the learning process: attention, that people need to be able to observe the behavior without distraction; retention, or the need to remember the behavior; reproduction or the ability to perform the action; and motivation, including past, promised, or vicarious reinforcement, which influence us to reproduce what we have learned

H1. Social learning: Newcomers whose friends share more content will go on to contribute more content themselves.

newcomers are singled out by their friends. For example, a friend might “tag” a newcomer in a photo, engage the newcomer in a chat session, or refer to the newcomer in a public status update

H2. Singling out: Newcomers who are singled out in content will contribute more content.

Feedback
differs from social learning, particularly singling out, in that feedback requires the newcomer produce some initial content, while newcomers can be singled out without taking any actions themselves.

H3. Feedback: Newcomers receiving more feedback on their initial content will go on to contribute more content.

H4. Distribution: Newcomers whose initial content is distributed widely will go on to contribute more content.

The most common form of content contribution within Facebook is photo uploading; the photo application draws more than twice as much traffic as the next three largest photo sharing websites

All variables were aggregated from sever logs using the Hadoop distributed computing system

Independent variables

Learning is represented by the number of photo-uploading stories the newcomers saw in their News Feeds during their first two weeks.

Singling out is represented by a binary variable indicating whether the newcomer was tagged in a photo during his or her first two weeks.

Feedback is measured by a binary variable indicating whether or not the newcomer received any comments on his or her initial photos during the first two weeks.

distribution is measured in terms of News Feed stories viewed by friends rather than pure friend count.

“‘Tag,’ to me, connotes graffiti. And that’s a negative to me.
Kind of a violation of privacy . . . that’s kind of stalking.” (P2,
a woman in her late 40s)
“Generally no [tagging] in family photos. Why would you
bother tagging, because we all know who they are?” (P6, a
woman in her late 40s)
”I’m just assuming that like a game of tag, I’m It, and now I’ve
tagged someone else. I’m done, and that means [my friend’s] It,
and I’m no longer It.” (P1, a man in his early 40s).
“I thought that only other people could tag you in their photos”
(P5, a woman in her mid 20s).

Burke, M., C. Marlow, et al. (2009). Feed me: motivating newcomer contribution in social network sites, ACM.

Cumulated Gain-Based Evaluation of IR Techniques/ What’s the value of TREC – is there a gap to jump or a chasm to bridge?
Examines the usefulness of a document based on their ranking provided by the information retrieval system. Three architectures are defined: one that is based on the ranked result list, the second is the same but provides a discount factor, and lastly “relative-to-the-ideal performance of IR techniques, based on the cumulative gain they are able to yield”. The gain is measured by the higher the relevancy value is more crucial the marginal values and the higher the ranked value the less important it is to the user. There is also case study that involves TREC-7, which includes over 500,000 documents and 1.9 terabytes. Contrary to this article though in regards to TREC is also the main topic of the second article which counter imposes the work of TREC for its use in generalization based on a distinction between the system requirement and environment and the task context.

Chapter 12 focuses on language models, which is a probabilistic model that uses the string of determined from the related documents to rank its importance to the users query. There are multiple types of language models, including: unigram language models and bigram language models. Once these models are in practice, other means must be taken into account to apply these language models. This can include (but is not limited too) query likelihood ratios, multinomial Naive Bayesian model, etc.
Half way through the chapter, an experiment is given based on these language models, which described the benefits of term weighting. Both of the authors were obviously favored this idea and brought much debate about how realistic and applicable their results were. This then leads into a discussion between IR and the language model presented.

Relating the New Language Models of Information Retrieval to the Traditional Retrieval Models
Relating the New Language Models of Information Retrieval to the Traditional Retrieval Models begins by providing a brief outline on most of the topics we have already stated in class or through the readings. But then provides some key new approaches in regards to relevance weighting and document ranking based on the Boolean operate used in user defined queries. From this point I did have some difficulties determining the exact nature of the process and how it could be applied to what we’ve discussed. The most I was able to take away from the reading was the algorithm, but not its application nor implementation requirements or needs.

Chapter 6: Scoring, term weighting and the vector space
This chapter looks at how scoring can be completed for finding matching documents. Also, if matches are found, how ranking these pages can provide more optimal results. This could be as simple as a score that takes into account only word frequency. Also, utilizing metadata for parametric and zone indexes, which could include the author(s) name, title, date. Weighing a document based on importance could also provide good, but all be it not perfect, results. This would provide all scores in a normalized manner, as in from [0,1] and could also included training data to fine tune results to meet the need of the user generated query. Lastly, the idea of vector space scoring is described. This method takes a set of document and creating vectors of these documents. Finding similarity would only necessitate using common vector algebra, like cosine similarity and find the magnitude of the vector. Again, normalization may also be used to create a common space to find relevant documents given a user query.

Chapter 2: Term Vocabulary and Postings Lists
The chapter examines the measures to process documents into lists. This is started by first identifying the type of document or format of the file we are processing. From there is has to be determine what size the document unit needs to be, which is the standardized measurement of the number of characters/pages/books/libraries used for the processing. For larger items, granularity issues arise. So limiting the work space to the chapter instead of a book is sometimes a necessity. Once the format and document unit is confirmed the text is then tokenized by breaking the piece into word or word phrases, where it is useful for some words to stay as a single case (currency, URL, etc.). Stop words can then be utilized to remove heavily used and of little valued words. These can be created from previous lists or collection frequency information. Normalization and stemming can then be used to help with various word spellings, linguistics, capitalization issues, or errors in spelling. Sets of words may also be taken in bioword indexes to help with finding phrases and searches can be reduced by using position indexes.

Chapter 3: Dictionaries and Tolerant Retrieval
The chapter is focused on the use of dictionary and retrieving information from tree structures of the document information. The dictionary is actually a data structure used to store words as they are processed. Search trees are then used to obviously search the information that is stored. Trees or B-trees are common for such trees. B-trees are actually by necessity even, which is ideal for search trees. Unevenness can cause errors in search or exponentially increase searching areas. Wildcards are import for searching. They allow user to designate multiple words and word meaning by their spelling and use. For wildcards at the end of words, normal B-trees are used, but for wildcards at the beginning of words, reverse B-trees are used. K-gram indexes are another method of retrieval when using wildcards. They are defined subwords of the words used in the tree. As in cas, ast, … for castle. Spelling mistakes are a great concern in IR and they can be handled by edit distance, or the differences from one word to another either by insertion, deletion, augmentation, and lastly k-gram index, which is the same as listed before but for corrections.

Chapter 4: Index Construction
This chapter covers index construction, which is a process that creates inverted indexes. The start of the chapter begins with hardware introductory and explaining some of the basics of a computer and grid processing. Next using the Reuters-RCV1 collection, block sort based indexing is examined, but for a large dataset this method is not efficient enough, therefore single pass in memory indexing is applied. Distributing indexed is then mentioned with an example of MapReduce, which is a grid based IR developed by Google. Lastly the chapter goes into dynamic indexing, which takes into account documents that keep changing.

Chapter 5: Index Compression
Chapter 5 goes over some of the basics on how to make the dictionary and indexing more efficient. This could be used to benefit storage and memory usage. This chapter even goes into detail by describing some of the effects of preprocessing and post processing statistics, which mentions a concept that is very familiar to database designers and that, is loseless and lossy compression. This concept is basically making sure that compression is done but without losing the meaning of the information. Heap’s law can be used to calculate the vocabulary size and Zipf’s law can be used for modeling the distribution of terms. The rest of the chapter examine the compression of the dictionary, pre and post processing.