Contents
- Introduction
- Contributors
- Use Case
- Action Items
- Time Line
- Data
- Resources and Links
Introduction
In January 2010, Ehren Foss of Prelude Interactive shared an update on his Social Actions Tuner project. The update led to a discussion about applying TF-IDF functionality to the Action Packs in order to more intelligently sort actions by cause area, geographic location, or theme. In non-technical terms, each action pack would 'learn' to display relevant actions on a cause area, geographic location, or theme based on the keywords contained in actions that have had a high number of click-throughs (and possibly RTs) in the past. On February 11, 2010, Social Actions convened an open call on the topic. This wiki page was setup in advance of the call to serve as an organizing tool for implementing the innovation.
Contributors
Technical Overview
- Filtering with Social Actions API only gets you so far
- Description of Social Actions Tuner
- Built on technology called LDA (latent _____ allocation)
- Attempts to figure out what the actions are. You don't say "I'm looking for literary actions." It's a data structure that consolidates that topic.
- Analyzes groups of text using the terms and their frequency and how they appear across all articles. Attempts to categorize data coming in.
- Used it to build a tool for someone who'd log into Social Actions Tuner, search by category/term/etc. Would vote results -- thumbs up if it was what you were looking for, thumbs down if not.
- Broad application -- if you can behind the scenes record what people are looking for, develop an idea of what was searched for and ultimately clicked on. Improve those search results, and you increase the chances they'll click through and get involved in an action.
- Lots of possible parameters - Number of terms to search, lots of ways to do this. Need to test it alot and hone it for your application. Can what the output.
- Term Frequency to Inverse D__________ Frequency
- TFIDF helps you figure out if X is part of your topic.
- Looks at terms used in particular documents that AREN'T used in all of the documents.
- You could used "literacy action" that uses books etc. and never use the word literacy. Both techniques help you find those actions.
- Most appropriate for Social Actions is most likely TFIDF. Both do similar work: pass in a corpus of data, tell it what you want to see.
Use Case
Our goal is to build intelligent, feedback-based Action Packs that don't necessarily match a keyword used when the action was created, but that reflect keywords that evolve as people click on an action following a search.
Ehren -- that jumps ahead. You could use TFIDF to see last 500 things that kicked out. Use those examples to develop a filter with either technology. If it stopped generating results, would have to adjust it to how people are now
Peter: Other use case: testing _______________. Would improve the actions flowing through the system.One approach: analyze what people have clicked on in last 30 days and say, this is what we want to build on.
Ehren: Great approach. Person doing the filtering wouldn't be biased.
LDA could tell you that alot of people are publishing things that fall between two action pack categories, such as literacy and music.
LDA helps you figure out the topics in your data.
We could announce, starting on this day, "please click-through tweets that have most compelling title."
Ehren: Great example. We can select based on body of action, can rely on community to select based on title. Would learn alot about what the titles are misleading, or not informative. We could find 5 of the most important words in the body. i.e. Grantwriter, vs. Grantwriter literacy children. Have you been recording clicks?
Available datapoints
- Click-through information including referral URLs for actions when viewed on the profile page of Climate Actions or Education Actions, etc. We're recording that on the Social Actions log.
- _______________.
- We also have retweet information. I was research the Twitter API options for actions people are retweeting. Those all might contribute to the filter. Also
- Friendfeed (clicking on it wouldn't contribute to click-through data) or anything similar (tweet this" sandbox to produce the filter)
Whether human review or click-throughs, the goal is to have 20-50 examples of what you do and don't want to see for each one.
What if there's no rhyme or reason to what is voted up or down? Is that a risk we run in voting up and down content?
- Ehren: It can be. Could be people click randomly, or difference between "good" and "bad" is difficult for computer to understand. Expect it will catch most obvious: business environment vs ocean environment returns. We could definitely get started and get totally irrelevant returns out of there.
Another question: it's tough to keep up with creating Action Packs at the rate we can conceive of them. You mentioned the engine identifying clusters of activity. What about the next step: using these tools to identify action packs? Here's where there clusters. "Trending topics producing action packs."
- Ehren: Very interesting, but user interface. Computer will recognize significant terms, but person will have to assign label. Or, giant tag cloud, people click on what they're interested in, program kicks out the closes topic filter. That may be too confusing. The way it's laid out on Social Actions is very clear. Nothing to it but to do it -- feed data into LDA, tell it to generate 80 topics. None of this will save you time upfront.
Time Line
(notes)
Action Items
Getting into deeper detail and analyzing some logs
Data
Social Actions API
- Title
- Description
- Action Source
- Action Type
- Hits
Social Actions API Click-Through Data
Twitter API
Resources and Links
Comments (0)
You don't have permission to comment on this page.