Winter Milestone 5 matchmaking

From crowdresearch
Jump to: navigation, search

System (for task feed)

Brief introduction of the system

The matchmaker algorithm is a system that utilizes a multi-tiered classification system. With every bit of input users have, the system learns more delicate nuances of the user that may not be uncovered otherwise.

How is the system solving critical problems

The system needs to build a knowledge base of not only what the user likes but why they like it. Reasons that a user may like certain tasks can be any one of an enormous assortment of things such as having a preference of 1 paged tasks over multiple paged tasks, tasks that have numerical input rather than word input, tasks that have a correct answer versus a task that is more opinion based and open ended. It would be unreasonable to ask the user to know this about themselves and effectively express it. By constantly obtaining input from the user by allowing them to tell the system “Show me more like this”, the system is constantly learning and is increasing in accuracy.

Introducing modules of the system

The curating of tasks happens on two different levels, one is system generated and the second is user generated. The system’s algorithm creates one giant curated list but users apply their own custom filters on top of it. This allow users to prioritize their own specific needs. One user may only have a hour or two a day to spend working and they may choose to prioritize the monetary value placed on tasks rather than doing tasks that fall under a certain category that they enjoy. Some users may want to strengthen a particular skill such as writing and will create a ‘writing filter’ to pursue this. Other users may like to just kill time and want to explore and broaden their task selection.

Module 1: System Curation

Problem/Limitations

The question is how self-aware users are in terms of what they think they like or don’t like and how good they are at expressing that through filters. Or what user think they are good at versus what they are actually good at. This matters because at the end of the day, most users want to get paid. An experiment to implement would be to have users create their own filtration system of tasks and examine their satisfaction level versus a computer generated filtration system. Satisfaction level may be very difficult to assess so perhaps it would be better to examine monetary differences instead, depending on how reliant satisfaction is to money. This method may be risky because the user may only allow the system to recommend X number of “wrong" suggestions before they feel like they cannot trust the system and leave. [2]

Module preview

Existing filtration systems tend to be too broad and unspecific. As users continue to complete more and more tasks and learn the system, the filtration system stays static rather than growing dynamically with the user. By utilizing a system of classification users will be better able to search, increase efficiency as well as accuracy [1] because it goes beyond typing words into a search field but plays off of their feelings immediately after they have finished a task, they simply have to ask themselves, “Did I like this task?”.

System details

Many companies provide users with methods of input and then suggestions based on previous input/selections they may have made. For example, Netflix recommends that I watch “The X-Files” because I watched “House, M.D.”. When I’m shopping on Amazon for a router, it also shows me modems at the bottom of the page. Pandora Radio uses a thumb up and thumb down system to allow users to give the system input. A thumb up meaning they liked the song (making the radio station play more of that song and songs like it) and a thumb down meaning they don’t want to hear that song anymore or songs similar to it. With every little bit of input the system learns more delicate nuances of the user that may not be uncovered otherwise. To achieve this, tasks need to be broken down into a more complex system of taxonomy. One way to do this may be to implement a more template based/drag and drop interface system that can supply some information to associating characteristics to the task. When creating this multi-tiered classification system, the top class can be left to task requesters but the sub class classification should be left to the system. We will have to experiment with how detailed to break down tasks in the classification system and the library of attributes we are able to assign tasks.


Module 2: User Filters

Problem/Limitations

The great thing about system curated task feeds are that they constantly evolve with the user. Users needs and wants are constantly changing and so should their task feed. The user generated system however, allows users to prioritize different things such as task type (i.e.: writing) or hourly wage.

System details

The user generated filtration system can be as specific or as broad as the user chooses. The process of finding tasks the user may enjoy/want to do has already been done by the system and the user is simply sorting through it.

[1] http://users.cis.fiu.edu/~taoli/pub/icassp-2005.pdf [2] http://files.grouplens.org/papers/www10_sarwar.pdf


Example of main dashboard.