Ongoing project for ad hoc work on our existing platform

We are building a credit model that sees beyond traditional FICO for auto loans.  We want to use real-time relavant metrics that measure past achievments with future potential.  Variables such as GPA, University Attended, SAT Score, Career Path, Income, etc to better underwrite risk and approve more thin file consumers who get declined due to lack of payment history.

Swoop is a fast-growing search advertising startup founded by a team with many previous IPO and M&A exits. We run Google AdWords campaigns on Web pages with the same targeting precision Google uses on search result pages. Not infrequently, we outperform Google AdWords. To do this at scale, we are solving a number of both standard and unique text analysis, search, classification and optimization problems.

The goal of this project is to categorize Web page content features into Google AdWords industry verticals. The end result of the project is to improve the matching between ads and pages.

Overview

Google AdWords allows advertising campaigns to be restricted to a set of verticals. For example, an advertiser may choose to show an ad for hiking boots only to a category “Hiking & Camping”. Categories are nodes in a category hierarchy managed by Google, e.g., “Hobbies & Leisure > Outdoors > Hiking & Camping”.

Swoop targets ads to various content features of Web pages. Improving the association of content features to AdWords verticals improves the quality of ad matching.

Because they come from Web pages, content features are associated with a URL and page-level meta-data of varying quality, based on the publisher (title, META keywords, OpenGraph meta-data, etc.). In terms of granularity, content features may be one of the following:

• The "clean" text of the Web page with most navigation, ads, widgets, comments and other "non-core" content removed.

• A snippet of the clean text. We are interested in snippets because, often, the categories of individual snippets on the same page vary. Snippet length varies from a few words (a recipe ingredient, a heading) to a paragraph or a section of content.

• A subset of the snippets on the page.
  
  

The content is overwhelmingly in English with a small portion in Spanish. Other languages are not important in the short run.

Goal

Swoop seeks an algorithm to determine the set of category memberships for a given content feature. The algorithm will be operationalized to work in an online system. Therefore, proposals must include descriptions of the process for both initial and ongoing training/tuning of the algorithm.

Measuring Performance

Algorithm performance will based on the following factors:

• Performance (ROC AUC) against a validation set of content features.

• Operational convenience, including but not limited to: setup, ongoing training & tuning costs, scalability characteristics and operating costs.

Bonus

As an optional bonus, we are also interested in algorithms with the capability to balance type I and type II errors based on the qualities of a particular category in the hierarchy. In categories where we see a lot more content features than we have ads to show, we may be interested in reducing Type II errors. In categories where we have more ad opportunities than content features to target, we may be interested in reducing Type I errors.

Input Data

Web page information:

• URL

• Contents of TITLE tag

• Attributes of all META tags

• High-level vertical of the site the page if from, e.g., auto, health, finance, food, lifestyle, news, entertainment, etc.
  
  

Content feature information:

• URL of Web page

• Snippet text or HTML, in whatever representation is best-suited for processing
  
  

If additional input data would be helpful, please describe it and the reasoning behind the need. We may have access to it already or we may be able to readily acquire it through third parties. An example would by referrer URLs (pages that users arrive from to a given URL).

Training data

If the project solution requires the development of a training data set, describe the process by which it should be efficiently created, the size of the training data set required and the expected cost to create it.

Project is Hourly

This project seeks the highest value work, not the lowest possible cost. Thus the hourly rate range we will consider is very broad, from $50-$200. We will evaluate proposals based on experience and qualifications, and the hourly rate will be based on merit.

Proposals

Please don't merely tell us about your background since we will have access to your profile once you apply. Your proposal should outline your methodology and provide insights into how you would solve the problem.

The Digital Youth Network is a non-profit project that supports organizations, educators and researchers in learning best practices to help develop our youths’ technical, creative, and analytical skills.  The DYN is seeking a Tableau consultant to work with its team to design and implement a suite of dashboards designed to provide timely reports for different stakeholders (i.e. funders, city leadership, program providers, mentors, and youth).  The consultant needs have experience creating highly customized data dashboards with Tableau, be knowledgeable about working with Amazon Redshift, partitioning access to data based upon user status, automating updates, creating reports for view via Tableau reader, and creating interactive visualizations that link to urls.

Summary:

Broadly, we'd like to build an algorithm that takes in "Member Profile Data" and returns a list of the "Top 10 Members You Should Meet", prioritized by best match. 

Background:

We run a mastermind group of 250 CEOs, executives, and investors. The group meets once a year, and we'd like to provide a list to each of them of the top 10 people they should make an effort to meet at the event. 

We collect extensive profile information on each member.

Here are the fields: 

Standard:

Name, Title, Company, Address/Location, Zipcode, Country, Gender, Area Code, Personal Website, LinkedIn Profile URL, Company Website, Number of Employees (Select a Range), Company Annual Revenue (Select a Range)

Non-Standard:

List of Crowd Tools Most Interested In: (Community Building, Crowdfunding, Crowdsourcing, Incentive Competitions)

List of Similar Organizations You Are Involved In: (Strategic Coach, Genius Network, Center RIng, Entrepreneur's Organization, YPO, Business Mastery, Maverick1000)

Number of Previous Events You've Attended: (0, 1, 2, or 3)

Which Global Grand Challenges Are You Interested In Solving: (Abundant Energy, Clean Water, Democracy, Global Healthcare, Hunger, Literacy, Poverty, Other)

Your Industry: (Select from Standard List of Industries)

Description of What Your Company Does In this Industry: (Comment box input)

Are You An Angel Investor: (Y/N)

How Many Companies Have You Started: (Number)

Top Two Exponential Technologies You Are Interested In: (AI, Internet of Things/Networks and Sensors, Robotics, 3D Printing, Cryptocurrency, Alternative Energy, Virtual Reality, Human Longevity)

How Did You Hear About This Event: (Comment box input)

Description of Where You Want Your Company to Be in 5 Years: (Comment box input)

Would You Consider Yourself and Early Adopter (Y/N)

Have you Attended a Singularity University Event?: (Y/N)

If yes to the above, which ones?: (EP, IPP, GSP, Other)

Number of Investment Deals Per Year: (Number)

Average Investment Size: (Select a Range)

Deliverable

The deliverable would be an algorithm that would take in this profile information and "match" the members. Think of a "match" as representative of "most likely to do business together" -- Proximity, similarity of business, similarity of business size/business revenues, similar interests in technology, similar grand challenges they want to solve, similar networks (ie. heard about the event from the same person), similar titles, similar angel investment sizes and frequencies. 

The output would be a ranked list of the top 5-10 best matches.

Note -- Perhaps there should be a randomization factor included here as well. For example, if we have a husband and a wife (or two business partners) attending the event and they both run a business together or work at the same place, we don't want the algorithm to list them as top matches. Perhaps, if they are too similar, they shouldn't be included in the output.

Note 2 -- We will also provide a basic "training set" or benchmark list of connections that we consider "valuable". In other words, we will provide a list of people and their top 5-10 connections (that we know of).

Note 3 -- Though it maybe out of the scope of this project, plugging into LinkedIn's API or crawling individual's personal websites for keywords ("Real-Estate", "Coach", etc) could add interesting dimensions to their similarity indices.  

Bonus -- Though not required, it would be nice if we could see the "Top Factors in Common" for each match (ie. Person A and Person B both want to solve this problem or A and B both run equal sized businesses out of New York).

Language: Python is preferrable, but not required.

User Interface -- We'd like to have a basic user-interface that would allow us to upload a file and see the output of matches. 

Input File -- CSV or Excel Document (All fields not 100% complete)

Output: CSV or Excel Document -- List of Top 10 Matches for Each of the 250 Members

+ README.md with detailed instructions and explanation of algorithm

Proposed Output Format:

Member: John Smith

Top Matches:

1) Name

2) Name

3) Name

4) Name

5) Name

6) Name

7) Name

8) Name

9) Name

10) Name

Member: Jack Stone

Top Matches:

1) Name

2) Name

3) Name

4) Name

5) Name

6) Name

...

[updated to add more clarity on the problem statement]

We have a large set of sentences. 

I 'm looking for the algorithm that will produce the *minimum* set of words (in descending order of appearence) such that each sentence contains *at least one* of these words.

For example:

For the sentence set {"black white red", "black white"} the minimum set of words is 

"black"

For the sentence set {"one two five", "three four six", "seven nine one"} the minimum set of words is

"one", "three"

(there are other correct answers, too, but none with less than 2 words)

We need to build an ingestion pipeline to bring data in from our advertising and editorial portals into Tableau; help create reports and dashboards; and explore options for sharing reports with our clients.

Milestone 1: Manual extraction and assembly of datasets from the portal, migrate some of the charts from the current POP reports and Quantitative reports into Tableau.

Milestone 2: Migrate any remaining charts from these reports, construction of some new analytic charts/dashboards, start prototyping database store for serving content into Tableau

Milestone 3: Automate data ingestion pipeline into Tableau

NOTE: The client has shorlisted candidates. PLEASE DO NOT SUMBIT NEW PROPOSALS.

Seeking technical resource with experience in AWS EMR (Elastic MapReduce) to provide 10-15 hours of consultation.  Need to be able to describe and explain how to perform variety of functions in EMR for purpose of comparison document. 

We have a new and sizeable sand-box environment leveraging Tableau, Redshift, and potentially Informatica Cloud. We need someone that could step in immediately and help us load data into Redshift.

We need to design and implement a system that will enable us to query a massive dataset in realtime.  We need a Big Data architect to help us design the system using appropriate technologies to scale to millions of users and implement the solution by working with our team.  We have the luxury and advantage of organizing and structuring the data set.  Here is the problem:

If a user wants to explorer a node of interest: x1 with and without filters, we would like to show incoming path and outgoing path with aggregations at each node

 

x12 —> x9 —> x1 —> x2 —> x1 is a path

x12 —> x9 —> x1 —> x3 —> x8 is a path

 

We want to show counts next to each node with and without filters like:  (for all x1 the filters f1, f2 and f3... and for x3: fa, fb, f2, f3… and so on)

 

an example of filter f45 is model=BMW

an example of x1 is action=TestDroveCar

 

The data could be thought of stored with filter attributes

x1   f1  f2  f56  f45

x9  fx fy

x12 fa fx f2

 

 //pseudo query

select count(x) from table where 
(
(x == x1 and f1 == f1v and f2=f2v) OR
(x == x2 and f45 == f45v) OR
(x == x3)
)
and
x1Timestamp < x2Timestamp < x3Timestamp

// alternatively we could use intersect