Master's capstone · 2019

CorrelaTech

Tech news articles classified as having positive, negative or neutral impact on a company, with a confidence score, set against that company's stock price over the same window. Built for CSE 593 at Arizona State University while at Intel, and written up as Classifying Tech News with Sentiment Analysis and Machine Learning.

Master’s capstone
CSE 593 · Arizona State
Intel Corporation
April 2019

Angular UI SailsJS API Supervised classifier over article text 5 companies — Amazon, Facebook, Google, Intel, Microsoft

The paper

20 PAGES · PDF

The argument, in short: sites like TechCrunch and The Verge published volumes of company news with no sentiment analysis attached, while algorithmic processing was already routine in financial trading. Classifying each article for impact and charting it against the stock gives a reader the trend around a company without asking them to read every headline — and doing it with a model rather than an editor keeps human bias out of the label. Read the paper (PDF, 800 KB).

Recorded demo

FOUR CLIPS · 5 MIN

1 · Overview and architecture — 0:34 The landing page, which doubles as the app's own documentation: how articles get ranked, and how correcting a ranking with the pencil feeds back as a training label. Below that, the build — one Angular page template driven by a directive per company, and a SailsJS API whose models map to the database columns behind each page.

2 · A company page — 2:03 Amazon, filtered by date range and sort order, with the AMZN price series beside the articles at one month, three, six or a year. The same page template serves every company in the sidebar.

3 · The classified feed — 1:46 Two hundred articles loaded for Facebook, each with its source, timestamp and the classifier's verdict, running alongside the FB chart for the same window.

4 · Ingestion — 0:59 The scraper running on the project VM, inserting articles company by company. The unique constraint on title does the deduplication, so re-running the job over the same news cycle is safe.

These clips came out of the original demo deck, re-encoded from 122 MB to 13 MB for the web. They are silent: the deck's title slide promised audio, but no audio track survived in the embedded video. The app itself ran at correlatech.azurewebsites.net and has long since been retired, so this recording and the paper are what's left of it.