1 / 8

Predicting Citations with Feature Construction and Model Class Selection

This study by Claudia Perlich, Foster Provost, and Sofus Macskassy from New York University in the KDD Cup of August 2003 focuses on predicting citations in research papers. The process involves constructing features related to the temporal shape of citation counts, seasonality of publishing, author and paper reputation, and more. The research delves into training considerations, including normalization, selection of training period and observations. The final model uses an Ordered Probit Model with specified categories and prediction curtailment. The assessment includes a comparison of participant scores and predictive constants.

berthar
Download Presentation

Predicting Citations with Feature Construction and Model Class Selection

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Task1: Predicting Citations Claudia Perlich, Foster Provost & Sofus Macskassy New York University KDD Cup, August 2003

  2. Citations

  3. Step 1: Feature Construction • Temporal “shape” of citation counts: + counts for last 6 quarters and number of missing • Seasonality of Publishing (conferences): + quarterly dummies • Age of paper: + publication quarter • Author reputation: + number of papers, total number of citations + median change at same paper “age” • Paper reputation: + total and average number of citations per quarter

  4. Step 2: First Impression • Very hard problem • Across quarters it is very hard to beat consistently a constant (Median -2 scores 1403) • Distribution of target is not symmetric • L1 vs. L2 matters • Strong but instable seasonality • The median delta varies • Scaling differences across papers • Max citations: 2400, Average: 15 • Only 10% of papers ever had more than 5 in a quarter

  5. Step 3: Training Considerations • Normalization • Large variation in average citation counts → Normalize by average of the last 3 quarterly counts • Selection of Training Period • Potentially non-stationary due to conference schedules, age of database, ... → Keep entire dataset • Selection of Training Observations • Evaluation only on papers with at least 6 citations → Only use observation where the last citation count was at least 6

  6. Step 4: Model Class Selection

  7. Ordered Probit Model Y* 4 3 2 1 0 x X

  8. Results and Comparison • Final Model: • 10 categories: -7, ..., +2 • Ordered Probit with special prediction • Curtail predictions between -4 and 0 • Results: • Winning Participant: 1329 • Best constant (-2): 1403 • Curtailing Predictions between -4 and 0: 1360

More Related