Reading Articles Online

TL;DR

This paper models the online article reading process as a decision problem balancing information gain and time constraints, proposing algorithms with competitive ratios linked to hint accuracy.

Contribution

It introduces a novel online reading model, connecting it to the Online Knapsack Problem, and develops algorithms with competitive guarantees based on hint accuracy.

Findings

Any alpha-competitive algorithm for Online Knapsack yields a (e + alpha)C-competitive algorithm for RAO.

Current best Online Knapsack algorithms imply a 3.45e C upper bound for RAO.

A threshold-based reading strategy is O(C)-competitive, constant when C is constant.

Abstract

We study the online problem of reading articles that are listed in an aggregated form in a dynamic stream, e.g., in news feeds, as abbreviated social media posts, or in the daily update of new articles on arXiv. In such a context, the brief information on an article in the listing only hints at its content. We consider readers who want to maximize their information gain within a limited time budget, hence either discarding an article right away based on the hint or accessing it for reading. The reader can decide at any point whether to continue with the current article or skip the remaining part irrevocably. In this regard, Reading Articles Online, RAO, does differ substantially from the Online Knapsack Problem, but also has its similarities. Under mild assumptions, we show that any $\alpha$-competitive algorithm for the Online Knapsack Problem in the random order model can be used as a black box to obtain an $(\mathrm{e} + \alpha)C$-competitive algorithm for RAO, where $C$ measures the accuracy of the hints with respect to the information profiles of the articles. Specifically, with the current best algorithm for Online Knapsack, which is $6.65<2.45\mathrm{e}$-competitive, we obtain an upper bound of $3.45\mathrm{e} C$ on the competitive ratio of RAO. Furthermore, we study a natural algorithm that decides whether or not to read an article based on a single threshold value, which can serve as a model of human readers. We show that this algorithmic technique is $O(C)$-competitive. Hence, our algorithms are constant-competitive whenever the accuracy $C$ is a constant.