This paper addresses the challenge of personalized news article recommendation by modeling it as a contextual bandit problem. The authors propose a new algorithm, LinUCB, which is computationally efficient and well-motivated from learning theory. The contributions of this work are threefold: first, they introduce LinUCB, a new contextual bandit algorithm; second, they argue that any bandit algorithm can be reliably evaluated offline using previously recorded random traffic; and third, they successfully apply LinUCB to a Yahoo! Front Page Today Module dataset, achieving a 12.5% click lift compared to a standard context-free bandit algorithm, with even greater advantages in data-scarce scenarios. The paper begins by formulating the problem as a contextual bandit, where the goal is to sequentially select articles based on user and article features to maximize total user clicks. It reviews existing bandit algorithms, including $\epsilon$-greedy and upper confidence bound (UCB) methods, and discusses their limitations. The authors then propose LinUCB, which efficiently computes confidence intervals for linear payoff models and works well for dynamic sets of arms. The evaluation methodology is described, showing that it is possible to evaluate bandit algorithms offline using logged data under certain assumptions. The experimental results demonstrate that LinUCB outperforms other algorithms, including context-free and warm-start methods, in both learning and deployment buckets. The use of user and article features is shown to significantly improve performance, with a CTR lift of around 10% compared to the baseline.This paper addresses the challenge of personalized news article recommendation by modeling it as a contextual bandit problem. The authors propose a new algorithm, LinUCB, which is computationally efficient and well-motivated from learning theory. The contributions of this work are threefold: first, they introduce LinUCB, a new contextual bandit algorithm; second, they argue that any bandit algorithm can be reliably evaluated offline using previously recorded random traffic; and third, they successfully apply LinUCB to a Yahoo! Front Page Today Module dataset, achieving a 12.5% click lift compared to a standard context-free bandit algorithm, with even greater advantages in data-scarce scenarios. The paper begins by formulating the problem as a contextual bandit, where the goal is to sequentially select articles based on user and article features to maximize total user clicks. It reviews existing bandit algorithms, including $\epsilon$-greedy and upper confidence bound (UCB) methods, and discusses their limitations. The authors then propose LinUCB, which efficiently computes confidence intervals for linear payoff models and works well for dynamic sets of arms. The evaluation methodology is described, showing that it is possible to evaluate bandit algorithms offline using logged data under certain assumptions. The experimental results demonstrate that LinUCB outperforms other algorithms, including context-free and warm-start methods, in both learning and deployment buckets. The use of user and article features is shown to significantly improve performance, with a CTR lift of around 10% compared to the baseline.