Context

Inferring a behavioral model from users’ navigation patterns in a web application helps application providers to understand their users’ interests. It is essential to automatically identify and generate such models as the volume of daily interactions with applications are enormous.

Objective

The goal of this paper is to incrementally generate such an automated user behavior model with no instrumentation for understanding users’ interests in large-scale mobile and desktop applications.

Method

We propose an approach to fully automate the behavioral model generation for large-scale web applications. Our proposed solution infers a reward augmented behavioral model using a reinforcement learning method by 1) dynamically generating a set of probabilistic Markov models from the users’ interactions, 2) augmenting the state of the model with reward values. Our analysis engine of the proposed solution evaluates the evolving properties of interaction patterns against the inferred behavioral models using probabilistic model checking.

Results

We evaluate the utility of our approach by using it on a large-scale mobile and desktop application. In order to show that it is assigning meaningful reward values, we compare these values with results from Google Analytics (as a state-of-the-art approach). Empirical results indicate that our approach is not only compatible with the results from Google Analytics, but also can provide information in situations, where Google Analytics data is not available.

Conclusion

In this paper, we present a novel stochastic approach to (1) generate user behavioral models for mobile and desktop web applications, (2) automatically calculate the states’ rewards, (3) annotate and analyze the models to verify their quantitative properties, and (4) address many limitations found in existing approaches.

中文翻译：

---

在大规模 Web 应用程序中自动推断用户行为模型

上下文

从 Web 应用程序中用户的导航模式推断行为模型有助于应用程序提供商了解其用户的兴趣。由于与应用程序的日常交互量是巨大的，因此自动识别和生成此类模型至关重要。

目标

本文的目标是逐步生成此类自动化用户行为模型，无需借助工具来了解用户对大型移动和桌面应用程序的兴趣。

方法

我们提出了一种完全自动化大规模 Web 应用程序行为模型生成的方法。我们提出的解决方案使用强化学习方法通​​过以下方式推断奖励增强行为模型：1) 从用户的交互中动态生成一组概率马尔可夫模型，2) 用奖励值增强模型的状态。我们对所提出的解决方案的分析引擎使用概率模型检查来评估交互模式的演变特性，以针对推断的行为模型。

结果

我们通过在大型移动和桌面应用程序上使用我们的方法来评估它的效用。为了表明它正在分配有意义的奖励值，我们将这些值与谷歌分析的结果进行比较（作为最先进的方法）。实证结果表明，我们的方法不仅与 Google Analytics 的结果兼容，而且可以在 Google Analytics 数据不可用的情况下提供信息。

结论

在本文中，我们提出了一种新颖的随机方法来 (1) 为移动和桌面 Web 应用程序生成用户行为模型，(2) 自动计算状态的奖励，(3) 注释和分析模型以验证其定量属性，以及(4) 解决现有方法中发现的许多限制。