Recommender Systems (RS) in e-commerce are typically used to suggest products to online shopping customers, and now play a key role in product marketing strategies for major online retailers, such as Walmart and Amazon. The main goal of such systems is to predict likely future customer desires and to trigger purchases through the timely provision of product recommendations. Therefore, RS have become indispensable tools for both customers and retailers. However, most existing RS recommend products from the point view of customers (i.e. likelihood of customer purchase) but ignore one of the most important business goals: the optimization of revenue. Consequently, there is an increasing need to learn utility patterns online and provide near real-time utility-based recommendations. To address these challenges, we first define the utility of recommendation sets and formulate the problem of real time utility-based recommendation. Next, we consider that online transaction streams are usually accompanied with flow fluctuation, and propose an Adaptive Online Top-K (RAOTK) high utility itemsets mining model to guide the utility-based recommendations. Additionally, three variants of this algorithm are described and we provide a structural comparison of the four algorithms with discussions on their advantages and limitations. Moreover, to make our model more personalized, we also take the buying power of customers into account and propose a simple but effective method to estimate the consumers' willingness to pay. Finally, extensive empirical results on real-world datasets show that the proposed model works effectively and outperforms several baselines.