As a bridge between the physical and cyber world, the Internet of Things (IoT) senses and collects a large amount of user data through different types of devices connected to it. As a general information filtering technology, the recommender systems can help to associate information with each other in the IoT and to recommend personalized services for users. However, in practical applications, the collected data is uncertain due to noise, sensor errors, transmission errors, etc., which in turn affects system performance. In order to solve the data uncertainty problem in the IoT-based recommender systems, we propose a new recommender framework with item dithering. In this framework, the list of recommendations generated by the recommender algorithm is stored in a newly opened storage space for the entire session of the interaction between the user and the system. When the user interacts with the system, the list is pushed to the user after being shaken. Based on the proposed framework, we designed IDither, an item-based dithering and recommendation algorithm to shake out irrelevant items through predetermined indicators, thereby retaining the items required by the user and recommending them to the user. Experiment evaluations on real datasets show that IDither is an effective solution for handling uncertainty in the IoT-based recommender systems. We also found that IDither can be viewed as a list updating tool to increase diversity and novelty. INDEX TERMS Recommender systems, Internet of Things, data uncertainty, dithering.

The ranking algorithm in the recommender system aims at optimizing accuracy during training so that it pays too much attention to the relevance of the individual and ignores the mutual influence between the items in the list. In response to this problem, we propose dither, a re‐ranking model for the recommender system. We deploy the re‐ranking algorithm as an independent module after the ranking algorithm to achieve the function of decoupling from it. Our model formalizes the re‐ranking problem as a multi‐objective optimization problem. It re‐ranks the initial ranking list by balancing multiple indicators to generate an improved list and updates the list during frequent user interactions with the system. Through a case study on the MovieLens 100 K data set, the workflow, and effects of the dither model are demonstrated. In addition, the re‐ranking algorithm shows the performance advantages of our model over existing methods.

No abstract