Author: Zen and the Art of Computer Programming
With the construction of smart cities, more and more attention has been paid to the convenience and effective use of road traffic. How to forecast, plan, and optimize traffic based on historical data is still a very important issue. At present, the research related to intelligent transportation mainly focuses on the prediction based on location information, that is, the location information collected by sensors is used to make predictions. However, the geographic location distribution in actual application scenarios is not necessarily accurate and credible, and even a large amount of data may be missing or incomplete, which makes it difficult for a prediction model based on location information to achieve true and reliable results. Therefore, the prediction based on sensor data has gradually turned to the method based on semi-supervised learning, and the training prediction is carried out by combining various types of data such as location information, geographic network data, and traffic index data. However, there are still some problems to be solved with this method:
- Incomplete or missing data prevents the model from learning useful information;
- There are many data sources, the data quality is uneven, and the noise affects the prediction effect;
- It is difficult to capture the relationship between different locations, and the model is difficult to cope with the complex traffic flow planning requirements in diverse environments;
- The model training period is long, it is difficult to adapt to the increasingly complex traffic system changes, and it poses a challenge to the real-time prediction ability;
- The optimization direction of the model is not conclusive, and the optimization methods and parameters required in different scenarios may be different. This paper will elaborate on the current research hotspots, challenges and latest progress in the field of intelligent traffic flow prediction and optimization based on semi-supervised learning from the above aspects. And try to give corresponding solutions and suggestions to these problems.
2. Explanation of basic concepts and terms
(1) Semi-supervised learning
Semi-Supervised Learning (SSL) is a machine learning task in which only some examples are labeled and the rest are unlabeled, and the system must be able to discover the characteristics of these unlabeled examples by itself. It can be considered as an extension of supervised learning, learning in data that is not fully labeled, or called "supervised" and "unsupervised" complement each other