In the Reality Mining data, each instance has eight attributes and the values change according to logging time. The combination of eight GS-9137 price attributes composes an individual instance. If a new instance is equal to one of the previous events, the instance is regarded as old. Otherwise it is assigned as new. For the whole data, the ratio of old events changes and is represented in Figure
6. As shown in Figure 6, the ratio of old events gradually increases up to 32.8%. Kim and Park found the regularity in human behaviors from Reality Mining data [46]. In lifelong experience, we postulated that the human behaviors are repeated so that old/new judgment from the event stream is an important task to determine the next process such as updating the model or expecting the next situation. Figure 6 The ratio of old instances among total encoded instances during incremental learning. The overall ratio of old instance is about 32.8%. We also found that the distribution of attributes changes by time. When we divide the whole data into seven sections with the same instances, each section has different distributions of attributes. Figure 7 shows a change of distribution for one of attributes, location. Among over 30 values for location, four specific locations are dominant in the distribution.
However, the distribution is changes by the logging time. If the attributes are modeled by probabilistic approach, each section needs a particular probability distribution table. Therefore, in human behavior modeling, we need to consider both the regularity of the overall event stream and the irregularity of local fluctuation inside the attribute. Figure 7 The distribution of an attributes, location,
among eight attributes in Reality Mining dataset. The distribution changes according to the logging time. The primary goal of the experiment is to evaluate the proposed memory model that represents the properties of human-like recognition memory. Whenever the Reality Mining data are encoded, the results of the recognition judgment were compared with human behavioral performance. In addition, the dataset contains contextual information so that when a partial data with missing attributes appears, the recognition memory completes the missing part and expects the next context from the previous experience. In the following experiments, we investigate the structural configuration of the proposed memory model to reveal the most similar human performance. Furthermore, we figure out the characteristics GSK-3 of the model in nonstationary environment and evaluate the performance of expectation in comparison with conventional probabilistic model, Bayesian networks. 4.2. Experiment 1: Find Optimal Edge Configuration In the first experiment, we find the optimal hyperedge condition to derive acceptable results for judgment. Based on the hypergraph theory, the experimental dataset could be constructed into various hyperedge structures.
No related posts.