基於關聯式分類法在手機上進行人類行為活動辨識

Abstract

人類行為的辨識一直以來都是一個很重要的議題，尤其近年來智慧型手機的普及化，這個議題更加的重要。在智慧型手機上辨識人類行為可以帶來許多的好處，舉個例子，如果我們可以辨識使用者現在正在開車，智慧型手機可以自動的切換到開車模式，此模式可以自動語音接電話、自動念出簡訊…等。這些應用可以讓手機更加的智慧並且讓人們生活更加的便利。正常來說，辨識系統就是資料探勘領域內分類的技術。辨識系統首先需要收集足夠多的使用者資料來建造一個分類器，並用這個分類器來判斷使用者的行為模式。所以選擇一個正確的分類器會很大程度影響系統的好壞，在這篇論文中，我們選擇以規則為基礎的分類器當作我們的核心模組，此種分類器的核心概念在於發掘隱藏在特徵點間之有價值的合作模式，這非常適合用於人類行為辨識，在文章中我們會詳列出為何我們選擇此種分類器和其好處。但為了使用規則分類器，我們必須先將收集的資料離散化，因為離散化後的資料型態是建立規則的基礎，經由我們的詳細分析，在離散化這部分我們採用一個著名的方法MDLP來完成。除此之外，實際建立辨識系統在智慧手機上還需要考慮的使用者體驗，為了讓使用者獲得最佳的體驗，我們又遇到了兩個問題，分別為模型大小的限制和冷啟動的問題。對於模型大小限制的問題，我們制定一個新的排序規則來使的我們的模型在限縮大小的情況下仍有優良的表現，基於新的排序方法我們提出SCCBA來使的我們的模型一定符合限制並有優良的表現。另外對於冷起動問題，我們提出CRR演算法來解決他，我們的演算法會利用現有之他人模型來幫助冷起動使用者建立一個不錯的個人化模型。最後我們收集了多人的真實資料來驗證我們的方法，實驗結果證明我們提出的辨識系統能準確的辨識人類行為並且能有效的解決上述兩個議題。

Human activity recognition is an important issue on using smart phones. Recognizing user activities on smart phones provides more suitable information, phone settings and recommendation for users. For example, if smart phones could infers users in driving activity, smart phones could set driving scenario setting, such as read text messages and oral dial the phone number without touching for users to achieve the purpose of safety driving. Without loss of generality, activity inference problem is the classification problem in data mining field. In the classification problem, the training dataset with some labels is provided for building classifiers. In activity inference, we implement a logger to collect context information, including location information and sensor's information. Moreover, each user needs to label his/her activity labels. From the training dataset collected, in this paper, we explore associative classifier as activity inference classifier. Note that the associative classifier aims at mining association rules among context information hidden in the training dataset. As such, associative classifier is a rule-based classifier. Thus, rule-based classifier is more suitable in smart phones due to the resource constraints on smart phones. To mine association rules among sensor's data, one challenge issue is to discretize numerical data into discrete data types. These discrete types are the basic units for association rule. Therefore, we adopt MDLP, which us the concept of data compression, as our discretization method. To build the model in smart phone, sometime model size constraint is another big challenge. For model size issue, we propose a classification method called model Size Constraint Classifier Based on Associated Rules (SCCBA). SCCBA can let model size fit the constraint and only lose less accuracy. To build a good recognition system need enough data but new user will not provide system these data. How to build a pretty good model for new user is called cold-start problem. For cold-start problem, we propose Cluster and Restructure Rules(CRR) algorithm which use other user's model to recognize new user's activity. To evaluate the performance of our proposed method, we collect real dataset from 50 participants in eight months. The experimental result shows that our framework has good recognition accuracy and a smaller size by using associative rule classifier.