Read the paper for a deep dive into our data collection.
Cyril Gorlla, Jared Thach, Hiroki Hoshida. Development of Input Libraries With Intel XLSDK to Capture Data for App Start Prediction. 2022. hal-03527679Users often face a myriad of minuscule slow-downs when navigating a computer. These slow-downs largely take the form of application loading times and we aimed to first collect user data by building custom data collectors in our previous work. The collected data can then be used to study user patterns and to extract insights via machine learning practices and models such as Hidden Markov Models (HMM) and Long Short-Term Memory (LSTM) Models. By leveraging these models, the inconvenience of accumulated application start-up times can be greatly reduced. Despite computer processor speeds increasing year after year, there are still high usage programs and processes that interrupt workflows in our daily lives. Whether that interruption be lag when opening Zoom to join a meeting, or waiting for Google Chrome to open a link embedded in a Word document, these micro (or in some cases, major) stutters can cause a process meant to be relatively smooth to be a large source of frustration in the daily lives of an end user.
One key point, however, is that these processes are actually often “scheduled” in a sense. That is, most people often repeat the same or similar tasks every day, and have a set routine. One example could be an office worker who opens Microsoft Excel almost every time they turns on their laptop at the office. These sorts of patterns can be studied, and by using machine learning, we can create predetermined schedules for users, allowing us to preload apps, or have processes ready before the user needs them. In other words, by analyzing user behavior, we can load an application the user would likely use next in advance, reducing wait and loading times.
By using Intel’s System Usage Reporting library and the accompanying XLSDK, we created “monitors” of user and computer activity, known as collectors or Input Libraries, and create activity logs which we can then analyze to provide preloading solutions for the user. In the previously linked paper, we have covered the development of these Input Libraries. We now briefly restate their descriptions. In order for us to create our own Input Libraries, the XLSDK provides a wide range of examples which can be used either as references or templates. Throughout the first ten weeks of working with Intel, we developed four different Input Libraries, each using a different template and measuring different categories of inputs from the user and computer. The first, the mouse_input Input Library, keeps a log of the cursor coordinates as the user moves the mouse around. Second, the user_waiting Input Library, keeps a timer based log of the cursor icon as the user uses the computer. The third, a foreground_window Input Library, creates a log entry whenever the foreground window (the window in front of all other windows) changes whether it be automatically (such as a notification pop up) or by user input (clicking the taskbar). Finally, the fourth Input Library is the desktop_mapper, which, when triggered by a change in the foreground window, maps all the windows on the desktop in z-order and stores pertinent information about each window e.g. position and size. Each of these Input Libraries are coded differently in fundamental ways, and measure changes in different ways as well. By using the data provided by Libraries like these, we can determine preloading schedules for the individual user.
