Creating a model for "classifying" behavior I am working on creating a model for "classifying" behavior on a scale from "good" to "bad". I have a largish dataset where each row represent an individual and each column an aspect of the behavior it can be amount of time spent, changes in time spent compared to a previous time period, number of distinct sessions, how much of some actions that an individual has performed etc. 
All in all there is currently 24 variables (a number which may change in the future). Now I want to classify the individuals on a scale from "good"/"normal" to "bad"/"abnormal" behavior (this scale must be continous and not descrete). 
My initial approach has been to do a principle component analysis (PCA) on the 24 variables and found that the first 3 PC's account for 70% of the variance in the data (adding two more I have about 90%). This seems promising to me since those individuals with "abnormal" behavior have larger values for at least some combination of the variables, e.g. more time spent, higher frequency of certain actions.
My problem is how I should go about reducing the dimensions further. I want to have all the individuals mapped on a single dimension from good to bad.
Since I am quite new to this type of modeling I would appreciate any hints/suggestions. Would it be a good idea to do a simple linear regression of the three princple components? Should I use more of the PC's to account for more of the variance in the data? Is there another method I should try to model this?
I am using R if that is relevant. 
 A: There is no "golden bullet" to solve such problem. Without prior knowledge about the nature of variables and their relationship with behavior the answer cannot be given. So what can be done at best. Let us think about the sources of variations in your data. The first reason is the behavior of individual. The second may be his temperament. The third - user skills, then time of day etc. I don't know the sources, since I'm not involved into problem, but you should know them. So think and write the list of possible variation sources. Then think about each your variable: what is the main factor that influence its value? Is it the kind of behavior? Or, maybe, gender (or any from your list of sources) has greater effect? Remain for analysis only that variables, where kind of behavior is supposed to have greater effect. Then do PCA. In such case the first PC (it corresponds to a source of greater variation) will incorporate the behavior. The scores of the first PC are the individuals' positions on your scale, and the loadings - relative influence of your variables on the behavior. Since the direction of PC vector is arbitrary, look where are "bad" values, and where are "good" (it may be that "good" behavior has positive values and "bad" - negative, and may be vice versa). Look on the loadings. The most of them will describe the right direction (e.g. the greater number of sessions corresponds to "better" behavior). But few can have unexpected (wrong) direction (due to chance or mistakes in your prior knowledge). Remove that variables with unexpected direction and repeat PCA once more. Now you have your behavior scores. Finish.
