My goal is to create a regression model with text data where encoded text predicts a value, (news headlines, or article summaries, predicting number of clicks). The y is very left-skewed (few articles having very many clicks and most articles having few clicks). My models so far are not able to predict anything and I'm wondering whether I should keep trying. Having read a few papers with similar scenarios, you rarely see text predicting a value, in one paper the authors said that would be too difficult a task and created categories for their y to do ordinal regression.The most often seen case is logistic regression where text predicts spam/not spam, positive/negative sentiment, is it really too difficult to get a reasonable model from text where your y is not binary?
6
-
$\begingroup$ Is there a reason to expect the text to be predictive? For instance, can a human read the text and make an accurate prediction? Consider something like the MNIST digits where a human can look at the images and get the right answer basically every time, because those pixels encode all of the needed information. // When you write that the authors broke up their variable into categories to do ordinal regression, could you please elaborate? Perhaps give the reference. $\endgroup$– DaveCommented Sep 28, 2023 at 12:23
-
$\begingroup$ Could you try being more precise about what kind of problems you have in mind? Binary classification is popular, you would also probably find some examples of multinomial classification, predicting something like the number of clicks or sentiment (as continuous score) should be doable. However, using linear regression for creating summaries of articles is unlikely to have much chance of working. $\endgroup$– TimCommented Sep 28, 2023 at 12:37
-
$\begingroup$ @Dave You are right, I don't have a good reason to think the text would be predictive of clicks. I was hoping to find something in the text, maybe some specific words that make the headline more clickable, so at the end I could say: with this new headline you can expect this many clicks. But, maybe there is no such relationship. My dataset might simply be too heterogenous. My ideal dataset should probably have headlines of a specific event/topic with clicks, and maybe then I could extract which wording of the headline results in more interest. $\endgroup$– user3722736Commented Sep 28, 2023 at 13:39
-
$\begingroup$ @Dave In the article I was referring to: "A classification task with three classes, representing low, moderate and high responses to tweets, was defined and addressed using four machine learning classifiers." mdpi.com/2073-8994/12/6/1054 Their models look complicated and I'm just starting my adventure with data science. $\endgroup$– user3722736Commented Sep 28, 2023 at 13:40
-
$\begingroup$ @Tim Apologies for making it confusing, I'm not trying to create summaries using regression. I am trying to predict clicks or views based on heading text (I also have access to short summaries but headings are probably more useful here because people likely click on something based on the heading) $\endgroup$– user3722736Commented Sep 28, 2023 at 13:40
|
Show 1 more comment