Dear Naeem A. ,

In this tutorial, I’ll show you how to create a single label classification model in Google AutoML. We’ll be using a dataset of AI-generated faces from generated.photos. We’ll be training our algorithm to determine whether a face is male or female. After that, we’ll deploy our model to the cloud AND create the web browser version of the algorithm.

https://medium.com/wao-ai/single-label-image-classification-model-with-google-automl-4e978bf259f1

Regards,

Shafagat

Dear, since you have some data belonging to a single label then others are not belonging to that label. In this way it is a binary classification having two classes belongs and does not belong

Dear Naeem A. :

The single-label classification occurs when input data comprises labeled data belonging to one of several classes.

As for as your case is concerned, you have two classes (effective against a target, Not effective against a target). In this way, you have two classes. You can use any binary classifier to solve this issue. Furthermore, You can learn from these links about the working of single-label classification.

• https://blog.workaround.vercel.app/blog/single-label-image-classification-with-keras

Regards;

Ehtisham

Shafagat Mahmudova the link you provided is unrelated to the except for its title.

Rohail Gulbaz and Ehtisham Lodhi , maybe I didn't phrase it well enough, my dataset isn't binary, I have the data for "effective against the target" but non for "others/not effective".

The link you provided started by labeling the faces first which is not possible in my cases (can't conclude effect or not by looking at the structure).

Dear, please use semi-supervised learning technique. Otherwise just ignore the label and use unsupervised learning.

In my opinion, you have to extract the particular "features" which make these specific compounds effective against a target (label 1). Use these features (not compound names) as the input features. You can then enlarge the dataset by adding other compounds (their respective features) which are NOT effective against the target (label 0).

ML or DL algorithms use the features to distinguish / classify different classes. If you provide the model with these features during training, the model will learn which features (or threshold values) make the compounds effective (1) / NOT effective (0) against the target.

Nairit Barkataki It's a machine learning task, so of course, the inputs are compound features and not their names, I don't think you can use compound names as inputs anyway :)

You're proposing adding "not effective" data to make it a binary classification but I've mentioned that I couldn't do that, the only way that would be feasible is if I purchased a few thousand compounds and tested them for efficacy in vitro, which I can't do atm.

Rohail Gulbaz I don't know how I could use an unsupervised approach for this, could you reach out to me if you had the knowledge and can afford the time?

Dear Naeem A.,

I understand that you will not be able to obtain 0-class data. Yet, as implied by several other comments, to infer class 1 probability you will need something to compare it against (there is no definition of light without the dark). Alternatively, you at least need variability in your class 1, hence, an idea which compound is more or less effective. Using clustering like k-means will produce sub-classes within your class 1, and I agree with you that this will not be very helpful in answering your question.

You can at any time use distance measures such as euclidian distance to calculate similarity between two points. This is essentially a pre-step of many clustering algorithms. However, this will make it very difficult to infer anything about effectiveness. The main reason being, that you do not know which (if any) of your features predict effectiveness. Even adding a few "unhelpful" features can render results entirely useless. In essence: Depending on how many of the available features are really connected to effectiveness, similarity might or might not have any predictive power.

So, here is what I would do (Disclaimer: I know nothing about chemical compounds and am sorry if this answers is not helpful); Look for any empirical data suggesting which feature (sets) are connected to the target label. Then only use these features to at least theoretically ensure the highest possible noise-to-information ratio. Depending on how big you data is, use raw features or fit a dimensionality reduction model (e.g., PCA) to try to extract the commonalities across all class 1 compounds and calculate distance metrics on these. By then transforming a set of "unkown" compounds with the same model or leaving them as is, you can say which of the unknown is most similar to the group of class 1 compounds. Thus, it will be possible to say than any certain unkown compound A is more similar to your 1 classes across all features than a certain compound B. This way you can at least identify the most promising candidates.

You do not have a single-label dataset (as opposed to a multi-label dataset). Yours is a zero-label dataset: in terms of information content, a dataset whose elements have the same label is strictly equivalent to a dataset without any label. In other words, it is completely useless, whether for supervised or for unsupervised learning. What you are trying to do is known in the literature as QSAR (Quantitative Structure-Activity Relationships), a field where machine learning has been extensively used, and where the central problem is, in general, the choice of relevant numerical features; in your case, however, you have the additional problem of simply not having relevant data. QSAR has made tremendous progress in recent years, but it will never work miracles.

Kirsten Zantvoort Thank you, your approach resembles my initial approach of using clustering methods to somehow differentiate among them but as you too mentioned, without an opposite class to compare it, it would lack usability and performance.

Hi, I had a similar case, in a different domain, at the beginning of Covid-19, I lead a research for predicting Covid from a sample of breath.

The labeled data in the training came from PCR results. The latter's false negative rate is known to be high, and although the undetected positives in the population was relatively small because anyway it is bounded by the proportion of the positive people, we needed a purer training set.

I developed an SVD model for the positive samples only, and used it as a single class classifier on the test set, by measuring the distance of the example test from the model.

For that end you have to set a threshold.

Think of how to learn a threshold on a cross validation set.

Hope that that gives you a starting point.

Best of luck,

Aviv

Aviv Gruber Thank you for your time, indeed a very interesting approach to addressing the problem. I'm not very familiar with SVD models but I can research them. Can you provide a reference for your approach or maybe your implementation (if it's open-source), thanks again :)

Hi, I think It's better to look at this problem from another angle. Apparently, there are some features that have been recorded. Try to find the most effective features with some tools like t-test and then try to find the relationships between features with unsupervised methods. If the new features don't follow the relationships, the label won't be 1. I know it's not clustering but in this case that all the data are valid, I think it's better the label to be ignored.