# Are there known techniques to transform features X classified as C to features Y classified as C'

I don't think the wording of my question is that clear myself, but I don't have any better words suitable for a title (on top of my head at least). I was wondering if given features X that is classified by a model M as class C, is there a way to find the features Y that is relatively "close" to X so that it would be classified as class C' by M.

I was thinking if some sort of clustering can help such as k-means and then getting the centroid of class C' and using that. The final idea is to get the difference between X and Y to be displayed. Does that sound reasonable? I'm not really a data scientist so want to check up on my thoughts.

If someone can suggest a paper or direction that would be much appreciated

EDIT: For clarification. The purpose of this is that I have a set of people's skills and their jobs and I want to be able to give an advice of what skills a person needs to cultivate for their desired job.

E.g I can program, have a CS degree, experienced with unix etc. and am classified as software developer (skills are codified into numerical values not text anymore) and I want to work as a chemical engineer. I want to know the skills I would need so that I can be classified as fit to be a chemical engineer.

So X is my set of skills, C is software developer, C' is chemical engineer, and Y is the set of skills appropriate for a chemical engineer that I am looking for.

• If this example is correct add a similar one for clarity: I have a set of profile (side) pictures of blue cars classified as 'car', I want a set of profile pictures of blue motorcycles classified as 'motorcycle'. Two sets of images (pixels as features) are close since they are both vehicles, blue, and photographed from the side. – Esmailian Mar 15 '19 at 10:54
• @Esmailian I added some more detail of what I am trying to do. Thanks! – Btara Truhandarien Mar 15 '19 at 18:31
• For this task you first need a sizable data set of (skills, job title) for many individuals to begin with, specially chemical engineers. Your personal information is not enough. Then, a fast approach would be to find a set of chemical engineers that the distance of their features to yours is the smallest. This is a good start. – Esmailian Mar 15 '19 at 18:41

This problem can be addressed from an Economics perspective, even with existing data and without complicated algorithms. There is some existing research in the Labor Economics field about transferability of skills between occupations. For example, this "The estimation methods of occupational skills transferability" paper: https://link.springer.com/article/10.1007/s12651-016-0216-y

There is one method it discusses, referencing another 2016 paper by Ormiston, that could be helpful to you. It involves calculating "the ratio of shared, standardized knowledge, skill and ability categories across occupations".

Here are some relevant excerpts:

It "employs a “skills” approach by explicitly estimating the proportion of knowledge, skills and abilities (KSAs) utilized in occupation i that can also be applied in occupation j."

"To estimate this proportion, it used data from O*NET – the U.S. Department of Labor’s occupational database – to examine the commonality of occupations across 120 standardized knowledge, skill and ability dimensions."

"O*NET examines each occupation and measures each component on the level (0–7 scale) of proficiency needed to function within the profession and the importance (1–5 scale) of each component to the occupation; multiplying these values together generates a comprehensive score, $$S_{im}$$ , for each component m of occupation i."

There is a formula given that

"estimates the skills transferability across 500 occupations based on the three-digit occupational codes and then aggregates these transferability estimates into 22 occupations in the two-digit level by taking the arithmetic mean within each two-digit category for the comparison with Shaw’s skills transferability. As a result, this study derives the 22 × 22 occupational skills transferability matrix based on two-digit U.S. census code, representing the estimated proportion of occupational human capital transferability."

There is an example given:

The data indicate that HR managers can apply only 16.01 of their 19.68 points utilized in their current profession into a new role as an industrial engineer. However, given differences in occupational skill portfolios, HR managers could apply all of their required knowledge in engineering and technology (1.03 points) and sales and marketing (5.99) in a new position as an industrial engineer. Summing the amount of shared points between occupations i and j (23.03) and dividing by the total points employed in occupation i (26.70) results in an estimated transferability rate of 86.30 % (t ij = 0.863).

You could use this data and methods to then figure out the areas where someone's skills are lacking and make recommendations from those findings.

• Interesting! I did use O*NET but didn't use their scale factors. I will look more into it, and see where I can go. Thank you! – Btara Truhandarien Apr 18 '19 at 4:39

As far as i can understand , you are interested in Adversarial training techniques. check this out: https://www.deeplearningbook.org/contents/regularization.html @7.13 Adversarial Training and the image example of the panda. Although the concept is opposite, i.e it finds out points near and similar to the existing classified regions which the model classifies wrong but sometimes the points are close to be identical like: the panda example in the reference at page 265.

• do you know anything that is closer to what I was describing? – Btara Truhandarien Feb 19 '19 at 5:12