I'm relatively new to data science / machine learning (yes, I know) and am experimenting with text analysis. I only want a relatively naive approach and am looking to know whether my approach is valid procedurally even if I may not get perfect results. I want to compare my model to a fulltext tf-idf search of specific skills/keywords against an index of resumes.
The task: Classify job descriptions, rather cluster them in an unsupervised way. Eventually I want to match resumes against them as a basic implementation of my idea. I was thinking that a naive implementation of word embeddings and kmeans would be fine. Basically, extract document_vectors from specific job descriptions, then cluster the job descriptions based on these vectors. This would create a ghetto "labeling" system so that I could then apply a classification algorithm Then I would use resume raw_text and extract a new vector, applying a simple classification it into one of these clusters (regression).
To summarize
1. Use a large database of job descriptions
2. Tokenize the text, and apply doc2vec and generate document vectors
3. Cluster the documents and group similar job descriptions
NOTE: as to how many clusters to use, that will have to be tuned
4. Tokenize the raw_text of a resume
5. Use regression (or some technique) to find the closest cluster based on the resume vector compared to the cluster vector.
Now, I realize there are most likely some huge holes in the way I'm doing this (I'm using intuition based on no training). For example, my assumption that I can even match resume to job description is based on the idea that these are even comparable text structures - and how good the matching is is also dependent on how good the clustering of documents is.
I would like to know if this approach is good for something naive or whether I'm missing anything critical. I'm using nltk, scikit-learn, and gensim Word2Vec
