I have a dataset which consists of one text column (it is related to nlp), four categorical columns as features and one target column which has almost 90.000 categories. Also, the dataset has millions of rows. I have to use PySpark library. I researched multi-label classification and got familiar with the subjects like Binary Relevance, Classifier Chains and Label Powerset. My company wants me to do multi-label classification in order to predict 90.000 target categories.
However, i saw multi-label classification examples that have four or five target categories. To me, predicting 90.000 target categories using methods like Binary Relevance, Classifier Chains and Label Powerset is impractical or almost impossible when we think of maintenance of 90.000 ml models or one model which tries to predict thousands of target categories.
What is the suitable number of categories in order to do multi-label classification?
How should i approach to this problem?
I suggest you to follow https://scikit-learn.org/stable/modules/multiclass.html
https://www.depends-on-the-definition.com/guide-to-multi-label-classification-with-neural-networks/
https://medium.com/@saugata.paul1010/a-detailed-case-study-on-multi-label-classification-with-machine-learning-algorithms-and-72031742c9aa
Hi Muhammad Ali ,
Thank you for sharing your links that are about multi-label classification.
I looked at the links. They explain the topic. But, i am interested in my own case which is the feasibility of predicting almost 90.000 categories using methods like Binary Relevance, Classifier Chains and Label Powerset.
In theory there is no upper bound for the number of categories as long as you have a reasonable amount of samples per category, although in practice there are of course limitations.
You might consider a hierarchical approach, i.e. building a first classifier with a few (k) macro-categories, then build k new classifiers for the samples in each macro-category to predict sub-categories, and so on until you reach the desired granularity.
This might be especially useful if the sample distribution is not balanced among categories, because you could merge several small classes together at the beginning to separate them from over-represented classes to obtain balanced groups.
If you don't have prior knowledge to define such macro-categories and sub-categories, you might consider unsupervised clustering and decomposition methods to observe how your samples cluster together.
Hi Paola Galdi,
Thank you for your answer.
It sounds interesting. I don't have prior knowledge to define such macro-categories and sub-categories. Should i cluster only features or only targets or both features and targets?
What I would try is to explore how your samples (described by your features) group together, and then see if there are clusters containing only samples from a subset of categories: this would form a macro-category.
Hi Paola Galdi,
After k-means clustering (creating macro-categories/cluster ids), should i build classification models for each cluster's target variables (sub-categories) ?
You can try tokenization that will splits a texts into tokens and hence each label can be represented as number that can be used as a feature vector in machine learning. Scikit-learn provides a CountVectorizer and a TfidfVectorizer to vectorize text. Keras also comes with several text preprocessing classes - one of these classes is the Tokenizer. The labels can also be encoded using MultiLabelBinarizer. For instance, if you have 100 labels then it would be represented as 100 binary elements.
Hi Hikmet Yavuz
yes my idea is that you can train a first classifier to assign samples to the clusters that you found, and then build an additional classifier per each cluster. But as I mentioned before you might need to iterate this process if your sub-categories are still too big.
Thank you so much Paola Galdi !!! You've made my day. I will offer this method to my company.
Hi Ritika Lohiya ,
Thank you for your answer.
I think your advice is for my text feature.Right?
- Badges
- Science topic