The final output file is ./output/dept_preds.csv. There are two columns:

• Job Titles: All the job titles provided in ./data/jobtitles_all.txt
• Department Predictions: The department predictions. There can be no, one or more than one department attributed to a job title.

Example

1. Clone the repo

   git clone https://github.com/msi1427/Leadbook-ML-Challenge.git
   cd Leadbook-ML-Challenge

2. Initialize and activate a virtual environment

   virtualenv --no-site-packages env
   source env/bin/activate

3. Install the dependencies

   pip install -r requirements.txt

4. Prepare the data

   python scripts/data_prep.py

   If you want to provide a different datafile with the same format of ./data/departments.json, but a different name:

   python scripts/data_prep.py --raw_dept_data <json_file_name>

5. Train the Phrase Similarity Model

   python scripts/train.py

   If you want to provide a different datafile with the same format of ./data/train.csv, but a different name:

   python scripts/train.py --train_data <csv_file_name>

   If you want to use a particular GPU for training:

   python scripts/train.py --gpu <gpu_index>

6. Inference on trained models

   python scripts/inference.py

   If you want to provide a different test datafile with the same format of ./data/jobtitles_all.txt, but a different name:

   python scripts/inference.py --test_data <txt_file_name>

   If you want to provide a different department datafile with the same format of ./data/departments_processed.json, but a different name:

   python scripts/inference.py --dept_data <json_file_name>

   If you want to use a particular GPU for training:

   python scripts/inference.py --gpu <gpu_index>

   I have trained 5 trained models. You can choose any of them for inference. By default, it will use 'paraphrase-MiniLM-L6-v2' model for inference.

   python scripts/inference.py --trained_model <model_name>

Problem Statement: Build a model which can classify given job title into departments. A job title may have more than one department or none.

Data: The data provided are two files:

• ./data/departments.json: There are 35 departments with some indicating phrases. The data format of that file is:

  [
  	{
  		"department_name1" : ['indicating_phrase_list']
  	},
  	{
  		"department_name2" : ['indicating_phrase_list']
  	}
  ]
  

• ./data/jobtitles_all.txt: There are 141564 job titles to attribute to departments. Each line is a particular job title.

Data Preparation: There are 2 parts in data preparation for training the phrase similarity models:

• Data Preprocessing: The ./data/departments.json file is converted to ./data/departments_processed.json file with the file format:

  {
  	"department_name1" : ['indicating_phrase_list'],
  	"department_name2" : ['indicating_phrase_list']
  }
  

  The phrases are preprocessed in 2 steps:

  1. The '/'s are replaced with 'or' and '&'s are replaced with 'and'.
  2. All the words are lowercased. This is done because the provided ./data/jobtitles_all.txt lines are lowercased.

• Train Data Preparation: The intuition here is that if two phrases are under a particular department, they are similar and if two phrases are not under a particular department, they are not similar. The train dataset is paired like that from ./data/departments_processed.json file. The ./data/train.csv have 113204 samples. The format of the train data is:

  phrase1	phrase2	score
  defence	space	0.9
  space	defense	0.9
  defence	animation	0.1
  defense	motion pictures and film	0.1

Training Phrase Similarity Model: I fine-tune Sentence Transformers to measure phrase similarity for my task. The train data ./data/train.csv is split into 90-10 as train-validation set. 5 best performing Sentence Transformer pretrained models were chosen as shown in the following table. While training, the batch size was 16 and the data were shuffled when building the DataLoaders. As the loss function, I use CosineSimilarityLoss and we use Cosine Pearson Index as metric. All the models are trained for 5 epochs and the model weights will be saved as ./model/{model_name}/. The performance on validation set is shown in the following table:

Inference on Trained Models: I use paraphrase-MiniLM-L6-v2 for my inference to get the output file ./output/dept_preds.csv because this model is much faster than other models while performing almost closer than other. In case of production, we need to consider these tradeoffs. The inference is done in 3 steps:

• Data Preprocessing: There are a lot of noises in the ./data/jobtitles_all.txt file. The preprocessing is done in 3 steps:

  1. The '/'s are replaced with 'or' and '&'s are replaced with 'and'.

  2. Remove all the punctuations.

  3. Remove all the noisy words. The noisy words I identified are the following:

     ["ceo ", "coo ", "cfo ", "cio ", "cmo ", "chro ", "cto ", "director ", "chief ", "president ", "vice president ", "vp ", "vice chair ", "board member ", "member ", "team member ", "team captain ", "owner ", "chairman ", "co - chair ", "co - chairman ", "senior "]
     

• Loading the Models: Since the trained models are of big size, they are currently stored in Google Drive. The models are automatically downloaded whenever specified. The default model is paraphrase-MiniLM-L6-v2. The corpus embeddings of department topics are generated.

• Inference: The inference is done in 3 steps:

  1. If all the words in the a job title is a noisy word, the title is ignored.
  2. If the job title is not English, the title is ignored. I used polyglot library to identify the language.
  3. The top 20 phrases similar to the job title is selected. If they have >=0.5 similarity score, the departments are taken into consideration and attributed to the job title.

  More about the output file is discussed in the first section.

• An annotated job title data as test set might be a good way to improve the performance further because as I mentioned in the last section, I used the top 20 phrases and >=0.5 similarity score for inference. Both of them were an educated guess from my previous experience. But an educated guess does not always ensure the results. An annotated set might help us tune these hyperparameters.
• I used some of the words as noisy words like ceo, director. This list of noisy words is also an educated guess and totally from my intuition. But, a comprehensive list by domain experts might give a better direction.
• A more extensive list of department and their topics might give us a better result in the long run.