Model

• Run grid search using Linear SVC
• move all gridsearch to a new notebook
• Select best parameters
• Evaluate model
• feature importance for phase2=0 and phase2=1
• Run by section
• save evaluation metrics, confusion matrix

Writeup

• Check false positives and false negatives
• Describe confusion matrix - nature of mergers, nace_code, year, article
• Describe confusion matrix using n-gram importance
• Writeup

Robustness

• change random seed
• change pre-process - lowercase, stopwords
• Compare metrics

Labels

1. Phase1 vs. Phase2
2. Conditions vs. W/o conditions
3. OPTIONAL: Competition vs. Anticompetitive (variation of conditions label added Art 8.3)

• N-grams, TFIDF
• word embeddings
• n-gram embeddings for SVC and get feature importance
• clustering by topic (BERTtopic?)
• check maximum token length

Option 1

• word embeddings + multilayer perceptron

Option 2

• legal-bert
• transfer learning
• how to finetune

• add function to parse by section for outliers
• change functions - extract_year, extract_simp_text (simp_text must be <=5 len and art6.1b)
• re-run parsing for all?

• option1: resample training data to oversample minority class
• option2: balance data by undersampling majority cases
• if option2: put excess to test set
• put simplified cases to test set

Visualization

• wordcloud per section using n-gram features
• wordcloud of all sections n-gram features for svm_wc main and robustness
• [ ]

Look for pattern on which cases are being correctly predicted or wrongly predicted.

• Check TP and TN
• Check FP and FN
• Check on both training and test sets

Baseline model

• Remove obvious words?
• Run grid search using LOGIT and SVM
• Select best parameters
• Evaluate model
• feature importance for phase2=0 and phase2=1
• For reproducibility: Create pipeline to vectorize, oversample/undersample, train, evaluate

Labels

1. Phase1 vs. Phase2
2. Conditions vs. W/o conditions
3. Competition vs. Anticompetitive (variation of conditions label added Art 8.3)

main research question

• single_bidder: 1 if there was no competition for the contract
• contract_eu_funded: 1 if the contract was supported by EU funding
• est_value: the raw estimated value of the contract (currency/VAT unclear)
• contract_value: the raw value of the contract (currency/VAT unclear )

• use preprocess_function
• save pre-processed df again to remove obvious words and duplicates plus other data cleaning in baseline notebook

Questions

• What to do with simplified decisions? probably disregard
• How many models to run? Main: logit, svm, legal_bert
• Only 1 logit for baseline full_text with phase2?
• Why use precision-recall, f1-score, instead of accuracy?

Experiments

• Experiment1: using label1:phase2 vs. label2:wc (with conditions) full text
• Experiment2: using single section texts from four main sections
• Experiment 3: using word embeddings for deep learning