Given the encouraging results with the bag-of-words (BOW) plus logistic regression model, we seek to replicate the performance with a simple neural network that mimics logistic regression.

The BOW plus logistic regression model (1) showed that most of the signal to discriminate between fake and real jobs is in some specific words. The sequence of words is not hugely consequential. Thus, we project that a simple neural network with a BOW (represented as a dense matrix) connected to an output node with sigmoid activation would produce similar results as the logistic regression model.

We build such a model in (2) and train it for 30 epochs. We then test the model on the held-out sample. This simple neural network model returns an accuracy of 98.70% and an F1 score of 87.18%. The results are nearly identical to those of BOW with logistic regression (accuracy 98.70% and F1 score 86.59%).

References

  1. Logistic regression plus bag-of-words model ↩

  2. Bag-of-words, neural network model with no hidden layer ↩