Wednesday (7/21) Response

(1) Overview of Model

• The dataset from this model looks to examine how various features of an individual can help to predict someone’s wealth. The specific features are rather simple, being age, gender, size, and education level. Due to there being so few features, setting the feature columns wasn’t that complex of a task. Age was the only feature that I placed in the numeric column. Age also was bucketized along with edu. The remaining features, size and gender, were both placed in indicator columns. The final step I did prior to training was create a crossed column for edu and age since they are highly correlated. After having som issues with loss appearing as nan, I changed the batch size to 132 and these issues were resolved.

(2) Wealth = 2

• The model appears to be very successful, with an impressive accuracy and minimal loss. The model does not deserve too much praise because most of its success can be attributed to the fact it is the least common wealth class by far. This makes it much easy to train because there is much less data to examine. With this being said this model should not be considered that impressive due to the limited amount of class 2 wealth points available to train.

(3) Wealth = 3

• This model was still ok, but far less successful than the previous wealth class. Both accuracy on the training group and validation group fell almost .1. This can be associated with the increase in individuals who fall within the wealth 3 range, making the model more complex and therefore harder to train.

(4) Wealth = 4

• This model was far, far worse than the previous two, reaching a peak accuracy of around 0.66. The training loss appeared to have a consistent decline, but the validation loss was a little unpredictable. With this odd movement in validation loss, the validation accuracy saw odd movement as well, with it ending lower than it started.

(5) Wealth = 5

• This model once again, had an awful accuracy. The validation and training accuracy were both subpar, peaking at around 0.62. The one improvement with this model compared to previous wealth classes is that this is the first time an increase in validation accuracy can be seen. In all the other model’s validation accuracy remained relatively stagnant, where here it begins at 0.54 and improves to 0.61 by the final epoch.

(6) Final Observations

• A definite trend can be seen within this model. As the target wealth class increases, the accuracy begins to plummet. I’m not sure if this factor can be attributed to the increase in sample size as the wealth class goes up, which would explain why 5 and 4 had similar accuracies since they had a similar number of people, or if there was some sort of error on my end in assigning feature columns. All in all this was an interesting model that with more editing of feature columns and other factors hopefully could become much more accurate.