Predicting-Sales

1. Load the CSV file - Result is a Pandas dataframe

   • Input:
     • Critic Score (Float with NaNs)
     • Publisher (Catagory)
     • User Score (Float with NaNs)
     • Genre (Catagory)
     • Name (String)
     • User Count (Float with NaNs)
     • Critic Count (Float with NaNs)
   • Output: Global Sales (Float)

2. Tokenize Names - Bag of Words

   • The loop that builds the vocab then counts how many times that is

   Put the Bag-of-Words code here later

   • Result: a list of Ints and how many Ints in that list is the size of the $n_b$ (this is vocab)

     • You will often see something like this - max vocab: 200 words words(drop infrequent) - This will keep the 200 most common words and drop the ones that don’t show up a lot

     NOTE: Sometimes you may see an error that comes up that looks something like this:

     Mat1 * Mat2, 
     	Shapes (1007, 2)
     	and (4, 97)
     

     This is why the shape of you data matters

3. Collect results from all rows (16720, $n_b$) - Save as tensor and call t_names

4. Take user_score column

   • Options:
     1. Remove rows with NaNs - this Pandas function is dropna()
     2. Imputation: inject mission values
        1. Use the average value to fill in all missing values - the Pandas function is fillna()
        2. Make it zero
        3. Train a model on the other columns to predict the user_score (do not use global_sales)
   • Result: tensor(16720, 1)

5. Take publisher column - One-Hot encoding without having to tokenize it - this Pandas function is get_dummies()

   • Result: (16720, $n_p$) - $p$ for publisher

6. Take genre

   • Result: 916720, $n_g$) - $g$ for genre

7. Figure out the total size

   torch.cat([t_names, t_user_score, ...])

   • Result: (16720, $n_{v} + 1 + 1 + 1 + n_p + n_g$)

8. Build a neuron

   model = nn.Linear(input_dim, 1) 

   Where input_dim refers to the number of input features that the model, expects to receive for each individual sample in the dataset.