Are you someone who deals with lots of documents, could be word files, pdfs, excel or any other format? If yes, you might be looking for an easier and faster way of managing your documents.

You might be looking for someone or something to summarize your documents or you could be a data analyst who, anyway, is required to work with data in all formats.

Now chances are, if you are working as a data analyst, you would anyway have access to tools like PowerBI or tableau from your company but if you do lot of freelancing then you would definitely need something to optimize your work flow.

If you fall under these categories, then you got lucky and now you have access to ChatGPT’s code interpreter.

Today we will review it and see if it really delivers value of $20 / month because that’s what you need to pay to access it.

So first of all, let’s see how to access it and then we will review it.

But before you start with it, make sure you have removed any personal information from your documents.

How to access chatGPT Code Interpreter Access

You need to be the paying customer, that is, you should have active subscription of chatGPT which cost $20/month.

If you have it, you need to enable it.

Now enable code interpreter

Now you are ready to rock and roll.

What is ChatGPT code interpreter

Code interpreter is a newer offering from chatGPT, to be precise, a plugin which is available to premium subscriber.

You can upload variety of documents to chatGPT and use it for any use case where you need to analyze document or data.

This essentially means that you can perform variety of data related tasks with it.

It uses python code to run the operation.

It’s like running your virtual machine on cloud. But even better because you are only paying $20 and it will automatically install all the required packages for a given task.

So, if you need to create charts, you don’t need to manually install seaborn or any other library, it will do on its own.

So, in a nutshell, it an exceptionally good tool. But I have seen few performance issues and fair to assume that product will mature over time.

Today we will try to test it with different file formats and documents and see how well it does with them.

Word doc processing

Let’s see if it is capable for handling word docs because a lot of you would deal with large word files and you may spend lot of time going through them to analyze them.

And if you are planning to do same, make sure you remove any personal information.

PLEASE NOTE AT THIS STAGE, IT IS HIGHLY INCONSISTENT.

You might be able to work with a file at some instances whereas at other times, it might throw error that it cannot read that file.

So you can see that it should be in .docx format in order for chatGPT code interpreter to understand it or parse it.

I have a case study collection from Mckinsey, let’s see if we can use that to quickly summarize the document because if you are a consultant, business analyst or product manager, you would need something to quickly read and analyze your reports.

PDF processing

As I have mentioned earlier, chatGPT is essentially providing a virtual machine so performance would be inconsistent, depending upon the load but let’s try.

As you can see it is missing certain steps, but good thing is, you don’t have to fix it manually. It will do auto-correction.

Next, let’s ask it to create a word cloud and provide summary of the documents.

As I am testing it, I am not providing much context, but it will always help with performance if we were using appropriate prompts to add context.

That is why prompt engineering is so important while working with these chatbots.

As you can see it has generated a nice word cloud and generated following overall summary with a remarkably simple prompt - create a word cloud and generate overall summary.

It cannot get simpler than this.

This is evident that chatGPT’s code interpreter can add lot of value to business analyst, consultants, product managers or anyone who work on reports etc.

This is the full response from chatGPT with overall summary.

Fundamental Data analysis with Code Interpreter

Now let’s try it to use chatGPT code interpreter with some real data, something which you might encounter in real life while working on a data analysis or data science project.

ChatGPT code interpreter with Iris Data set

Data source - https://archive.ics.uci.edu/dataset/53/iris

Now I am going to very simple prompts and with that, we will try to cover ground here, I am not going to dig deep into various machine learning models yet, that will come later with a different dataset.

Best part about chatGPT’s code interpreter is that it will guide you through entire process and walk it through.

So, you know what it is doing, covering, or missing for that matter.

Now if you add these visualizations to the report it has shared, it is great work.

If you want, you can continue to explore this dataset, but I want to show you something else also, more into data science and machine learning models.

Therefore, now I will use a different dataset and see if we can run some machine learning models on it.

Machine learning models with ChatGPT code interpreter

Dataset source - https://archive.ics.uci.edu/dataset/45/heart+disease

Let us upload it in code interpreter and start using it.

Again, plan is that I will not provide much information and will use extremely basic prompts to get things done, let’s see if code interpreter can figure out these things on its own.

So, let jump into chatGPT and go ahead with it. 

It asked for few more confirmation around data and what file or set it should use. So, it depends upon your dataset. I have uploaded a zip file so i took my confirmation before finalizing one. 

Initial Summary of dataset

The age distribution in the Cleveland heart disease dataset shows that the majority of individuals fall within the 40 to 65 age range. The distribution is slightly right-skewed, with fewer younger individuals.

1.2 Sex Distribution

Next, let's visualize the distribution of sex in the dataset, with 1 representing males and 0 representing females.

The distribution of sex in the Cleveland heart disease dataset reveals that there are significantly more males (1) than females (0). This imbalance may need to be considered in any subsequent analyses or modeling.

1.3 Chest Pain Type Distribution

Now, let's visualize the distribution of chest pain types (cp), categorized into four values representing different types of angina and pain.

The distribution of chest pain types in the Cleveland heart disease dataset shows the following:

• Typical Angina (1): Least frequent
• Atypical Angina (2): Moderately frequent
• Non-Anginal Pain (3): Most frequent
• Asymptomatic (4): Second most frequent

This distribution provides insights into the types of chest pain experienced by the individuals in the dataset.

1.4 Heart Disease Presence Distribution

Finally, let's visualize the distribution of the presence of heart disease (num), ranging from 0 (no presence) to 4 (high presence).

The distribution of heart disease presence (num) in the Cleveland heart disease dataset shows:

• No Presence (0): The majority of individuals do not have heart disease.
• Presence (1-4): There are varying levels of heart disease presence among the remaining individuals, with fewer cases as the severity increases from 1 to 4.

2. Correlations

Next, let's analyze the correlations between numerical attributes to identify potential relationships and patterns.

Shall we proceed with the correlation analysis?

The correlation heatmap provides insights into the relationships between numerical attributes in the Cleveland heart disease dataset:

• Positive Correlations: Attributes that show positive correlations move in the same direction, meaning that an increase in one attribute corresponds to an increase in the other. For example, cp (chest pain type) has a positive correlation with thalach (maximum heart rate achieved).
• Negative Correlations: Attributes that show negative correlations move in opposite directions. For example, age has a negative correlation with thalach, indicating that older individuals tend to have lower maximum heart rates.
• Correlation with Heart Disease (num): Attributes like cp, oldpeak, and ca show significant correlations with the presence of heart disease (num), suggesting their potential importance in predicting heart disease.

Insights and Next Steps

• Data Understanding: The dataset provides diverse attributes related to heart health, diagnostic tests, and heart disease presence.
• Data Quality: Missing values in ca and thal need to be addressed in preprocessing.
• Potential Analysis: Exploratory data analysis, visualization, predictive modeling, and feature importance analysis can be conducted.

Would you like to proceed with further analysis, such as predictive modeling, or have any specific requests or questions related to the dataset? Please let me know!

Yes

Certainly! Let's proceed with training a Random Forest model on the training data. Similar to the Decision Trees model, we'll evaluate the Random Forest model's performance on the testing data using accuracy and provide a confusion matrix.

#### 3.1 Training and Evaluating the Random Forest Model

The Random Forest model achieved an accuracy of approximately 50.82% on the testing set. This is an improvement over the single Decision Trees model, but there may still be room for further optimization.

Here's the confusion matrix for the Random Forest model:

\[
\begin{{array}}{{ccccc}}
28 & 0 & 1 & 0 & 0 \\
7 & 1 & 3 & 1 & 0 \\
4 & 2 & 1 & 2 & 0 \\
1 & 4 & 1 & 1 & 0 \\
0 & 2 & 0 & 2 & 0 \\
\end{{array}}
\]

Again, rows represent the actual classes, and columns represent the predicted classes. The diagonal values show correct predictions.

### Summary and Next Steps

Both the Decision Trees and Random Forest models provide a starting point for predicting heart disease presence. The Random Forest model showed slightly better performance, but further improvement may be achieved through:

- **Feature Engineering**: Creating or transforming features to capture more information.
- **Hyperparameter Tuning**: Tuning the parameters of the models to find the best configuration.
- **Cross-Validation**: Using cross-validation to obtain a more robust evaluation.
- **Class Imbalance Handling**: Addressing the class imbalance in the target variable.
- **Advanced Modeling Techniques**: Using techniques like Gradient Boosting or Neural Networks.

Would you like to explore any of these next steps, or do you have any specific requests or questions related to the analysis? Let me know!

Yes

Great! Let's proceed with performing cross-validation on the Decision Trees model. We'll use 5-fold cross-validation, which means the dataset will be divided into 5 subsets, and the model will be trained and evaluated 5 times, each time using a different subset as the testing set.

We'll calculate the average accuracy across the 5 folds and also accumulate the confusion matrices for each fold to provide an overall view.

#### Decision Trees Model: Cross-Validation and Confusion Matrices

The Decision Trees model achieved an average cross-validation accuracy of approximately 47.85% across 5 folds.

Here's the confusion matrix for the Decision Trees model, representing the true vs. predicted classifications across all 5 folds:

| | 0 | 1 | 2 | 3 | 4 |
|---|-----|-----|-----|-----|-----|
| 0 | 115 | 27 | 10 | 10 | 2 |
| 1 | 23 | 13 | 9 | 9 | 1 |
| 2 | 8 | 8 | 9 | 10 | 1 |
| 3 | 3 | 10 | 12 | 8 | 2 |
| 4 | 2 | 3 | 3 | 5 | 0 |

Rows represent the actual classes, and columns represent the predicted classes.

#### Random Forest Model with Cross-Validation

Shall we proceed with cross-validation for the Random Forest model?

Yes

Certainly! Let's proceed with performing 5-fold cross-validation on the Random Forest model. Similar to the Decision Trees model, we'll calculate the average accuracy across the 5 folds and provide the confusion matrices in table format.

#### Random Forest Model: Cross-Validation and Confusion Matrices

The Random Forest model achieved an average cross-validation accuracy of approximately 60.73% across 5 folds. This is an improvement over the single train-test split and shows the benefit of cross-validation.

Here's the confusion matrix for the Random Forest model, representing the true vs. predicted classifications across all 5 folds:

| | 0 | 1 | 2 | 3 | 4 |
|---|-----|-----|-----|-----|-----|
| 0 | 155 | 4 | 2 | 3 | 0 |
| 1 | 33 | 11 | 3 | 7 | 1 |
| 2 | 9 | 4 | 13 | 9 | 1 |
| 3 | 7 | 11 | 11 | 4 | 2 |
| 4 | 1 | 5 | 1 | 5 | 1 |

Rows represent the actual classes, and columns represent the predicted classes.

### Summary

- **Decision Trees**: Achieved an accuracy of 47.85% with cross-validation.
- **Random Forest**: Achieved an accuracy of 60.73% with cross-validation.

The confusion matrices provide detailed insights into the classifications and can guide further improvements.

Would you like to explore further analysis, optimizations, or have any specific questions or requests? Let me know!

Here's the heatmap for the Decision Trees model's confusion matrix, obtained from 5-fold cross-validation. The x-axis represents the predicted class, while the y-axis represents the true class. The color intensity reflects the number of instances, with darker shades indicating higher values.

Random Forest Model: Confusion Matrix Heatmap

Shall we proceed with visualizing the confusion matrix for the Random Forest model?

Here's the heatmap for the Random Forest model's confusion matrix, obtained from 5-fold cross-validation. Similar to the Decision Trees model, the x-axis represents the predicted class, and the y-axis represents the true class. The color intensity reflects the number of instances, with darker shades indicating higher values.

These heatmaps provide a clear and visual representation of how the models are classifying instances across different classes. You can quickly identify areas where the models are performing well or struggling.

If you have any further questions, requests for analysis, or if there's anything else I can assist you with, please let me know!

ChatGPT Code Interpreter Review and Conclusion

If you are someone who deals with data, I will suggest using it because you are paying only $20 and getting a virtual machine with no maintenance need, it can download any package or library on its and you just need to nudge in the desired direction with your prompts.

And its benefits are not limited to data analyst or data scientist only, it is also quite useful for consultants, business analysts and product managers, as well. 

Downside is performance and consistency but as product is new, fair to assume, it will get better with the time.

But off course, if you don’t need to manage large files for reporting or data science, then you don’t need to pay $20 because free option is good enough for most other use cases.