Index

1. Abstract
   • Features
   • Tech Stack
2. Usage
3. Endpoints
4. Search Space
5. Screenshots

Abstract

Problem Statement: Extracting meaningful insights from raw datasets and selecting the most effective machine learning models remains a significant challenge for both non-technical and technical users. Non-technical users and analysts often face steep learning curves due to complex tools and the need for coding expertise. On the other hand, technical users struggle with fragmented workflows that lack intuitive interfaces for rapid experimentation, hyper-parameter tuning, and performance comparison. This disconnect hinders efficient model development and slows down decision-making across teams

Context and Background: In the current data-driven era, organizations and individuals increasingly rely on data analysis for strategic actions. However, most available tools require programming knowledge or familiarity with data science workflows. This creates a barrier for non-technical users and business professionals who need to make sense of data without specialized skills. Even technical users encounter inefficiencies due to disjointed tools and unintuitive interfaces, making it harder to iterate quickly, fine-tune models, and compare results effectively

Purpose and Contribution: Synapse aims to democratize data analysis by developing a no-code, web-based platform that enables users to upload datasets, perform exploratory data analysis (EDA), and select the most appropriate machine learning model through a simple, conversational interface. The system bridges the gap between usability and advanced analytics by combining automation with natural language interaction

Methods and Approach: Synapse includes a user-friendly web interface with two modes: a visual dashboard for EDA and bayesian optimization, and a chatbot for natural language queries. Upon uploading a dataset file, the system automatically handles data preprocessing such as cleaning, encoding, and scaling. Users can visualize the dataset and interact with the chatbot. For model selection, Bayesian optimization is used to identify the best-fit algorithm for classification

Results and Conclusion: Synapse successfully simplifies complex data tasks, enabling users to analyze and interpret their datasets without writing code. It demonstrates that combining automation, natural language processing, and model optimization can make machine learning more accessible, thereby enhancing decision-making for users across technical and non-technical domains

Features

1. Engineered a full-stack real-time ML platform enabling the automation of ML workflows
2. Integrated Bayesian Optimization to autonomously tune hyperparameters for diverse models
3. Implemented an automated EDA pipeline generating insightful and interactive visualizations
4. Developed a robust customisable preprocessing engine with intelligent missing value handling, feature selection, and scaling
5. Embedded an AI chatbot to provide data-driven insights and statistical interpretations for technical and non-technical users

Tech Stack

1. Backend & Frameworks
   • Python (Flask): The core web framework used to build the application
   • Flask-SocketIO: Enables real-time, bi-directional communication for the EDA and training logs
   • Flask-SQLAlchemy: ORM for database management
   • Flask-Dance: Handles Google OAuth 2.0 authentication
2. Machine Learning & AI
   • Scikit-Learn: Used for standard algorithms (SVM, KNN, Random Forest, etc.) and metrics
   • Scikit-Optimize (skopt): Powers the Bayesian Optimization engine for hyperparameter tuning
   • XGBoost & LightGBM: Advanced gradient boosting frameworks integrated into the pipeline
3. Visualization
   • Matplotlib & Seaborn: Generates static charts like correlation heatmaps and pairplots (rendered to Base64)
   • Pygal: Used for interactive vector-based (SVG) visualizations
4. Frontend
   • HTML5 / CSS3 / JavaScript: Core technologies for the user interface
   • GSAP (GreenSock): Used for advanced animations and scroll triggers
   • Motion (Motion One): A modern animation library for UI transitions
   • JSZip & FileSaver.js: Allows users to zip and download generated charts directly from the browser
5. Database
   • SQLite: A fast and simple database used for storing user data and task information

Usage

To run the application locally, follow these steps:

First of all, ensure that you have git and Python 3.8+ installed on your machine. Then, run the following commands:

# Clone the repository
git clone https://github.com/msr8/synapse
cd synapse/src
# Install the required dependencies
pip install -r requirements.txt
# Run the flask application
python run.py

The application will be accessible at http://127.0.0.1:5000 in your web browser

Warning

These instructions are intended for local deployment only. For production deployment, use a production-ready server like Gunicorn or uWSGI, and consider using a reverse proxy like Nginx

Endpoints

URL Path	Description
/	Landing page
/learn-more	Information page about the project
/dashboard	User dashboard displaying tasks
/login	User login page
/signup	User registration page
/logout	Logs the user out
/login/google-authorised/	Google OAuth callback URL
/task/<int:task_id>	Main interface for a specific task
/api/auth/login	API to handle user login
/api/auth/signup	API to handle user registration
/api/auth/change-username	API to update the current user's username
/api/auth/change-password	API to update the current user's password
/api/upload	API to handle dataset uploads
/api/task/set-target	API to set the target column for a task
/api/task/change-taskname	API to rename a specific task
/api/task/delete-task	API to delete a task
/api/task/chatbot/initialise	API to start the LLM chat session
/api/task/chatbot/chat	API to send a message to the chatbot
/api/task/chatbot/reset	API to clear chat history

Search Space

We optimise over the following classification models using Bayesian optimization to find the best model and hyperparameters for a given dataset:

1) K-Nearest-Neighbours

Hyperparameter	Description	Type	Range / Values
n_neighbors	Number of neighbors to use	Integer	1 to 30
weights	Weight function used in prediction	Categorical	uniform, distance
metric	Distance metric to use	Categorical	chebyshev, cosine, euclidean, manhattan, minkowski, sqeuclidean

2) Support Vector Machine

Hyperparameter	Description	Type	Range / Values
C	Regularization parameter	Float	1e-4 to 1e+4 (log-uniform)
kernel	Kernel type to be used	Categorical	rbf,sigmoid, poly
degree	Degree of the polynomial kernel	Integer	1 to 3
gamma	Kernel coefficient	Categorical	scale

3) Logistic Regression

Hyperparameter	Description	Type	Range / Values
C	Inverse of regularization strength	Float	1e-6 to 1e+6 (log-uniform)
penalty	Norm used in penalization	Categorical	l1, l2
solver	Optimization algorithm	Categorical	liblinear, saga

4) Decision Tree

Hyperparameter	Description	Type	Range / Values
criterion	Function to measure split quality	Categorical	gini, entropy
splitter	Strategy used to choose split	Categorical	best, random
max_depth	Maximum depth of the tree	Integer	1 to 10
min_samples_split	Min samples required to split node	Integer	2 to 10
min_samples_leaf	Min samples required at leaf node	Integer	1 to 10
max_features	Number of features to consider	Categorical	None, sqrt, log2

5) Random Forest

Hyperparameter	Description	Type	Range / Values
n_estimators	Number of trees in the forest	Integer	10 to 100
criterion	Function to measure split quality	Categorical	gini, entropy
max_depth	Maximum depth of the tree	Integer	1 to 10
min_samples_split	Min samples required to split node	Integer	2 to 10
min_samples_leaf	Min samples required at leaf node	Integer	1 to 10
max_features	Number of features to consider	Categorical	None, sqrt, log2

6) Extra Trees

Hyperparameter	Description	Type	Range / Values
n_estimators	Number of trees in the forest	Integer	10 to 100
criterion	Function to measure split quality	Categorical	gini, entropy
max_depth	Maximum depth of the tree	Integer	1 to 10
min_samples_split	Min samples required to split node	Integer	2 to 10
min_samples_leaf	Min samples required at leaf node	Integer	1 to 10
max_features	Number of features to consider	Categorical	None, sqrt, log2

7) Gradient Boosting

Hyperparameter	Description	Type	Range / Values
n_estimators	Number of boosting stages	Integer	10 to 100
learning_rate	Shrinks contribution of each tree	Float	1e-6 to 1 (log-uniform)
max_depth	Maximum depth of estimators	Integer	1 to 10
min_samples_split	Min samples required to split node	Integer	2 to 10
min_samples_leaf	Min samples required at leaf node	Integer	1 to 10
max_features	Number of features to consider	Categorical	None, sqrt, log2

8) Light Gradient Boosting Machine (LGBM)

Hyperparameter	Description	Type	Range / Values
n_estimators	Number of boosted trees	Integer	10 to 100
learning_rate	Boosting learning rate	Float	1e-6 to 1 (log-uniform)
max_depth	Maximum tree depth	Integer	-1 to 15
num_leaves	Max tree leaves for base learners	Integer	10 to 50
min_child_samples	Min data needed in a leaf	Integer	5 to 20
subsample	Subsample ratio of training instance	Float	0.5 to 1.0
colsample_bytree	Subsample ratio of columns per tree	Float	0.5 to 1.0
reg_alpha	L1 regularization term	Float	0.0 to 5.0
reg_lambda	L2 regularization term	Float	0.0 to 5.0

9) Ada Boost

Hyperparameter	Description	Type	Range / Values
n_estimators	Maximum number of estimators	Integer	10 to 100
learning_rate	Weight applied to each classifier	Float	1e-6 to 1 (log-uniform)

10) Bagging

Hyperparameter	Description	Type	Range / Values
n_estimators	Number of base estimators	Integer	10 to 100
max_samples	Number of samples to draw	Float	0.1 to 1.0
max_features	Number of features to draw	Float	0.1 to 1.0
bootstrap	Draw samples with replacement	Boolean	True, False
bootstrap_features	Draw features with replacement	Boolean	True, False

Screenshots

Figure 1: Landing Page

Figure 2: Configurable Options for EDA, Preprocessing, and Bayesian Optimization

Figure 3: Target Column Selection

Figure 4: Feature Columns Distributions

Figure 5: Correlation Heatmap and Mutual Information Heatmap

Figure 6: Pairplot Visualization

Figure 7: Real-time Bayesian Optimization Results

Figure 8: Chatbot Interface

Figure 9: User Dashboard

Figure 10: Learn More Page

Figure 11: FAQs