SAI VYSHALI VAKATI

• PROJECTS
• STORYTELLING
• ABOUT
• CONTACT

In the dynamic world of e-commerce, coupons are a powerful tool to drive sales and enhance customer loyalty. However, without a personalized approach, these coupons risk being wasted on customers who may not need them or failing to engage those who would benefit the most. This case study examines how targeted coupon strategies can be tailored to specific customer segments—gold, silver, and bronze. By focusing on customers with lower satisfaction or purchase rates, while minimizing offers to those who are likely to buy regardless, the aim is to ensure coupons are used effectively to both increase sales and improve customer satisfaction.

I conducted a permutation test to assess the impact of applying discounts on total spending for gold customers. Customers were divided into a control group (no discount applied) and a treatment group (discount applied). The test aimed to determine if there was a statistically significant difference in total spending between the two groups.Null Hypothesis (H0):
H0: There is no significant difference in total spending between the treatment group (discount applied) and the control group (no discount applied).Alternative Hypothesis (H1):
H1: There is a significant difference in total spending between the treatment group (discount applied) and the control group (no discount applied).

Coupons did not increase total purchases for gold customers, as they are regular purchasers who tend to buy consistently with or without discounts.Customer satisfaction levels from Key Insight 1 showed that most gold customers are already satisfied with their purchasing experience.Recommendation: No need to offer coupons to gold customers, as they are likely to make purchases without any incentive.Recommendation for Business Benefit:Coupons can be reallocated to other customer groups (those with lower satisfaction or lower purchase activity).This approach can help attract new customers or drive more purchases from less engaged groups.The company can save money by avoiding unnecessary discounts for gold customers, ensuring a more efficient use of resources.

Key Insight 3: Permutation Test for Silver Customers
For silver customers, I performed a permutation test comparing the control group (no discount) and the treatment group (with discount). The observed difference in total spending was −115.10, with a p-value of 0.0, indicating a significant difference.The 95% confidence interval for the difference in means was (-21.96, 22.20), and the Cliff's Delta was 1.0, showing a strong negative effect. Similar to the gold customers, silver customers with discounts spent less than those without, suggesting that discounts may be reducing total purchases for this group as well.Key Takeaway for Silver Customers:1.Observed Difference: The significant difference in total spending for silver customers is larger than for gold customers but still negative, indicating that discounts did not increase total purchases for this group either.2.Satisfaction Levels: However, Key Insight 1 revealed that silver customers have low satisfaction levels compared to gold customers.Recommendation:Despite the negative spending impact, it might be beneficial to continue offering coupons to silver customers as a way to improve satisfaction and potentially drive future engagement.Recommendation for Business Benefit:Targeting low-satisfaction customers with incentives like coupons can help boost their overall satisfaction, potentially leading to more loyalty and engagement over time.By focusing on improving the experience of less satisfied customers, the company could foster long-term customer retention in this group.
Key Insight 4: Permutation Test for Bronze Customers
For bronze customers, I performed a permutation test comparing the control group (no discount) and the treatment group (with discount). The observed difference in total spending was 52.23, with a p-value of 0.0, indicating a significant difference in favor of customers who received discounts.The 95% confidence interval for the difference in means was (-11.42, 11.53), and the Cliff's Delta was -0.993, indicating a strong effect. This means bronze customers with discounts spent significantly more than those without discounts, highlighting that coupons are effective for this customer segment
Takeaway for Bronze Customers:
1.Observed Difference: The significant positive difference in total spending for bronze customers indicates that coupons have a positive impact on their purchases.2.Satisfaction Levels: Bronze customers have low satisfaction levels and low purchasing activity, making them ideal candidates for receiving discounts.3.Recommendation: Coupons are essential for bronze customers, as they not only boost total purchases but could also help improve customer satisfaction and engagement.Recommendation for Business Benefit:Offering coupons to bronze customers can increase their spending and help retain a less engaged segment.Targeting this group with coupons can improve satisfaction and drive longer-term customer loyalty.Data Limitation and Challenge:
1.Performing an exhaustive permutation test was not feasible due to a lack of computational resources, as it requires significant processing power to calculate all possible combinations.2.Additionally, a t-test was not suitable because the purchasing data wasn't normally distributed, limiting the statistical options available for analysis.3.The dataset was also limited to 350 records, but we still tried to derive as many insights as possible from the available data.

• Link

Project 2
Unveiling Market Sentiments: Positive, Negative, or Neutral Sentiment Analysis
INTRODUCTION

Refining Baseline Models: Impact of Stopwords in Financial ContextModel Experimentation: I tested two baseline models with different text preprocessing strategies using CBOW for Word2Vec embeddings to determine their impact on sentiment analysis:1.Model with Stopwords: This version included stopwords, under the hypothesis that in financial contexts, common words might carry significant sentiment cues.2.Model without Stopwords: This model excluded stopwords to concentrate on the main lexical content.Preprocessing Insights: Lemmatization was trialed but showed no beneficial effects on the results and was therefore omitted from further processing.Model Architecture: Both models employed LSTM networks to capture the sequential nature of text data. I integrated early stopping with a patience of three epochs to optimize training and avoid overfitting.Key Findings: Interestingly, the model that included stopwords demonstrated improved performance. This suggests that in financial texts, stopwords can hold contextual significance that enhances sentiment analysis accuracy. This outcome guided my subsequent adjustments and optimizations to the sentiment analysis model.

Improving the Baseline Model: Handling Unknown Words and Hyperparameter Tuning
Unknown Word Handling:My approach included two strategies:
Learning from Context: The model dynamically learns embeddings during training, utilizing the contextual information available within financial documents to infer meaning.
Random Initialization: For isolated or rare words, where contextual clues are sparse, I employed random initialization.

Optimal Configuration Achieved:
The combination of parameters — vector size of 300, window size of 25, negative sampling rate of 20, 256 LSTM units, and a dropout rate of 0.4 — demonstrated the best performance across all tested configurations. This optimal setup achieved a peak accuracy of 0.7313, outperforming all other parameter combinations in our hyperparameter tuning process.

Key Takeaways
1,Custom-Built Embeddings: Without relying on pretrained embeddings, I developed word representations from scratch. This approach allowed the model to adapt specifically to the unique language and context of financial text.2.Effective Handling of Unknown Words:By dynamically learning representations for unknown words and using random initialization when necessary, the model became resilient to new or rare financial terms—essential for real-world financial applications.3.Optimal Hyperparameter Configuration: Extensive tuning led to the ideal combination of vector size, window size, negative sampling, LSTM units, and dropout rate, achieving a peak accuracy of 0.7313 and maximizing model performance.4.Enhanced Financial Text Analysis: This custom-built model effectively captures sentiment in financial texts, making it a valuable tool for insights in finance, from market sentiment analysis to informed decision-making.5.Future Improvements: With everything built from scratch, the model provides a strong foundation for future enhancements. Potential next steps include experimenting with transformer models or incorporating additional financial datasets.

This project explores the potential of a data-driven, personalized coupon recommendation system that utilizes in-vehicle data to deliver relevant offers to users during their journeys. By analyzing user preferences, past purchases, and real-time location data, this recommendation system aspires to deliver timely discounts and promotions that align with users’ current contexts and needs. Our goal is to increase user engagement, boost customer satisfaction, and drive sales, all while ensuring that recommendations feel organic rather than intrusive.

Exploring Alternative Approaches: Logistic Regression and SVM

As part of my journey to build an effective recommendation system, I experimented with Logistic Regression and Support Vector Machine (SVM) models.
Logistic Regression served as a linear model that could reveal the influence of each feature on coupon acceptance. To account for class imbalance, I used SMOTE to boost representation of the minority class in the training set, ensuring the model had a balanced view of both outcomes. I also tuned key hyperparameters to optimize its learning rate and convergence. Logistic Regression, while interpretable, showed limited effectiveness with an accuracy of about 61% and struggled with complex patterns in the data.Soft Margin SVM was an attempt to explore a non-linear decision boundary for this classification task. Given the computational expense, I applied Stratified Sampling to work with a subset of the data and used ANOVA F-Value for feature selection, choosing the most relevant features. While SVM’s theoretical strength lies in separating classes with an optimal margin, it didn’t perform well on this categorical dataset, achieving an accuracy of around 42%. The categorical nature of the data, along with SVM’s sensitivity to scaling, made it less suitable for this problem, but it highlighted the need for models that handle non-linearities more effectively.Neural Network:Capturing Complex Patterns in User Behavior.
After establishing a baseline with Naive Bayes, I introduced a Neural Network to capture the complex, non-linear relationships in user behavior, improving the recommendation system's predictive accuracy.The Neural Network consisted of:1.An input layer representing all features, including contextual data like destination, weather, and time.2.Two hidden layers with 128 and 64 neurons, using ReLU activation to detect intricate patterns and interactions between features.3.A sigmoid output layer that produced a probability score for coupon acceptance, ideal for binary classification.To ensure the model generalized well, I incorporated dropout layers (20% rate) to prevent overfitting by disabling random neurons during training and used batch normalization to stabilize learning and improve convergence.Key Takeaways
1.Accuracy of approximately 70%, comparable to Naive Bayes but with enhanced adaptability.
2.Precision of around 68-70%, indicating relevant and focused recommendations.
3.Recall of 81%, making it highly effective at capturing cases of coupon acceptance—a critical feature for recommendation systems.

Imagine a platform that not only classifies text by topic but also recommends related content based on the specific themes a user has already explored. This project, "Recommend System: Text Analysis with SVM and Content-Based Recommendation," does just that. By combining the power of Support Vector Machines (SVM) for classification and a content-based recommendation system, we aim to create a system that both categorizes and recommends content effectively.

Building a Smart Text Classifier: Preprocessing, TF-IDF, and SVMFiltering relevant content from massive datasets can be a game-changer. To tackle this challenge, I set out to build a classifier capable of understanding and categorizing text from the 20 Newsgroups dataset, a classic collection of documents covering a diverse set of topics like technology, sports, religion, and more. Our goal? Achieve an accurate model that could categorize any given text into one of 20 categories.Preprocessing - Making the Text Ready for the ModelRaw text data is often noisy and unstructured, so we set out to clean and standardize it. Here’s how we prepared the documents:1.Lowercasing: To ensure consistency, I converted all text to lowercase.
2.Removing Punctuation and Special Characters:Next, I removed punctuation and symbols, focusing purely on the words themselves.
3.Tokenization and Stopword Removal: I split each document into individual words (tokens) and removed common words like "the," "is," and "in" using NLTK’s stopword list. This step reduced the dataset’s dimensionality, leaving behind only meaningful words that might actually help distinguish between categories.
4.Experimenting with Lemmatization
Initially, I thought lemmatization would help improve the model’s accuracy. By reducing "running" to "run" and "better" to "good," lemmatization can often make text more consistent and improve model performance. However, in this case, adding lemmatization didn’t yield significant improvements. This was likely because the dataset already contained specific terminology, and reducing words to their roots wasn’t as impactful as expected.5.Feature Extraction with TF-IDF
With the text cleaned and ready, I moved on to feature extraction. To represent text numerically, I used TF-IDF (Term Frequency-Inverse Document Frequency), which captures the importance of words in each document while down-weighting common words across all documents. I limited the vocabulary to the top 5,000 features to keep the model efficient while still capturing the essential information.

Adding a Personalized Touch: Content-Based RecommendationsAfter achieving a solid classification accuracy of 88% with the SVM model, I wanted to take this project a step further. Beyond just categorizing documents, I wanted to help users discover related content based on their interests. This led to building a content-based recommendation system that would suggest documents similar to a user’s input.Implementing the Recommendation SystemTo make recommendations, I designed a function that could take any text input and return the most relevant documents from the dataset. Here’s how it worked:Transforming the Input Text: First, I used the same TF-IDF vectorizer that had been trained on the document dataset to convert the input text into a numerical vector.
Calculating Similarity: With the input vector ready, I calculated the cosine similarity between the input and each document in the dataset. Cosine similarity helped identify documents with similar content by measuring the angle between their vectors—a smaller angle meant more similarity.
Retrieving the Most Relevant Documents: Once I had similarity scores, I sorted them in descending order and selected the top documents as recommendations. This approach allowed the system to recommend documents that were the closest match to the user’s input, whether they were looking for "Mac hardware issues" or "sports updates."
Presenting the Recommendations:
For each recommended document, I displayed the category and a snippet of the content to give users a quick preview. The goal was to provide a seamless experience where users could explore related topics without needing to specify exact keywords.
Results from Recommendation System

PROJECT 5
Understanding Customer Intent: A Journey Through Banking Conversations
INTRODUCTION

In the fast-paced world of banking, each customer query—whether it’s checking a balance, applying for a loan, or reporting a lost card—reveals a distinct intent. Yet, as the volume of customer interactions grows, deciphering these intents swiftly becomes challenging. To address this, I set out to create an intent classification model that could act as a bridge between customer needs and efficient service. By transforming vast amounts of unstructured text data into actionable insights, my model identifies patterns in language, categorizes intents, and ultimately enables banks to deliver more personalized and timely support. This journey through natural language processing in banking brings customer service one step closer to understanding and anticipating customer needs.

Lemmatization, or reducing words to their base forms, was skipped in this project. I realized that in banking, the specific forms of words carry unique meanings; “transferring” and “transfer,” for example, might signal slightly different intents. Retaining the original word forms allowed the model to recognize these nuances, enhancing its ability to accurately classify intents.Handling out-of-vocabulary (OOV) words was another critical consideration. To ensure the model wasn’t thrown off by unfamiliar terms, I introduced a placeholder token, <UNK>, to represent unknown words. This helped the model focus on known vocabulary while still capturing the essence of unfamiliar phrases through surrounding context. I also expanded contractions, such as turning “can’t” into “cannot,” which reduced the likelihood of missing out on critical intent signals hidden within shortened phrases.Finally, I created word embeddings using Word2Vec, employing the Continuous Bag of Words (CBOW) model to map each word into a dense vector. These embeddings enriched the model’s understanding of word relationships, especially in banking-specific terminology where context is key. To avoid potential biases from sequentially ordered data, I shuffled the batch data during training, ensuring the model learned underlying patterns rather than memorizing any order.

Building an Intent Classifier with LSTM: A Model Designed for Banking ConversationsTo tackle the task of understanding customer intents within banking queries, I needed a model architecture that could handle sequential data effectively. Given the context-heavy nature of language, especially in banking, I chose a Long Short-Term Memory (LSTM) network—a type of recurrent neural network well-suited for retaining contextual information over sequences.I started by stacking multiple LSTM layers, allowing the model to learn both simple and complex language patterns. This layered structure enabled the model to grasp everything from basic word associations to more nuanced meanings, crucial for accurately predicting customer intent. To prevent the model from overfitting, I incorporated dropout layers between the LSTM layers, randomly disabling some neurons during training. This approach helped ensure that the model could generalize well, rather than memorizing specific patterns in the training data.In addition, I used padding to make all input sequences the same length. This step was essential because customer queries vary in length; padding allowed the model to process them consistently, keeping its focus on content rather than structural inconsistencies.With the LSTM framework established, I proceeded with hyperparameter tuning. Using a random search strategy, I tested various configurations, fine-tuning parameters like the number of LSTM units, learning rate, batch size, and dropout rate. Adjusting these settings was like tuning an instrument, with each parameter contributing to a balance between learning speed, accuracy, and model robustness. I also optimized the embedding dimensions and window size in the Word2Vec model used for word embeddings. These embeddings captured word relationships, making the LSTM more effective at distinguishing intents based on contextual clues.

Experimenting with Bidirectional LSTM (BiLSTM)Finally, I experimented with a Bidirectional LSTM (BiLSTM) layer. Unlike traditional LSTMs, the BiLSTM processes information from both directions, allowing the model to understand words in the context of what comes before and after them. This bidirectional approach enhanced the model’s ability to interpret full intent more effectively, capturing subtle patterns across different intents within the banking domain. After rigorous training, this BiLSTM-based architecture, coupled with carefully tuned parameters, achieved the highest accuracy among the configurations tested, demonstrating its suitability for handling complex customer queries in banking.

Achieving Accuracy with Fine-Tuned Models: Results of Intent ClassificationThrough extensive experimentation and model tuning, the final BiLSTM model achieved impressive accuracy on the BANKING77 dataset. I tested various configurations and preprocessing strategies to find the optimal setup for accurately identifying customer intents in banking.After refining text handling and ensuring a balanced architecture, the final configuration with punctuation and stop words retained and a carefully adjusted BiLSTM achieved an accuracy of 85%. This result reflected the model’s ability to capture context effectively, distinguishing between nuanced banking queries with a high degree of precision.During training, both training and validation accuracy steadily increased over epochs, and loss values decreased consistently, showing that the model not only learned from the data but also generalized well to new examples. This final accuracy and low error rate confirmed the effectiveness of using BiLSTM architecture with tuned hyperparameters, creating a robust intent classification system that can handle the complex demands of customer interactions in banking.

Prediction of the Intent

Key Takeaways
Data-Driven Design with EDA:Exploratory data analysis revealed key patterns, such as the brevity of queries and frequently occurring bigrams. These insights informed the model's design, helping align it with typical customer behaviors in banking.Targeted Preprocessing: Customized text preprocessing steps, including the retention of punctuation and stop words, preserved essential context in banking queries. This approach allowed the model to interpret subtle nuances in customer language, which is crucial for accurate intent classification.Word Embeddings for Enhanced Understanding: Using Word2Vec embeddings provided dense vector representations of words, enriching the model’s understanding of banking-specific terms and relationships—valuable for recognizing customer intents.Effective BiLSTM Model: The bidirectional LSTM (BiLSTM) architecture enabled the model to understand each query in its full context, improving classification accuracy. Processing queries both forwards and backwards helped capture nuanced relationships in customer intents.Strong and Reliable Performance: The final model achieved 85% accuracy with minimal overfitting, as shown by closely aligned training and validation metrics. This performance validates the model's robustness and its suitability for real-world banking applications.

Hello! I’m Vyshali

CONTACT