AI CHAT BOT FOR eCOMMERCE
Introduction 1
Challenges 2
Proposed Solution 2
Brief Conclusion 2
Business Understanding: 3
Problem Statement: 3
Main objective. 3
Specific objectives 4
Business and Data Understanding 4
Data Source 4
Data Description. 4
Data Analysis 5
Modeling 5
Evaluation 5
Recommendations 5
Introduction
Artificial intelligence (AI) chatbots are apps or interfaces that can carry on human-like
conversation using natural language understanding (NLU) or natural language
processing (NLP) and machine learning (ML). They use large language models (LLMs)
to generate responses to text. This can be used as a virtual response assistant.
In the ever-evolving landscape of artificial intelligence, the integration of AI chatbots
has emerged as a transformative solution with wide-ranging applications. This project
centers around the development of versatile AI chatbots that serve as a responders. By
harnessing the power of natural language processing and machine learning, we aim to
create intelligent conversational agents capable of enhancing user experiences in
diverse domains.
Project overview
The AI Chatbot for Ecommerce Capstone project is designed to revolutionize
client-business interactions by introducing a sophisticated Frequently Asked
Questions (FAQ) system. Our primary goal is to offer seamless and efficient
assistance to both clients and business owners through the implementation of
advanced machine learning algorithms, including Naive Bayes and the Random
Forest Classifier. To augment the chatbot's capabilities, we plan to integrate
pretrained models that will ensure a comprehensive understanding of user
queries, enabling accurate and timely responses. The project will be executed in
several phases, starting with rigorous data preprocessing to ensure data quality
and consistency. We will then proceed to identify and extract relevant features
from integrated ecommerce data to enhance model training. Utilizing historical
data with known client-business interactions, we will train machine learning
models, with a focus on optimizing the system for scalability to accommodate a
growing user base. Emphasis will be placed on providing transparent and
understandable model outputs, facilitating effective decision-making.Our
success criteria include measuring the chatbot's accuracy and responsiveness,
evaluating scalability.
Challenges
eCommerce’s popularity has increased in recent years. It’s now fast becoming the go-to
shopping platform for many people. A report by Statista reveals that its global sales are
projected to reach a whopping $8.1 trillion by 2026. These staggering numbers show
eCommerce’s great potential. As your eCommerce business grows, you’ll receive more
customer inquiries via email, phone, and social media. It can be challenging to manage
these inquiries efficiently, especially if you’re a small business with limited resources.
For example, a customer who purchased a product from an eCommerce website has
questions about its specifications. He sends an email to the customer service team
asking for clarification. However, the inquiry was not responded to on time due to
mismanagement, as a result, the client felt neglected and unheard, leading the customer
to perceive the brand negatively.
Proposed Solution
Investing in a customer service chatbot that seamlessly integrates all communication
channels is a strategic move for any ecommerce store. By adopting such technology,
you streamline your customer support operations, offering a centralized solution for
efficiently managing client inquiries across various channels. This not only enhances
the customer experience but also boosts team productivity. With features like canned
responses and templates, customer enquiries can swiftly address common questions,
saving valuable time and ensuring consistent responses. Additionally, the platform,
ensures timely resolution and improved customer satisfaction. In essence, embracing
a chatbot-integrated customer service platform empowers your ecommerce store to
deliver exceptional support, build customer trust, and ultimately drive business growth.
Brief Conclusion
In conclusion, implementing a chatbot for customer support offers a transformative
solution for businesses, streamlining communication processes and enhancing
customer satisfaction. By leveraging artificial intelligence and natural language
processing capabilities, chatbots provide immediate assistance to customers,
addressing their inquiries and resolving issues efficiently around the clock. This not
only reduces response times but also enables businesses to handle a high volume of
queries simultaneously, scaling support operations effectively. Moreover, chatbots can
offer personalized recommendations, gather feedback, and analyze customer
interactions, allowing businesses to continuously improve their services. Overall,
integrating chatbots into customer support strategies empowers businesses to deliver
seamless, proactive, and tailored assistance, fostering stronger customer relationships
and driving long-term success.
Business Understanding:
In the dynamic realm of artificial intelligence, the utilization of AI chatbots has become
increasingly integral, presenting numerous opportunities for transformative
applications. This project focuses on the development of versatile AI chatbots
designed to serve as customer support. Leveraging advanced natural language
processing and machine learning techniques, the goal is to craft intelligent
conversational agents capable of elevating user experiences across a spectrum of
domains.
Problem Statement:
In today's rapidly advancing digital landscape, the integration of chatbots has become
a pervasive trend, offering enhanced efficiency and convenience. In the dynamic world,
chatbots help do the day-to-day running of customer support.The current landscape of
chatbots often falls short in adaptability, lacking the versatility required for diverse
roles. The challenge at hand is to address this limitation by creating an AI chatbot
framework that can effortlessly resolve customer concerns, customer queries and
have meaningful interactions with the end user. Time is a precious commodity and
having enough in a day to do the things that matter is important. Time to answer all
business messages can be overwhelming, leading to a disconnect between the
business and the customer . As a result, they may feel unheard and miss out on
valuable interactions. Thus,enhancing the engagement experience for customers,
through AI-powered chatbot designed to provide responsive, context-aware assistance
without the limitations of human-operated systems.
Main objective.
To engineer a dynamic and user-centric AI chatbot system for businesses and
corporate intranets. This system should not only grasp context effectively but also
adapt to user preferences, providing a comprehensive solution for varied scenarios.
By pushing the boundaries of virtual interactions, the project aims to redefine user
experiences, offering innovative AI-driven chatbot functionalities.
Specific objectives
1.To enhance customer engagement.
2.To increase operational efficiency by avoiding bottlenecks when it comes to
response time (Wu et al., 2020).
3. Sales Support in ecommerce by providing product information to potential
customers.
Business and Data Understanding
Data Source
Amazon question/answer data https://cseweb.ucsd.edu/~jmcauley/datasets/amazon/qa/
Reference
Wu, Q., Ma, J., & Wu, Z. (2020, April). Consumer-driven e-commerce: A study on C2B
applications. In 2020 International Conference on E-Commerce and Internet Technology (ECIT)
(pp. 50-53). IEEE. DOI: 10.1109/ECIT50008.2020.00019
Data Description.
This dataset is stored in the "QA_Beauty.json.gz". It contains Question and Answer
data comprising of 32936 rows and 6 columns.Here is a brief description of each
column:
1. asin: This column represents the ID of the product, serving as a unique identifier,
such as "B000050B6Z".
2. questionType: It denotes the type of question and can take on values like 'yes/no' or
'open-ended', indicating the nature of the inquiry.
3. askerID: This column contains the ID associated with the individual asking the
question, providing a reference to the person initiating the inquiry.
4. questionText: It holds the text of the question, providing insight into the specific
queries users have about the products.
5. answers: This column encapsulates the responses to the questions posed. It could
contain a variety of information, including yes/no answers or more detailed responses
for open-ended questions.
6. questionTime: This column contains timestamp information representing the date
and time when each question was asked.
Data Analysis
The graph presented is a univariate bar plot that displays the distribution of question types within a
dataset. This dataset appears to be related to an e-commerce setting question/answer data.
From the bar plot, we can observe two categories of question types:
1. Open-ended: This category has the largest number of occurrences, indicating that it is the most
common type of question in the dataset. The bar representing open-ended questions is significantly
higher than the one for yes/no questions, suggesting that users prefer asking questions that require a
more detailed response rather than just a simple affirmative or negative. This could mean that
customers are looking f~~or more in-depth information about products or have queries that cannot be
answered with a simple yes or no.
2. Yes/no: This category has a much lower count in comparison to open-ended questions, indicating
that such questions are less frequent in the dataset. Yes/no questions are likely to be those that
require a straightforward answer without the need for elaboration.
The 'open-ended' question type exhibits the highest number of occurrences in the dataset, suggesting
that users frequently engage in inquiries that prompt detailed and unrestricted responses. This
prevalence highlights the users' inclination towards seeking comprehensive information or
explanations.
In a customer service chatbot context, this data could be used to prioritize the development of
features that support open-ended questions, ensuring that the bot is capable of handling detailed and
complex user queries effectively.
The histogram shows that the majority of questions receive a specific number of answers, with the
number two being the most common, as indicated by the highest bar on the graph.
With the added context in the caption, it suggests that the dataset contains a significant number of
questions that receive exactly two answers, pointing towards a pattern of binary or dual-response
format. This could imply that many users are satisfied with two answers to their questions, or that the
system or process that generates these answers tends to provide two by default or as a limit.
The prevalence of two answers could be due to several reasons, such as:
● Users often pose questions that can be sufficiently answered with a couple of responses, or
they tend to prefer a second opinion.
● The platform may encourage or limit users to provide two answers, which could be a design
choice to prevent information overload or to simplify the decision-making process for readers.
● The nature of the questions could be such that they naturally elicit a binary response,
representing a comparison or a choice between two options.
Understanding this pattern is crucial for the design and implementation of an AI chatbot for customer
service in e-commerce.
The visualization displays the top 20 most frequently occurring words found in questions from the
dataset. These types of words are typically used in text analysis to understand the common language
or topics that are being discussed.
Here are some key observations:
● "the" and "this" are the most common words, with "the" being the most frequent. These are
articles in English and are commonly used in many types of sentences, which explains their
prevalence.
● The list is dominated by function words, such as articles ("the", "a"), pronouns ("this", "it"),
conjunctions ("and"), prepositions ("to", "for", "of", "on"), and auxiliary verbs ("is", "does"). These
words are often used to form questions and are typically frequent in most English language
corpora.
● There are also some content words like "use", which may be relevant to the e-commerce
context of the dataset, potentially indicating that questions about how to use products are
common.
● The presence of words like "have", "does", and "use" towards the end of the list suggests that
questions in the dataset may involve ownership ("have"), inquiries about functions or features
("does... have"), and usage ("use").
This chart is useful for understanding the structure of questions that customers are asking.
Most Common Bi-grams: The left chart shows the frequency of bi-grams:
● "is this" is the most frequent bi-gram, suggesting that many questions start with an
inquiry directly about a product or feature.
● Following "is this", the next most common bi-grams include "is the", "does this", "have
it", "in the", and "of the". These bi-grams reflect common question patterns, where users
are asking about specific attributes of a product, its availability, or seeking to
understand its placement within a certain context.
● The least common bi-gram among the top shown is "is it", which could be part of
questions about product confirmation or verification.
Most Common Tri-grams: The right chart displays the frequency of tri-grams:
● "what is the" is the most frequent tri-gram, which is often a lead into a detailed question
about a product or service.
● Following this, the other common tri-grams include "is this the", "is the this", and "what
is this". These suggest that users are looking for specific information or verification
about products.
● The least common tri-gram shown is "or is it", which might be used in comparative or
alternative inquiries about products or features.
Notable Words: The caption also mentions notable words such as "product", "use", "work",
"will", "Thanks", and "One". These content-specific words can provide insights into common
themes in customer queries, such as questions about how to use a product ("use"), its
functionality ("work"), future availability or action ("will"), gratitude or follow-up ("Thanks"), or
selecting an option ("One").
Implications for AI Chatbot Development: This n-gram analysis is very useful for designing an
AI chatbot for e-commerce customer service:
● Understanding common bi-grams and tri-grams can help in training the chatbot to
recognize frequent patterns in customer inquiries.
● The chatbot can be programmed to trigger certain responses based on these common
n-grams, enhancing its ability to address the most frequent types of questions.
● Content words like "product", "use", "work", "will", "Thanks", and "One" might be critical in
understanding the intent behind a customer's question and providing accurate
responses.
Thus, the n-gram analysis helps in understanding the structure and subject of customer
queries, which is crucial for building an effective AI chatbot that can provide relevant and
helpful answers.
In this word cloud, we can identify several prominent words:
● "product": Its large size suggests that it is one of the most frequently mentioned words,
indicating that many questions are directly related to specific products.
● "use": Also quite large, implying that questions about how to use items or the uses of different
products are common.
● "work": Likely related to questions about the functionality or effectiveness of products.
● "will": This could be part of questions about the future effects of using a product or the
product's longevity.
● "Thanks": This appears to be a common sign-off in questions, showing politeness from the
askers.
● "One": Might be used in the context of selecting a product or choosing between options.
Other noticeable words include "size", "color", "ingredient", "buy", "face", "skin", "brush", "cream",
"shampoo", "makeup", "long", "need", and "help". These words suggest that the questions are related to
personal care products and that customers are concerned with product attributes (like size and color),
ingredients, purchase decisions, and usage instructions.
From an AI chatbot development perspective, the word cloud informs the areas where the chatbot
should have robust knowledge. For instance, the chatbot should be able to answer questions about
product specifications ("size", "color"), ingredients, how to use products ("use", "apply"), and possibly
the efficacy ("work") of products. It also suggests that customers appreciate courteous interaction
("Thanks"), which should be part of the chatbot's response templates.
Modeling
Multinomial Naive Bayes Model
Accuracy: 0.06140983486328244
The MultinomialNB model makes an independence assumption and suits discrete data.
Evaluation
The low accuracy indicates it did not generalize well. A look into the classification report confirms the
low scores across all the intent classes. The macro average precision is 0.05, recall is 0.06 and F1 is
0.06, highlighting that the model struggled to correctly classify examples from the minority classes.
Most classes have precision and recall scores in the 0.02 to 0.07 range, with just the "Personal Care"
intent slightly higher at 0.12 precision and 0.20 recall. Such large class imbalances showcase the
model's inability to learn effectively from skewed datasets. The model's average cross-validation
accuracy over 5 folds is 0.07, which aligns with the low test accuracy. This confirms the model's
subpar performance irrespective of train-test split.
Random Forest
Random Forest Test Accuracy: 0.8800475059382423
In Random Forest classifier is trained and evaluated after processing the data, with the model focusing on
the features 'questionText' and 'answerText' to predict the tag 'questionType'. The dataset is split into
training and testing sets, and the classifier is created with 100 trees. The accuracy is then calculated using
the test set, resulting in a high accuracy of 88.004%. This indicates that the model, trained specifically on
the textual content of questions and answers, performs well in classifying instances into different question
types. It's important to note that the success of the model relies on the features chosen for training, and the
achieved accuracy provides insight into the effectiveness of the classifier in this particular context.
Nonetheless, considering other evaluation metrics and potential hyperparameter tuning might further
optimize the model for the specific characteristics of the classification problem.
Sequential Neural Network Model
Neural Network Test Accuracy: 0.8640924096107483
Sequential neural network model is constructed using the TensorFlow Keras library for a
'questionType' classification task based on textual data from 'questionText' and 'answerText.' The text
is preprocessed by tokenizing and lemmatizing using the NLTK library. The lemmatized text is then
converted into bags of words, and the output rows are one-hot encoded. The Count Vectorization
technique is employed to convert the text data into numerical format, utilizing a reduced vocabulary
size. The resulting numerical data is split into features (X) and labels (Y), followed by a further division
into training and testing sets. The neural network model is designed with an input layer of 128 neurons,
a dropout layer to mitigate overfitting, a hidden layer with 64 neurons, another dropout layer, and an
output layer with softmax activation for multiclass classification. The model is compiled with
categorical crossentropy loss, the Adam optimizer, and accuracy as the metric. Subsequently, the
model is trained over five epochs, with a batch size of 32 and a validation split of 20%. The trained
model is then evaluated on the test set, yielding a neural network test accuracy of approximately
86.41%. This accuracy metric underscores the proficiency of the neural network in effectively
classifying 'questionType' based on the provided textual features.
Recommendations
1. Optimize for Scalability: As the eCommerce platform grows, the volume of customer inquiries
will increase. It's vital to ensure that the chatbot system is scalable, capable of handling a large
number of queries without compromising response time or accuracy.
2. Personalization: Implement features that allow the chatbot to offer personalized
recommendations and responses based on the user's browsing and purchase history. This can
enhance the customer experience and potentially increase sales conversions.
3. Feedback Loop Integration: Establish a mechanism for collecting user feedback on chatbot
interactions. This data can be invaluable for continuous improvement of the chatbot's
responses and functionalities.
4. Multilingual Support: Considering the global reach of eCommerce, adding multilingual support
to the chatbot can make it accessible to a wider audience, thereby enhancing customer
support and inclusivity.
Github Repository: https://github.com/MwangiWambugu/AI-chatbot-for-eCommerce-Store
