Use Case: Human Activity Recognition using Smartphone Dataset

Team

1. N.RAVINDRA BABU, VTU19868, vtu19868@veltech.edu.in

2. G.LAKSHMANUDU, VTU19924, vtu19924@veltech.edu.in
3. R.VIJAY KUMAR, VTU19972, vtu19972@veltech.edu.in
4. C.VENKATA BAVESH SAI, VTU19977, vtu19977@veltech.edu.in

Abstract:

In modern lifestyles, monitoring fitness activities has become essential for promoting healthier
living. This project proposes the development of a precision activity recognition system using
smartphone sensor data to accurately identify human fitness activities. The project aims to
address the multi- classification challenge inherent in distinguishing various fitness activities
such as walking, running, cycling, and more. The project workflow involves several key stages.
Firstly, a comprehensive dataset comprising accelerometer and gyroscope readings collected
from smartphones during diverse fitness activities is obtained and preprocessed. Data
preprocessing involves cleaning, normalization, and feature extraction to capture relevant
patterns and characteristics. Next, machine learning algorithms such as Support Vector
Machines (SVM), Random Forest, and conventional Neural Networks are employed to build
classification models. These models are trained on the preprocessed dataset to learn the
complex relationships between sensor data and fitness activities. To enhance model
performance, hyperparameter tuning and cross-validation techniques are applied to optimize
model parameters and mitigate overfitting. Additionally, ensemble learning methods may be
explored to combine the strengths of multiple models for improved accuracy and robustness.
The performance of the developed classification models is rigorously evaluated using standard
metrics such as accuracy, precision, recall, and F1-score. Moreover, confusion matrices and
ROC curves are utilized to assess the models' ability to correctly classify different fitness
activities.The project aims to deliver a scalable and efficient activity recognition system that
can seamlessly integrate with smartphones for realtime fitness monitoring. The system's
potential applications extend beyond individual fitness tracking to include personalized
coaching, health assessments, and activity- based recommendations. In conclusion, this project
offers valuable insights into solving multiclassification problems in the context of fitness
activity recognition using smartphone sensor data. These models are trained on the
preprocessed dataset to learn the complex relationships between sensor data and fitness
activities.To enhance model performance, hyperparameter tuning and cross-validation
techniques are applied to optimize model parameters and mitigate overfitting. Additionally,
ensemble learning methods may be explored to combine the strengths of multiple models for
improved accuracy and robustness. The performance of the developed classification models is
rigorously evaluated using standard metrics such as accuracy, precision, recall, and F1-score.
Moreover, confusion matrices and ROC curves are utilized to assess the models' ability to
correctly classify different fitness activities

Keywords: Machine Learning, Support Vector Machine, Cross Validation, Data Preprocessing
Classification Model
1. Introduction:

In today's digital age, the vast ocean of available content can often overwhelm users, leaving
them in search of tailored recommendations to navigate through the abundance of choices. One
such domain where personalized recommendation systems are paramount is the movie
industry. With millions of movies spanning various genres, directors, and actors, users often
find themselves seeking guidance to discover content aligned with their preferences. In
response to this need, our project focuses on the design and implementation of a movie
recommender system leveraging the Movie Lens dataset, an invaluable resource in the realm
of recommendation systems.

The MovieLens dataset serves as a cornerstone for research and development in the field of
recommendation systems, providing a comprehensive collection of movie ratings contributed
by users alongside detailed movie metadata. This dataset encompasses a diverse range of user
preferences and interactions, making it an ideal foundation for training machine learning
models to generate personalized movie recommendations.

At the heart of our project lies the application of machine learning techniques, specifically
collaborative filtering methods, to analyze user-item interactions and derive meaningful
recommendations. Collaborative filtering leverages the collective wisdom of user ratings to
identify similarities among users and recommend items based on the preferences of similar
users. Within collaborative filtering, two primary approaches are employed: user-based and
item-based. User-based collaborative filtering identifies similar users based on their past
interactions and recommends items favored by those similar users. Conversely, item-based
collaborative filtering identifies similar items based on user ratings and recommends items
similar to those already liked by the user.

In addition to collaborative filtering, we explore the utilization of matrix factorization

techniques such as Singular Value Decomposition (SVD) and Alternating Least Squares (ALS).
These methods aim to decompose the user-item interaction matrix into lower-dimensional
matrices representing latent factors. By capturing the underlying patterns and relationships
between users and items, matrix factorization techniques enable us to generate accurate
recommendations even in the presence of sparse data.

Furthermore, our project extends beyond traditional collaborative filtering approaches to

incorporate content-based filtering methods. Content-based filtering leverages movie attributes
such as genre, director, and cast to recommend items similar to those previously liked by the
user. By combining collaborative and content-based filtering techniques, our recommender
system aims to provide users with diverse and accurate movie recommendations tailored to
their unique preferences.

The efficacy of the recommender system is evaluated using established machine learning
metrics such as precision, recall, and Mean Absolute Error (MAE). These metrics enable us to
assess the accuracy and effectiveness of the recommendation algorithms, guiding us in fine-
tuning and optimizing the system for enhanced performance.

Ultimately, the culmination of this project will be the deployment of a user-friendly interface
that seamlessly delivers personalized movie recommendations to users. By leveraging the
power of machine learning and the rich insights provided by the MovieLens dataset, our
recommender system aims to enrich the movie-watching experiences of users and facilitate
content discovery in an era characterized by an abundance of digital media.
2. Related Works:

[1] Koren, Bell, and Volinsky (2009) et.al., delve into matrix factorization techniques within
recommender systems. By decomposing user-item interaction matrices, latent factors are
unveiled, leading to improved recommendation accuracy. This paper is crucial for
understanding the foundational principles behind matrix factorization and its application in
enhancing recommendation systems.

[2] Resnick and Varian (1997) present a seminal overview of recommender systems,
emphasizing their significance in aiding users' content discovery. The paper discusses
collaborative filtering and content-based filtering, providing a broad understanding of the
different approaches employed in recommendation systems.

[3] Sarwar et al. (2001) introduce item-based collaborative filtering algorithms, which
recommend items based on their similarity. By addressing the "cold start" problem, these
algorithms offer practical solutions for generating recommendations even for new or less-rated
items.

[4] Breese, Heckerman, and Kadie (1998) et.al., conduct an empirical analysis of predictive
algorithms in collaborative filtering. Their study provides valuable insights into the
performance and effectiveness of various collaborative filtering approaches, offering guidance
for implementing recommendation systems.

[5] Bell and Koren (2007) reflect on the lessons learned from the Netflix Prize challenge,
highlighting the efficacy of collaborative filtering techniques such as matrix factorization. This
paper offers practical insights into improving recommendation systems based on real-world
competition experiences.

[6] Goldberg (1992) et.al., introduce collaborative filtering as a method for generating
personalized recommendations. By leveraging the preferences of similar users, collaborative
filtering weaves an "information tapestry," facilitating content discovery for users.

[7] Ding, Li, and Jordan (2006) propose a method for solving non-negative matrix factorization
(NMF) problems using alternating non-negativity-constrained least squares. NMF is a vital
technique in recommendation systems, and this paper offers a valuable contribution to its
optimization and implementation.

[8] Desrosiers and Karypis (2011) et.al., provide a comprehensive survey of neighborhood-
based recommendation methods, detailing various algorithms based on user-user or item-item
similarity. This survey offers a thorough understanding of neighborhood-based approaches and
their applications in recommendation systems.

[9] Park, S. T., & Chu, W. (2009) et.al., Pairwise preference regression for cold-start
recommendation. In Proceedings of the third ACM conference on Recommender systems (pp.
21-28).Park and Chu introduce pairwise preference regression, a technique addressing the cold-
start problem in recommendation systems. By leveraging pairwise preferences, this method
effectively recommends items even for new users or items with sparse data.

[10] Hu, Y., Koren, Y., & Volinsky, C. (2008) et.al., collaborative filtering for implicit feedback
datasets. In 2008 Eighth IEEE International Conference on Data Mining (pp. 263-272) Hu,
Koren, and Volinsky present collaborative filtering techniques tailored for implicit feedback
datasets, where user interactions are inferred rather than explicitly provided.
3. Proposed Methodology:

Dataset : This Dataset is Movie Lens dataset which Contains movie ratings, metadata, user
IDs, timestamps which is Used for recommender system research and algorithm development.

Preprocessing: Before preprocessing, the Movie Lens dataset typically consists of raw user
ratings and movie metadata. This raw data may contain inconsistencies, missing values, or
noise that could affect the performance of recommendation algorithms.

After preprocessing, the dataset undergoes cleaning and transformation steps to address these
issues. This includes handling missing values, removing duplicates, standardizing formats, and
encoding categorical variables. Preprocessing may also involve feature engineering to extract
additional information or enhance the dataset's quality.

Sample Before Preprocessing:

- Raw user ratings:
- User 1 rates "Movie A" as 5 stars.
- User 2 rates "Movie B" as 3 stars.
- Raw movie metadata:
- "Movie A" has genres Comedy and Drama.
- "Movie B" has genres Action and Adventure.

Sample After Preprocessing:

- Cleaned user ratings:
- User 1 rates "Movie A" as 5 stars.
- User 2 rates "Movie B" as 3 stars.
- Cleaned movie metadata:
- "Movie A" is labeled with genres Comedy and Drama.
- "Movie B" is labeled with genres Action and Adventure.

4. Results and Discussion:

5. Conclusion:

In this project, we worked on a K-Means clustering algorithm , knn algorithm and Affinity
Propagation Clustering Algorithm to make movie recommendations as good as possible. User
rating and preference have been considered while building the system. Our system will
calculate the common in predicting rating from user information which may be used to analyze
that movie ought to suggest to new users using three Machine Learning Algorithm. This proves
that our system may be a valid one for predicting within the field of movies. Finally, we
compare the result of the three algorithm and found that we got nearly same result with different
execution time. So we conclude that Affinity Propagation Clustering Algorithm is more
efficient than the other two algorithm
References:

[1] Koren, Bell, and Volinsky (2009) et.al., delve into matrix factorization techniques within
recommender systems. By decomposing user-item interaction matrices, latent factors are
unveiled, leading to improved recommendation accuracy. This paper is crucial for
understanding the foundational principles behind matrix factorization and its application in
enhancing recommendation systems.

[2] Resnick and Varian (1997) present a seminal overview of recommender systems,
emphasizing their significance in aiding users' content discovery. The paper discusses
collaborative filtering and content-based filtering, providing a broad understanding of the
different approaches employed in recommendation systems.

[3] Sarwar et al. (2001) introduce item-based collaborative filtering algorithms, which
recommend items based on their similarity. By addressing the "cold start" problem, these
algorithms offer practical solutions for generating recommendations even for new or less-rated
items.

[4] Breese, Heckerman, and Kadie (1998) et.al., conduct an empirical analysis of predictive
algorithms in collaborative filtering. Their study provides valuable insights into the
performance and effectiveness of various collaborative filtering approaches, offering guidance
for implementing recommendation systems.

[5] Bell and Koren (2007) reflect on the lessons learned from the Netflix Prize challenge,
highlighting the efficacy of collaborative filtering techniques such as matrix factorization. This
paper offers practical insights into improving recommendation systems based on real-world
competition experiences.

[6] Goldberg (1992) et.al., introduce collaborative filtering as a method for generating
personalized recommendations. By leveraging the preferences of similar users, collaborative
filtering weaves an "information tapestry," facilitating content discovery for users.

[7] Ding, Li, and Jordan (2006) propose a method for solving non-negative matrix factorization
(NMF) problems using alternating non-negativity-constrained least squares. NMF is a vital
technique in recommendation systems, and this paper offers a valuable contribution to its
optimization and implementation.

[8] Desrosiers and Karypis (2011) et.al., provide a comprehensive survey of neighborhood-
based recommendation methods, detailing various algorithms based on user-user or item-item
similarity. This survey offers a thorough understanding of neighborhood-based approaches and
their applications in recommendation systems.

[9] Park, S. T., & Chu, W. (2009) et.al., Pairwise preference regression for cold-start
recommendation. In Proceedings of the third ACM conference on Recommender systems (pp.
21-28).Park and Chu introduce pairwise preference regression, a technique addressing the cold-
start problem in recommendation systems. By leveraging pairwise preferences, this method
effectively recommends items even for new users or items with sparse data.