AF302 Assignment: Deep Learning and Recommendation Systems Review

This literature review explores the advancements of deep learning in recommendation systems. It begins by defining deep learning and machine learning, highlighting their applications in analyzing data and identifying patterns for decision-making. The review emphasizes the importance of deep neural networks in representing complex user-item interactions, reducing the need for hand-crafted feature design, and enhancing the quality of recommendations. It then discusses the lifecycle of deep learning in recommendation, including training and inference phases, and how deep learning models leverage techniques like embeddings and collaborative filtering to improve accuracy. Various deep learning approaches, such as Wide and Deep Learning and Deep Collaborative Filtering, are examined, along with the use of explicit and implicit feedback. The review also references relevant research papers and journals, providing a comprehensive overview of the topic.

Literature review

LITERATURE REVIEW
According to the viewpoint of Le, (2019), Deep learning is the model that represent the
nonlinear interactions in the data with nonlinear activations. Machine learning is the method for
the Analysis of data that automates analytical model building. It is a suitable branch of artificial
intelligence is based on suitable Idea and system that can learn from the data identify pattern and
leads to make decisions with the minimal human invention. Machine learning is the pattern and
the theory that computer can learn without being programmed for the particular task and the
researchers are also interested in the artificial intelligence wanted to see the computers that can
learn from the data.
It leads to make things possible to capture Complex and Critical user item interaction
pattern. In regard of this the conventional method like Matrix factorization machines are
important in the linear models (Steiner and et. al., 2019). It is well developed neural network it is
able to approximate any continuous function with the arbitrary precision that vary according to
the choices and combinations. In relation with is the deep learning is considered as the efficient
learning underline the suitable factors and useful to presentations from the input data.
Making proper use of this information tends to provide an effective way to advance the
understanding of the users and items. It is considered as the natural choice in order to apply the
dip neural networks regarding the learning for the recommended models. The benefits of
considering deep neural network are helpful in representing the learning considering the
reduction of the efforts for the hand craft feature design and also enable recommendation model
(Razzak, Naz and Zaib, 2018).
Along with this considering the growth for the volume of data available to the power
recommender system helps in providing speed to the data that increasing turning from more
traditional machine learning methods to the highly expressive deep learning models in order to
enhance the quality of the recommendations.
Apart from this the life cycle of deep learning for recommendation is classified into two
phases including training and inference. Considering the training factor this model is trained in
order to analyse user-item interaction probability by presenting it with the examples among the
users and items from the past (Agarap, 2018). The deep learning recommender model helps in
developing the existing techniques like factorization to the model with interactions among
several embedding and variables in order to handle categorised variables.

An embedding is a learnt vector of integers that reflect entity attributes, allowing similar
entities (people or items) to be separated in the vector space by similar distances. A deep
learning approach to collaborative filtering, for example, learns user and item embeddings (latent
feature vectors) from their interactions with a neural network (Shrestha and Mahmood, 2019). To
train on massive volumes of data, harness the power of deep learning for feature extraction, and
develop more expressive models, DL approaches draw on the huge and rapidly increasing
innovative network topologies and optimization algorithms.
The user model, product model, and recommendation algorithm are the three main
components of a recommendation system. Image recognition and natural language processing
have seen significant gains because too deep learning. Deep learning techniques are being used
to recommender systems in new manner, thus increasing the use cases of neural networks. Some
of these cutting-edge systems are already delivering cutting-edge results and high-quality
recommendations (Nguyen and et. al., 2018). Deep learning methods, in comparison to
traditional models, can provide a deeper understanding of user’s desires, item qualities, and past
relationships between them.
Collaborative filtering, which is mathematically explained below, is a typical solution to the
recommendation problem (including regularisation designed to keep the weight space smaller).
Techniques like stochastic gradient descent can be used by data scientists to solve the objective
function. The same dynamics that have made data mining and Bayesian analysis more popular
than ever are driving renewed interest in machine learning. Things like increasing data volumes
and variety, cheaper and more powerful computing processing, and economical data storage.
Explicit feedback is direct and quantifiable data acquired from consumers in the context of
recommender systems. Amazon, for example, allows customers to score their purchases on a
scale of one to ten (Nguyen and et. al., 2018). Users offer these ratings directly to Amazon, and
the scale allows the company to quantify user preferences. Implicit feedback, on the other hand,
is gathered indirectly from user interactions and serves as a proxy for user preference. Implicit
feedback has the virtue of being plentiful.
Implicit feedback-based recommender systems also allow us to customise recommendations
in real time, with each click and interaction. Today’s online recommender systems incorporate
implicit input, which allows the system to fine-tune its recommendations in real time as users
interact with it. Deep learning may accomplish amazing things, but the applications are

sometimes veiled in research papers or demand computational capabilities that are only available
to major enterprises.
Nonetheless, there are deep learning applications that may be performed on a desktop
computer without the need for a graduate degree. Embeddings are a technique for converting
discrete, categorical variables into continuous vectors. In contrast to one-hot encoding, neural
network embeddings are low-dimensional and learnt, bringing related items closer together in the
embedding space (Sannino and De Pietro, 2018). Wide and Deep Learning is a useful
methodology for solving both regression and classification issues, which was first introduced in
Google Play for app recommendation.
A single layer perceptron, commonly known as a generalised linear model, is used in the
wide learning component. The recommender can capture both memorising and generalisation by
combining these two learning approaches. Because the dot product of matrix factorization does
not meet the triangle inequality of the distance function, Collaborative Metric Learning replaces
it with Euclidean distance. The distance between users and their hated items is maximised, and
the distance between users and their favourite items is minimised, to learn the user and item
embeddings (Park and et. al., 2018).
Deep Collaborative Filtering is a framework for combining deep learning and a
collaborative filtering paradigm. The framework makes it easy to develop hybrid collaborative
models using deep feature learning techniques. The paper Comparative Deep Learning of Hybrid
Representations for Picture Recommendations suggests a deep learning model for image
recommendation that is based on CNNs.

REFERENCES
Books and Journals
Agarap, A.F., 2018. Deep learning using rectified linear units (relu). arXiv preprint
arXiv:1803.08375.
Nguyen, H and et. al., 2018. Deep learning methods in transportation domain: a review. IET
Intelligent Transport Systems, 12(9), pp.998-1004.
Nguyen, K.K and et. al., 2018, April. Cyberattack detection in mobile cloud computing: A deep
learning approach. In 2018 IEEE Wireless Communications and Networking Conference
(WCNC) (pp. 1-6). IEEE.
Park, J and et. al., 2018. Deep learning inference in facebook data centers: Characterization,
performance optimizations and hardware implications. arXiv preprint
arXiv:1811.09886.
Razzak, M.I., Naz, S. and Zaib, A., 2018. Deep learning for medical image processing:
Overview, challenges and the future. Classification in BioApps, pp.323-350.
Sannino, G. and De Pietro, G., 2018. A deep learning approach for ECG-based heartbeat
classification for arrhythmia detection. Future Generation Computer Systems, 86,
pp.446-455.
Shrestha, A. and Mahmood, A., 2019. Review of deep learning algorithms and
architectures. IEEE Access, 7, pp.53040-53065.
Steiner, B and et. al., 2019. Pytorch: An imperative style, high-performance deep learning
library.
Online
Recommendation System Series Part 2: The 10 Categories of Deep Recommendation Systems
That Academic Researchers Should Pay Attention To. 2019. [Online] Available through:
< https://towardsdatascience.com/recommendation-system-series-part-2-the-10-
categories-of-deep-recommendation-systems-that-189d60287b58>./