Overview On

Movie Recommender System

Jones G, Zahra Fathima, Prof. Devi Sivasankari, Prof. Naseela Jehan,

Zaiba Afreen, Uzma Farheen, Department Of ISE,
Department Of ISE, HKBK College Of Engineering,
HKBK College Of Engineering, Visveswaraya Technological University,
Visveswaraya Technological University, Bangalore, India
Bangalore, India

Abstract-- A novel unified framework for deep learning- II. LITERATURE SURVEY
based movie recommender systems, DeepMovRS, is
proposed. The framework proposed accepts different
various heterogeneous inputs from user and movie entities, Broadly speaking, the general properties of deep neural
and their knowledge to external and implicit feedbacks. In networks are solutions for the traditional methods’ weaknesses
order to ensure the unified deep architecture of the and the requirements of recommenders. For example, the classic
framework, so that it is easier for retrieving and ranking methods can’t handle the massive datasets efficiently, whereas
movies, it uses suitable machine learning tools to improve the more training data we have, the more accurate deep model
the quality of recommendations. The proposed framework we get.
has an additional feature which is flexible and modular,
and it can be generalized and distributed easily, and hence 1) Classic methods weaknesses: In the age of Big Data,
it turns out to be a rational choice for the recommendation classic algorithms like [6], can’t handle massive datasets
of movies for movie recommender systems. And this can efficiently. In the recommender systems in which data
further be extended for other entities. are sparse, maybe new set of algorithms can handle
sparsity better. Usually the notorious cold-start problem
of recommenders systems solved by leveraging auxiliary
I. INTRODUCTION
informations, so is there a new set of algorithms by
which this problem become less severe?
In the movie recommendation it the system application
has two entities: users and items. This paper focuses on the
movie recommender systems which are the core usage 2) Requirements: In order to have a high quality results,
functionalities of websites and e-commerce applications, i.e. long-term static and short-term temporal user prefer-
items=movies. In order to overcome the drawbacks, such as ences must be modeled. The diversity vs. the accuracy of
scalability, sparsity and cold-start problems. Although this recommendation is a hard trade-off that hasn’t an
framework is intended for movie recommender systems, it can obvious solution. Distributed recommender systems that
be easily extended to other domains such as hospital operate in open networks and recommenders designed to
recommendation system. In such movie recommender systems, operate in mobile devices and usage contexts, are
users have preferences for certain items, and these preferences required. Accurate recommendations even for cold start
must be obtained from the data [8]. And the one main difficulty ones, is a goal of these systems.
is in focal point of designing features (e.g. genre in the movie
recommenders) especially for a huge amount of items manually, 3) Deep learning models strengths:
is intractable. In such issues, the concept of machine learning a) Have dramatically improved the state-of-the-art in
plays an important role. And as obvious as it is in Artificial speech recognition, visual object recognition, ob-
Intelligence, Deep Learning, which in the recent emerging of ject detection, drug discovery and genomics [4], [7].
machine learning, there is an approach mainly for b) Can generalize to previously unseen query-item
recommender systems. In this paper, we propose a novel feature pairs.
unified framework which has certain advantages in contrast c) No or Less burden of feature engineering.
with the current frameworks. This has future evolved the
recommendation system, and in this case a movie d) Arbitrary continuous and categorical features can be
recommendation system. easily added to the model.
e) Layers of depth were shown to effectively model
In the above section i.e., Introduction we spoke about non-linear interactions between hundreds of fea-
our proposal and the main notion of the implementation of tures.
this paper and its proposal. In the further sections we would f) Non-linear mapping leads to compact representa-
like to give a brief about Literature Survey, i.e., Section II, tion, compact representation capture lots of seman-
and the main feature or key area is Implementation which tics.
would be g) Are able to discover non-linear latent variables with
discussed in the Section III, and the Conclusion would be heterogeneous data.
finally discussed in the Section IV .
 III.IMPLEMENTATION Many branches and a lot of research groups have been worked

→ embedding of words
on word embeddings. For instance, a team at Google created
word2vec, a word embedding toolkit which can train vector space
In natural language processing (NLP),
models faster than the previous approaches [9]. Recently, word
are a set of feature learning and language modeling techniques
embedding algorithms have relied on neural network architecture
where c e r t a i n words or phrases from the vocabulary are
instead of some traditional n-gram mod- els and unsupervised
mapped to or are assigned to vectors of real numbers ( : V
ℝ ). In other words, it is a imbibing of mathematical
learning.
It should be mentioned that, software for training and utilizing
embedding in the space from one dimension for a word to a word embeddings includes Word2vec, Stanford Uni- versity’s
spontaneous vector space with a dimension that is lower. GloVe, Gensim, Indra and Deeplearning4j. Princi- pal
Methods to generate this mapping include neural networks, Component Analysis (PCA) and T-Distributed Stochastic
dimension reduction on the word co-occurrence matrix, Neighbour Embedding (t-SNE) are both used to reduce the
probabilistic models, and explicit representation in terms of the dimensionality of word vector spaces and visualize word
context in which words appear [4], [12]. embeddings and clusters. For more on embedding, see, e.g. [4].
In linguistics word and its embedding were studied in the
research area of distributional semantics, it aims to quantify and BIAS/VARIANCE: METHODS OF EVALUATION
categorize semantic similarities between linguistic items based The machine learning goal of solving a task not only on the
on their distributional properties in large samples of language training set but also to generalize, can be achieved using many
data.

TABLE I
and deep neural networks combines the benefits of
NOTATIONS memorization and generalization for recommender systems.
Symbols Definitions and Descriptions
ℱ� ∣ �� User � ∣ Item � features In this section, a novel unified framework which is called
�� ∣ �� User � ∣ Item � communities
DeepMovRS is proposed. Based on the above discussion,
� ∣ � User � ∣ Item � reviews
V All reviews (= movies). This lead to a good representation of users and users
R Explicit feedback
H Implicit feedback matrix
are mapped to a common latent space, as well as items items.
Φ� ∣ Φ � User � ∣ Item � representation Then, this latent representation are used for retrieving
f�∣g�, c�∣d�, j�∣s�, p�∣d� User � ∣ Item � latent factors and ranking items. On the other hand, the framework consist of
two deep neural networks, one for generating top- candidate
subset of videos from the video corpus, and the other one for
tools provided by the field of statistics. Foundational concepts ranking results from the candidate set generated by the former.
such as bias and variance are useful to formally characterize While the common frameworks were different and trained
notions of generalization, underfitting and overfitting [1], [2], candidate generation and ranking networks separately, in our
[4]. proposed framework, we unify them, as well as in- formation
∙ Variance: Variance is corresponding to the precision of source processing modules. The recommendation problem can
the model. Low variance leads to a good memorization be viewed as a classification problem or the final objective
and high variance leads to the overfitting problem. Wide function can be a weighted sum of each module’s objective
linear models can effectively memorize sparse feature function. The overall framework will be trained via
interactions using cross-product feature transformations. backpropagation. Consequently, the users and the videos will
∙ Bias: Bias is corresponding to the accuracy of the model. map to user and video latent space respectively. In other words,
Low bias leads to a good generalization and high bias the framework learn an effective representation of the users and
leads to the underfitting problem. Deep neural networks the videos. DeepMovRS for some of information sources
can generalize to previously unseen feature interactions is presented in Figure 2. The mathematical notations used in this
through low-dimensional embeddings. presentation are summarized in Table I.
Novelty Our framework is different from some of the current
As it is shown in Figure 1, jointly training wide linear models
frameworks like [10] and [11]. Especially, we train the candidate
generation and the ranking modules together with the presented in [5].
embedding neural networks jointly, while in [10] and [11] they
don’t consider the candidate generation and the ranking Advantages The advantages of the proposed framework can be
modules in training. Furthermore some current frameworks can summarized as follows:
be viewed as a special case of our framework. For instance, if ∙ Arbitrary continuous and categorical features can be easily
the information sources be limited to the implicit feedbacks, added to the model.
then our framework will be reduced to the framework ∙ It combines and uses rich heterogeneous auxiliary info.

Fig. 2. The proposed framework: unified deep learning-based movie recommender system

representation space.

It is modular and easily can be distributed. Each module get an

information source and map it to a latent space. In order to
improve the quality of recommendations, anyone can design a
new module and append the result of this new module to the
 IV. CONCLUSION

A novel unified deep learning-based framework for movie

recommender systems was proposed. The main contributions of
the paper include:
∙ DeepMovRS unifies the candidate generation, ranking,
community detection, matrix factorization, review miner
modules.
∙ DeepMovRS leverages a broad spectrum of information
resources.
The proposed framework is generic and with a few
manipulations, can be used for different scenarios other
than movie recommendation (e.g. items= applications, books
or tags). Adding a new information source to the framework is
straightforward. For example, one can sample some number of
frames from each movie, feed them to a convolutional neural
network and add the final feature map to the representation
vector of the movie.

REFERENCES

[1] C. Bishop, Pattern Recognition and Machine Learning, Springer-

Verlag New York, (2006).

[2] H.-T. Cheng, L. Koc, J. Harmsen, T. Shaked, T. Chandra, H.

Aradhye, G. Anderson, G. Corrado, W. Chai, M. Ispir, R. Anil, Z.
Haque, L. Hong, V. Jain, X. Liu and H. Shah, Wide & Deep Learning
for Recommender Systems, In Proceedings of the 1st Workshop on
Deep Learning for Recommender Systems (DLRS 2016), (2016). ACM,
New York, NY, USA, 7-10.

[3] P. Covington, J. Adams and E. Sargin, Deep Neural Networks for

YouTube Recommendations, In Proceedings of the 10th ACM
Conference on Recommender Systems, (2016). ACM, New York, NY,
USA, 191-198.

[4] I. Goodfellow, Y. Bengio and A. Courville, Deep Learning, MIT

Press, http://www.deeplearningbook.org, (2016).

[5] X. He, L. Liao, H. Zhang, L. Nie, X. Hu and T.-S. Chua, Neural

Collaborative Filtering, In Proceedings of the 26th International
Conference on World Wide Web (WWW ’17), (2017). International
World Wide Web Conferences Steering Committee, Republic and
Canton of Geneva, Switzerland, 173-182.

[6] Y. Koren, R. Bell and C. Volinsky, Matrix Factorization

Techniques for Recommender Systems, Computer, 42, 8 (August 2009),
30-37.

[7] Y. LeCun, Y. Bengio and G. Hinton, Deep Learning, Nature,

521(7553), 436 (May 2015), 436444.

[8] J. Leskovec, A. Rajaraman and J. D. Ullman, Mining of Massive

Datasets, Cambridge University Press, (2014).

[9] T. Mikolov, I. Sutskever, K. Chen, G. Corrado and J. Dean,

Distributed representations of words and phrases and their
compositionality, In Proceedings of the 26th International Conference
on Neural Information Processing Systems - Volume 2 (NIPS’13),
(2013). Vol. 2. Curran Associates Inc., USA, 3111-3119.

[10] [10] H. Wang, N. Wang and D.-Y. Yeung, Collaborative Deep

Learning for Recommender Systems, In Proceedings of the 21th ACM
SIGKDD International Conference on Knowledge Discovery and Data
Mining (KDD ’15), (2015). ACM, New York, NY, USA, 1235-1244