Application of deep learning recommender system

Aspire Thought Leadership! Wondering What is deep learning recommender system? Find out more on deep learning recommender system and how fortune 500 c

The advancement in the field of information technology in recent times has made the access of huge amounts of data very easy. There are a vast variety of products and services whose description is easily available along with their comments and reviews. Due to this overload of information, it becomes very difficult for the user to choose an appropriate product according to his requirements [impact of ai in business].

To deal with the above problems, recommendation systems are used. Deep learning recommender Systems provide the users with personalized recommendations. There are many areas where recommendation system is being used, such as music, books, movies, shopping etc. Most of the online vendors have a recommendation engine already equipped [artificial intelligence business]. Recommendations are done on the basis of user’s previous items choice or the items preferred by similar user or on the basis of the description of item. Here the item refers to the product or services which is to be recommended. Recommendation system is classified into three parts based on their approach: Content based, collaborative filtering and hybrid.

Recently, the use of artificial neural networks has become very popular in the problems which require complex computations and huge amount of input data [Data science workflow]. Deep learning is a part of ANN architectural models of deep neural network are efficiently built and trained. Deep neural networks have its applications in various fields such as speech recognition, image processing, object recognition, image processing, NLP tasks etc. Due to various advantages of deep learning, researchers have been encouraged to use its associated techniques in the field of recommendation system also.

Deep learning recommender system

Deep learning is being used successfully in recommendation system as well as many other fields in computer science and has shown significant improvements in the existing models. In 2007, a collaborative filtering method for movie recommendation system was given by which utilized the hierarchical model of deep learning [Reinforcement Learning]. In 2015, with the use of auto encoders, predicted the values which were missing in the user-item matrix.

The sparsity issue in recommendation system in collaborative filtering was addressed by Various surveys have been done in the area of deep neural network based recommendation system. The state-of-theart survey for deep recommendation system has been done. In 2017, a comprehensive review on deep learning based recommendation system has been done. The paper proposed the classifications on the basis of their structure that is neural network models and integration models.

Background and terminologies

A recommendation system is used for information filtering and outputs a list of specific products in a personalized manner. For examining how the two fields that are recommendation system and deep learning are integrated together, one should know the basics of both these fields. This section of the paper includes a brief description about the fundamental classifications and challenges of both the fields. First, there is an introduction of types of recommendation system and then the details about deep learning methods.

Deep learning recommender system

Various recommendation system approaches are evolved over time and applied in various applications. User finds to filter the huge amount of data available as important task, to find a useful, tailored and relevant content [components of data science]. The recommendations that are predicted by the recommendation System helps the users in taking a decision.

In a traditional recommendation system, the recommendations can be made in two different manners, i.e., predicting a particular item or preparing a ranking list of items for a particular user. The recommendation model is categorized into Content based model, Collaborative filtering based model and hybrid recommendation framework.

1. Content based recommendation framework: In this model, the items which are same in content is searched. The profile of user is established on the basis of items on which the user is interested in. According to the profile generated, the recommendation system searches the database for the appropriate items using the descriptive attributes of the item. If we use this recommendation system for an item which is newly added, then content based recommendation system works very efficiently. The problem with new inserted item is that it may not have any rating, but still the algorithm works since it uses descriptive information for recommendations. The limitation of this method is that it cannot recommend diverse range of products since the algorithm does not take the information from similar users.
2. Collaborative filtering recommendation systems: This recommendation system assumes that users’, who have previously preferred same items, would have same choice in future also. In this system, the recommendations are done on the basis of similar users’ pattern rather than descriptive features of items. A correlation among the users is determined, depending upon the choice of similar users, the items are recommended.

There are two methods which are followed by Collaborative filtering algorithm: memory based algorithm and model-based methods.

In memory based algorithm the complete user-item matrix is taken into consideration for identifying similarity. After finding out the nearest neighbor, on the basis of neighbors past rating, the recommendations are provided. In model based algorithms, an offline model is built with the use of machine learning algorithms and data mining methods [What is data science?]. These models include clustering models, Decision models, Bayesian model and singular value decomposition model.

Hybrid Recommendation Framework: It is an integration of content-based model and collaborative filtering model; it incorporates the benefits of both the methods and tries to eliminate the limitations of the above models. There are many hybrid systems proposed, some of them are as following:

Cascade: The output of one approach is later on used by other approach

Switching: Recommendation output is given on the basis of current situation and either of the one approach is used. Weighted: the combination of the scores of various approaches is used for recommendation.

Deep Learning recommender system techniques

Deep learning is based on learning many layers of representations with the help of artificial neural networks, and is a part of machine learning [Machine Learning introduction]. Deep learning has its applications in variety of fields like Computer vision, recognition of speech, natural language synthesis etc. The important factors which increase the importance of deep neural network as the state-of-the-art machine learning [supervised machine learning & unsupervised machine learning] methods are as following:

• Big data: As the amount of data increases, better representations are learnt by the deep learning model [what is big data?]. 
• Computational power: The complex computations of the deep learning model is done using the GPU In this section, there is a description of various deep learning models which are used in recommendation systems. 

Autoencoder

Autoencoder is an unsupervised learning technique in which neural networks are applied for the task of representation learning. Specifically, a neural network architecture is designed which imposes a bottleneck in the network which forces a compressed knowledge representation of the original input as shown in figure 1. In autoencoder some representations from the encoded input are found by the training, so that the input can be restored back from these representations. Autoencoder have three layers arranged as the input, hidden and output layer. There are equivalent numbers of neurons in the input and output layer. The compressed representations are obtained from the hidden layer, and with the help of these representations, the first layer inputs are reconstructed at the output layer of autoencoder.

In the learning process, two mappings are used, with the help of encoder and decoder. An encoder is fully connected deep neural network which transforms the input into a latent space representation. Decoder is also having a similar structure as encoder and it is responsible for reconstructing the input back to the original form from the hidden layer outputs. There are many variants of autoencoders such as sparse autoencoder, stacked autoencoder, denoising autoencoder and variational autoencoder etc. Data Compression, Image Denoising, Dimensionality Reduction, Feature Extraction, Image Generation and Image colorization are few important applications of autoencoders.

Wide & deep learning for recommender systems

Recurrent Neural Network

Sequence prediction issues have been around for quite a while. They are considered as probably the most difficult issue to settle in the information science industry. These incorporate a wide scope of issues, from foreseeing deals to discovering designs in financial exchanges. In the ordinary feed-forward neural organizations, all experiments are viewed as free. That is when fitting the model for a specific day; there is no thought at the stock costs on the earlier days. This reliance on time is accomplished by means of Recurrent Neural Networks. An ordinary RNN structure glances as demonstrated in figure 2.

Recommendation system deep learning

In RNN, the sequence of values, i.e., x(0) , x(1) ,..., x(t) is processed. On each element of the sequence, the same task is performed and the output is based on previous computations. RNN uses an internal memory to hold the values of previous computations, so that it may be used later. Recurrent Neural Networks work moderately okay when we are dealing with short-term dependencies. But RNN fails to deal with large term dependencies. The reason behind this is the problem of Vanishing Gradient and exploding gradient. To remove this problem, Long Short Term Memory (LSTM) and gated recurrent unit is used. It is used when the processing is to be done for predicting events which has comparatively longer interval and delays. LSTM has a processor to distinguish the useful information from the information which are not useful. This processor is known as cell. There are three gates in LSTM namely input gate, output gate and forget gate. The structure of LSTM is shown below in figure 3.

A forget gate is answerable for eliminating data from the cell state. The data that is not, at this point needed for the LSTM to get things or the data that is of less significance is taken out by means of augmentation of a channel. This is needed for advancing the exhibition of the LSTM organization. The info entryway is answerable for the expansion of data to the cell state.

Deep learning for recommender systems

The work of choosing helpful data from the current cell state and showing it out as a yield is done by means of the yield gate.RNN are valuable while managing worldly elements of communications and when the client's conduct has a specific arrangement of pattern.

Convolution Neural Network

A CNN or convolutional neural network is feed-forward neural organization that is by and large used to dissect visual pictures by preparing information with matrix like geography. It's otherwise called a ConvNet. A convolutional neural organization is utilized to recognize and characterize objects in a picture. A convolution neural organization has numerous secret layers that help in extricating data from a picture. 

The four significant layers in CNN are:

1. Convolution layer
2. ReLU layer
3. Pooling layer
4. Fully connected layer

This is the initial phase during the time spent separating significant highlights from a picture. A convolution layer has a few channels that play out the convolution activity. ReLU represents the corrected direct unit. When the component maps are extricated, the subsequent stage is to move them to a ReLU layer. ReLU plays out a component insightful activity and sets every one of the negative pixels to 0. It acquaints nonlinearity with the organization, and the produced yield is an amended element map. Pooling is a down-inspecting activity that decreases the dimensionality of the component map. The corrected component map presently goes through a pooling layer to create a pooled highlight map. The following stage in the process is called straightening. Leveling is utilized to change over every one of the resultant 2-Dimensional exhibits from pooled highlight maps into a solitary long consistent direct vector. The smoothed lattice is taken care of as contribution to the completely associated layer to arrange the picture. There are many applications in which CNNs are applied such as object recognition, self-driving cars, audio processing etc. while transforming the input to output. The structure of the typical CNN model is shown in figure 4. 

Neural network recommender system

Restricted Boltzmann Machine

RBM is developed by Geoffrey Hinton, and are stochastic neural organizations that can gain from a likelihood conveyance over a bunch of sources of info. It comprises of two layers of neural organization in particular an obvious and a secret layer. Each noticeable unit is associated with all secret units. RBMs have a predisposition unit that is associated with every one of the noticeable units and the secret units, and they have no yield hubs. The construction of the RBM model is appeared in figure 5.

Deep learning based recommender system

The complex computations and learning in RBM is based on characteristic articulation of information. The word “restricted’ is used for intra-layer communication as it is not present in both hidden layer and visible layer. Due to this restriction, the learning efficiency increases. There is a full connection between the nodes of different layers which are stacked together and there is no connection between the nodes of same layer. Applications of RBM includes dimensionality decrease, characterization, relapse, community separating, highlight learning, and theme displaying. Since RBM utilizes a basic forward encoding activity, so it is quick when contrasted with different models, for example, autoencoder.

Application of deep learning in recommendation system

DL techniques are widely used in various real world applications such as sentiment analysis, speech recognition image classification, text classification etc. Various researchers also include deep neural network based techniques in the field of recommendation system to improve its performance as compared to traditional recommendations systems. Traditional recommendation methods use matrix factorization methods which have some limitations such as: 

1. The trouble of utilizing side highlights that may influence the suggestion like the U/PG rating of a film or the nation of the client. We can just utilize the thing ID and the client ID on account of network factorization. It likewise keeps us from questioning a thing or client not present in the preparation set. 
2. The matrix factorization additionally had the virus start issue because of the way that it had no component vector or inserting for the new things.
3. Matrix factorization regularly will in general prescribe mainstream things to everybody which doesn't generally mirror the particular client interests for the most part when Dot items are utilized.
4. Matrix factorization deals with the straightforward internal result of the User and thing highlight embeddings, it is regularly insufficient to catch and address the mind boggling relations in the client and things.

Deep Neural networks are planned and used to address these weaknesses of the matrix factorization strategies. This section describes the various categories recommendation systems which are based on deep learning.

The categorization is based on the types of recommendation systems used which is as follows:

1. Collaborative filtering recommendation system based on deep neural network.
2. Content-based recommendation system based on deep neural network.
3. Hybrid recommendation system based on deep neural network.
4. Social network-based recommendation system based on deep neural network.
5. Context aware recommendation system based on deep neural network.

Hybrid model and neural network model are the two categories of deep neural network-based recommendation system. Integration model is further divided into two categories on the basis of whether it combines any traditional recommendation system model with deep neural network technique or depends solely on deep learning method. 

Neural network model is also divided into two categories on the basis of deep neural network based technique used: models which uses single deep neural network based technique and deep neural network based composite model. In deep neural network based composite model, different deep neural network techniques are used to build a hybrid system having more capability.

Collaborative Filtering Recommendation Systems Based on Deep Neural Networks

Collaborative filtering (CF) is one of the usually implemented techniques in recommendation systems in order to tackle various real-life issues. The state of the art CF-based methods uses the rating matrix for recommending the items. But this approach faces the problem of data sparseness and cold start problem. The sparsity of the user-item matrix, the learned features is not effective which reduces the performance of recommendation system. Various researchers propose deep neural network based collaborative filtering techniques to enhance its effectiveness in recommendation.

Collaborative Filtering Method Based on Generative Adversarial Network

Generative Adversarial Network is a neural network which is generative and having discriminator and generator functions. These both functions are simultaneously trained in competition with one another in architecture of minimax game. The first model to implement GAN in the field of Information Retrieval is (IRGAN) which stands for Information retrieval generative adversarial network. The state of the art GAN model has two modules a discriminator and a generator. The generative retrieval module predicts appropriate documents with given query, whereas discriminative retrieval module predicts relevancy given with a pair of query and document.

The IRGAN model combines above two Information Retrieval models in order to play a minimax game with them: the generative retrieval model produces (or selects) relevant documents that are relevant documents like ground truth, while the discriminating retrieval model separates the relevant documents from those generated by the generative retrieval model. concentration is on the semantic-rich client thing communications in a recommender framework and propose a novel generative adversarial network (GAN) named Convolutional Generative Collaborative Filtering (Conv-GCF). They build up a powerful irritation system (ill-disposed commotion layer) for convolutional neural organizations (CNN), in light of which a generator is planned with lingering squares to combine client thing collaborations.

Recurrent Neural Network Based Collaborative Filtering Method

In order to deal with the information in sequential form, recurrent neural network (RNN) proves to be a very effective network. Concepts of loops are used in place of feedforward network to remember sequences. Variants of RNN viz. Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) network are used to deal with the problems of long term dependencies and vanishing gradient problem. In collaborative filtering method based on RNN, the impact of user historical pattern is modelled on the current behavior of user, recommendation is performed and user’s behavior is predicted. Figure 7 shows the framework of collaborative filtering method based on RNN. Let the input set is {I1, I2 . . . It}, and output is Ot = σ (f (W·ht−1 + V·It)·V), σ represents a softmax function, f represents the activation function, which specifies the selection probability of any item at time t. ht represents the hidden state vector.

Collaborative filtering model based on RNN

Collaborative filtering method based on Autoencoder

Collaborative Filtering Method Bssed on Autoencoders

The first ever developed autoencoder-based collaborative recommendation model is Autoencoder based Collaborative filtering. It decomposes the vectors by integer ratings. The model proposed by takes client or thing based evaluations as contributions to rating matrix R. The output is produced by the process of encoding and decoding by optimizing the parameters of model and reducing the reconstruction error. Consider an example, if the range of integers [1-5] represents the rating score, then each rui can be divided into five vectors. 

Above figure represents the 1 to 5 rating scale in which blue boxes represents the user rated item. The cost function which is to be reduced is taken as Mean Square Error. The rating prediction in this approach is found by making the summary of each of the five vectors, which are scaled by rating K. Pretraining of parameters and local optimum avoidance is performed by RBM. Stacking multiple autoencoder collectively shows the slight improvement in accuracy. This method based on autoencoder suffers from the problem of dealing with noninteger ratings and sparseness of input data due to decomposition of partial observed vectors.

Collaborative Denoising Auto-Encoder is primarily used for prediction rankings. User feedback is taken as input to the CDAE. If the user enjoys a movie, the input value is 1 otherwise it is 0. It shows the vector preference to display the user's interest in some item. Gaussian noise corrupts the CDAE input.

Collaborative Filtering Method Based on Restricted Boltzmann Machine

Restricted Boltzmann Machine (RBM) is a double-layer neural network capable to deal with typical learning based problems. The efficiency of learning is improved by removing the connections between same layers.

Further a conditional RBM model is proposed to consider information in form of feedback. The visible layer of RBM can take only binary values so only one hot vector can be used to represent the rating score. The architecture of RBM based model is represented as shown in Figure 8.

Collaborative filtering model based on RBM

Content-Based Recommendation Systems Based on Deep Neural Networks

Content-based recommendation systems recommend on the basis of descriptive attributes of items and users’ profiles such as texts, pictures and videos. The use of deep neural network in content-based recommendation systems is to capture the non-linear relationships between user and item. It also captures the intra-relationship between data, from large available data sources.

Hybrid Recommendation System Based on Deep Neural Networks

The state-of-the-art Collaborative Filtering-based approach uses the ranking matrix to suggest products. But this method suffers from a data sparsity and a cold start crisis. Due to the scarce existence of the useritem matrix, the features learned are not accurate, which decreases the efficiency of the recommendation framework. Hybrid recommendation model based on deep neural networks incorporates a content-based recommendation model with collective recommendation-based filtering models, which combines the mechanism of feature learning and recommendation into a single model.

An autoencoder is a variation of neural organization, having encoder and decoder as two components. The encoder changes over the contribution to its secret portrayal, while the decoder changes over the secret portrayal back to the re-established input structure. The boundaries comparing to the autoencoder are prepared to limit the mistake because of the recreation, which is estimated by the loss work or misfortune work. Denoising autoencoder (DAE) attempts to recreate the contribution from a ruined adaptation for improved portrayal from the info. More variations of autoencoder have been created for better results. 

The crossover suggestion model dependent on the stacked denoising autoencoder utilizes both the rating network and the side data and coordinates both the SDAE and the grid factorisation. Lattice factorization is a broadly utilized model with improved exactness, and SDAE is an incredible model for separating significant level highlights from inputs. The combination of the over two model will turn into an amazing model for additional advantages.

Social Network-Based Recommendation System Using Deep Neural Networks

Conventional recommendation models never consider social connections among the users. But we always take verbal recommendations from our friends in reality. These verbal recommendations are termed as social recommendation which occurs daily. Hence, for improved recommendation systems and for more personalized recommendations, social network must be employed among users. Every user will interact with various types of social relationships.

The quality of recommendation system is very crucial task which can be achieved by implementing the effect of social relationship among the users. Items with location attributes and sequential pattern of user behaviour in spatial and temporal frame are used to form spatiotemporal pattern which is used to improve recommendation accuracy. Recently, a very few recommendation techniques have been proposed which is based on the users’ trust relations improve conventional recommendation systems. These trust based recommendation models proves to be an effective move in the field of recommendation system models.

In current scenario an integration of deep learning and social network based recommendation system provides a platform for various research solutions. The limitations which are inherent to the social recommendation must be addressed in the future research.

Context-Aware Recommendation Systems Based on Deep Neural Networks

A context-aware recommendation system, integrates context based information into a recommendation model. This integration is effectively performed by the deep learning techniques in different conditions of recommending items. Deep neural network based methods are used to extract the latent space presentation from the context based information. Deep learning based model can be integrated into diverse data to reduce data sparsity problem.

Sequential nature of data plays a significant part in implementing user behaviours. Recently, recurrent neural networks (RNNs) are commonly used in a variety of sequential simulation activities. However, for real-world implementations, these approaches have trouble modeling contextual knowledge, which has been shown to be very essential for behavioural modelling.

Currently this method based on deep neural networks focused towards situation information. A novel approach is proposed called context-aware recurrent neural networks. It uses two types of matrices: input and transition matrices. They both are specific to the context and adaptive in nature. Input matrices are used to extract various situations such as time, place, weather condition where actually the user behaves.

6. Applications

As deep learning plays a significant role in most of the fields as it has the capability of dealing with large and complex problems with improved results. Deep learning technology also contributes in the field of recommendation system for improved customer satisfaction. Deep learning technology overcomes the shortcomings of traditional models to get high quality recommendations.

All the above discussed methods of recommendation systems use deep neural networks and hence also achieve the quality in recommending items to the users. Different recommendation models use different deep learning methods to obtain improve results.