Survey on Generations of Networks Recommender Systems

In online e-commerce websites, recommender engines collect the preferences and tastes of customers by considering the product ratings. It is all about Machine Learning (ML) algorithms and data science. People and commodities interact in various ways, and these recommendation engines are constructed on top of these interactions, which include visitors, readers, shoppers, app users, and more [1-3].

This section explores the different first-generation networks recommender systems – content based, Collaborative Filtering (CF), and hybrid systems.

Content-based recommendation systems

The items are recommended based on a user’s profile and the item’s content, which is dependent on the user’s rating. The user ratings and opinions are used to make this comparison. It is possible to rate an item in two ways: directly by clicking like or dislike buttons or giving it a rating; and implicitly by simply looking at reading, ordering, adding, or purchasing the item [2-4]. In contrast, getting an acceptable form of feedback can be challenging. The interaction matrix is constructed for the objects and users using cosine similarity to compare the data. Item-based recommender systems are another name for these systems. Users’ dependency and complicated behavior are the key drawbacks of these systems.

CF systems

Recommender systems that consider “user behavior” are used this way. To propose goods to new users, the information should be collected about previous users and things in the form of ratings, transaction history, purchase, and selection data. The mapping between new and current users is created, which is used to provide suggestions. The neighborhood technique might be used to finish this mapping [3]. KNN (K-Nearest Neighbor) and APRIORI may be used to develop the neighborhood approach. KNNs are sometimes called “slow learners” due to their inability to recall knowledge quickly. The similarity is once again calculated using the cosine similarity and correlation functions. These systems have a few drawbacks, including a slow response time for new users since their ratings have yet to be established and lack good recommendations.

Hybrid recommender systems

Personality-based recommender systems or hybrid systems solve the flaws of all earlier strategies. In addition to methodologies, the term “hybrid” refers to the combination of user-related elements, such as the user’s history and personal behavior. All of the aforementioned approaches may be combined into one recommender system. Facebook’s rotating hybrid method uses the hybrid technique.

This section explores the various second-generation recommendation systems – matrix factorization, web usage mining and personality based recommenders [3-5].

Matrix factorization-based approach

Latent factors are features that may be utilized to connect a piece’s content to its genre. Another kind of latent component model is based on neural networks. The predictions are made based on similarities between user and object components. The input data matrix depicts one dimension, and the objects themselves, represent the other. Items and users are integrated to create a combined latent factor space. The intrinsic product in that area shapes the interactions between users and objects. Evaluating the item’s vector can figure out how much it will cost to obtain these components. When determining the user’s degree of interest in a set of items, the vector is used to relate the user to that user’s interest. Using the dot products, a sense of how much interest a user has in a specific item is gained. This is how dot products may be used to produce recommendation ratings, which can then be used to forecast future events. The predictions are based on factors that may contain latent components that are factored into two vectors. Based on how a user rates something, latent factors may be used to identify what kind of content an item or an article could be, and this information can be used to determine what type of content an item might be.

Web usage mining-based recommendation

Using data science techniques, web usage mining extracts information from many web users. It is utilized for various purposes, including website modifications, system upgrades, and personalization. There are both offline and online components to the procedure. The knowledge databases are framed by the offline component, which forecasts previously available information that the online component then utilizes. Many recommender systems, notably suggest 3.0, use this method that delivers relevant information about sites of interest to the user. The person who has never visited before is linked in real-time, maybe because of an interest, the user may have shown. Without the requirement for any offline components, this recommender system may handle real-time websites made of individual pages.

Personality-based recommenders

Personality-based recommenders take into account both a customer’s purchasing history and their personality when making suggestions. That is how the automated personality classifier is linked to information gleaned from a user’s purchase history. This approach alleviates the issue of portal sales being harmed by user distraction. The identification of a user’s personality may be made in two ways: It is possible to get an idea of a person’s character by having them fill out a personality questionnaire. However, this technique has a significant fault in that it will be challenging to deal with missing data if one user fails to respond appropriately or if any user does not finish the questionnaire. The user’s personality is determined by analyzing publicly available data. Using Alexandra Roshchina’s second evaluation approach, TWIN (Tell Me What I Need) creates personalized recommendations depending on the user’s personality. To build a map of the user’s personality and writing, psychologists previously spent a lot of time and effort extracting specific chunks of text from the user’s writing. The system is built on the results of these researchers. It is the writing style used to analyze nature since it is based on the personality created from the content provided by the users. Hybrid recommenders combine CF and content-based approaches.

Frank Rosenblatt, a psychologist, set out to build an artificial brain in 1958 to mimic the human brain’s reciprocal behavior. In the same way the human brain processes information, the perceptron can learn, think, and count. AI software beat a top GO player in 2016. Lee Sedol, one of the game’s most accomplished players, was beaten by AlphaGo. Recommender Systems increasingly rely on Machine Learning (ML), which has led to a new generation of systems in networking applications [5-7].

Deep learning (DL)

DL or ML is a branch of AI fuelled by the human brain’s activity, incorporates composite engineering, and is used to duplicate and train models. The Neural Networks (NN) are the foundation of DL. A NN is an artificial model of the human brain’s network of neurons. To achieve DL, various types and sizes of NNs are used. Multilayer Perceptron, Convolutional NN (CNN), and Recurrent NN (RNN) are just a few examples of NNs. By tackling the challenge of physical feature abstraction, CNN revolutionized digital image processing. In addition to face and object identification, they have been utilized well for image classification, segmentation, and other tasks. In the 1980s, the first RNN was developed. Time series data may be analyzed using RNNs.

CF and restricted boltzmann machines

For the Harmonium project, Paul Smolensky developed the restricted Boltzmann machines (RBM) in 1986. There are two layers in this model. Starting with evaluating particular scenarios, we employ user evaluations on items in the system. Users may now be used to train the systems if each user scores each item. Due to one user failing to rate certain items, ratings are currently missing. The solution here is to create a separate RBM for each user. As previously shown, if the reciprocal user only rated a small number of items, their RBM has local connections. CF methods may be utilized with RBMs to create a model. Gradient ascent is used to maximize log-likelihood throughout the learning process. If the concealed unit’s states are computed, and then the anticipated value of expecting an item is calculated, then predictions may be produced. It is a downside of this technique because it does not employ reviews and user profiles to leverage content-related information. As a consequence, the cold start issue needs to be better addressed

CFDL

CFDL provides a robust solution to the cold start issue. Stack Denoise Auto-Encoder and Collaborative Topic Regression (CTR) are used to learn the item’s latent components from the contents linked with the reviews. The bag-of-words technique is used to reproduce the reviews in a matrix for CFDL’s learning process from item evaluations. The matrices that service the K items are represented by the latent variables linked with items learned using an SDAE network and a matrix composed of bag-of-words vectors. An alternative approach is to combine two Gaussian distributions, one for each user’s latent characteristics, to create a more accurate model for predicting ratings. This model’s latent elements and network parameters must be altered during training. Rating models are linked to the rest of the network through latent variables generated by intermediate layers. To calculate the CFDL log probability, particular hyper parameters are used, which may be optimized using an EM-style technique. In-text modeling applications, where knowledge of “word order” is critical, the most significant drawback is that it is lacking.

Recommendation systems based on deep content

The most common use of this technology is to give users personalized music recommendations. The cold start is the most typical difficulty, and an implied review on a music-focused website or app is the second. It’s a simple reality that a recommendation engine should suggest anonymous content. In contrast to standard matrix factorization, users are more likely to offer implicit feedback than explicit input, which results in a different framework. The weighted matrix factorization, a mutant matrix factorization, is used in this circumstance. Implicit reviews are included in this approach.

Product-based recommendation system

When constructing production-based recommenders, systems must be dynamic in every way possible. Because we create so much data rapidly, the current data we have to deal with is reasonably current. To retain “Freshness,” a recommender system based on production should be dynamic (to the fullest extent feasible). Inadequately methodized data may be found in a massive volume of data and the metadata that goes along with it. Before making suggestions to the customer, we gather data from them by engaging them with a service like a website or an app. The “Front End” comes into play at this point. The data transport model is then used to convey the data imported from consumers. The “Storage” model is then used to store this data, which may subsequently be used to provide projections and make recommendations. Now comes the training portion, which is divided into offline and online components. Methods such as offline training, which uses machine learning to test the recommendation algorithm, are the most common. It entails developing a model via the use of widely accepted data science techniques.

On the other hand, DL models of learning might also be used in this situation. The online component updates the batch with a new picture since the data flow is dynamic. A matrix factorizationbased approach like ALS (Alternating Least Squares) may be used, which is dynamic and straightforward to implement. The last step is servicing, which comprises putting the model into action and providing the consumer with suggestions. This is based on the output of the trained model. It’s about more than just promoting excellent content; it’s about making the recommender system dynamic, responsive, and challenging. In the selection of an algorithm, we should place more emphasis on ease of use.

Networking recommender systems are becoming more crucial at a time when the internet is experiencing its most prosperous period ever. New developments in networking recommender systems are required as the world changes, and the advent of DL in this area has already begun this process [8-9]. However, this introductory process is still in its early phases, and there is much more to study, rehearse, and perform to solve real-world networking problems through different soft computing strategies [10-23].

1. Zhang S, Yao L, Sun A, Tay Y (2019) Deep learning based recommender system: A survey and new perspectives. ACM Comput Surv 52(1): 1-38.
2. Zhang Y, Chen X (2020) Explainable recommendation: A survey and new perspectives. arXiv preprint arXiv 14(1): 1-101.
3. Lu J, Zhang Q, Zhang G (2020) Recommender systems: Advanced developments. World Scientific p. 352.
4. Zhang Q, Liao W, Zhang G, Yuan B, Lu J (2021) A deep dual adversarial network for cross-domain recommendation. IEEE Transactions on Knowledge and Data Engineering p. 1.
5. Liao W, Zhang Q, Yuan B, Zhang G, Lu J (2022) Heterogeneous multidomain recommender system through adversarial learning. IEEE Transactions on Neural Networks and Learning Systems pp. 1-13.
6. Zhao Z, Zhang X, Zhou H, Li C, Gong M, et al. (2020) Hetnerec: heterogeneous network embedding based recommendation. Knowledge-Based Systems 204: 106218
7. Guo Q, Zhuang F, Qin C, Zhu H, Xie X, et al. (2022) A survey on knowledge graph-based recommender systems. IEEE Transactions on Knowledge and Data Engineering 34: 3549-3568.
8. Bhaskaran S, Marappan R, Santhi B (2021) Design and analysis of a cluster-based intelligent hybrid recommendation system for e-learning applications. Mathematics 9(2): 197.
9. Bhaskaran S, Marappan R, Santhi B (2020) Design and comparative analysis of new personalized recommender algorithms with specific features for large scale datasets. Mathematics 8: 1106.
10. Marappan R, Sethumadhavan G (2020) Complexity analysis and stochastic convergence of some well-known evolutionary operators for solving graph coloring problem. Mathematics 8(3): 303.
11. Marappan R, Sethumadhavan G (2018) Solution to graph coloring using genetic and tabu search procedures. Arab J Sci Eng 43: 525-542.
12. Marappan R (2021) A new multi-objective optimization in solving graph coloring and wireless networks channels allocation problems. Int. J. Advanced Networking and Applications 13(2): 4891-4895 (2021)
13. Sethumadhavan G, Marappan R (2013) A genetic algorithm for graph coloring using single parent conflict gene crossover and mutation with conflict gene removal procedure. 2013 IEEE International Conference on Computational Intelligence and Computing Research, Enathi, India, pp. 1-6.
14. Marappan R, Sethumadhavan G (2013) A new genetic algorithm for graph coloring. 2013 Fifth International Conference on Computational Intelligence, Modelling and Simulation, Seoul, Korea (South), pp. 49-54.
15. Marappan R, Bhaskaran S (2022) Analysis of network modeling for real-world recommender systems. International Journal of Mathematical, Engineering, Biological and Applied Computing 1(1): 1-7.
16. Marappan R, Bhaskaran S (2022) Movie recommendation system modeling using machine learning. International Journal of Mathematical, Engineering, Biological and Applied Computing 1(1): 12-16.
17. Marappan R, Bhaskaran S (2022) Analysis of recent trends in e-learning personalization techniques. The Educational Review 6(5): 167-170.
18. Marappan R (2022) Classification and analysis of recommender systems. International Journal of Mathematical, Engineering, Biological and Applied Computing 1(1): 17-21.
19. Marappan R, Bhaskaran S (2022) Datasets finders and best public datasets for machine learning and data science applications. COJ Rob Artificial Intel 2(1).
20. Marappan R (2022) Recommender system for movielens datasets using an item-based collaborative filtering in python. International Journal of Mathematical, Engineering, Biological and Applied Computing 1(1): 42-43.
21. Marappan R (2022) Create a book recommendation system using collaborative filtering. International Journal of Mathematical, Engineering, Biological and Applied Computing 1(1): 44-46.
22. Marappan R (2022) Open-source datasets for recommender systems analysis. International Journal of Mathematical, Engineering, Biological and Applied Computing 1(2): 49-51
23. Marappan R (2022) Heart disease prediction analysis using machine learning algorithms. Journal of Applied Mathematics and Computation 6(3): 273-281.