Artificial intelligence (AI)-based recommendation engines/recommender systems utilize different AI techniques to analyze substantial amounts of data and make personalized user recommendations. These systems have been designed to assist users in finding valuable and relevant products or content based on their history, behavior, and preferences. This article discusses the AI techniques used in recommender systems and the major use cases.
Image credit: Photon photo/Shutterstock
AI Techniques in Recommendation Engines
Recommendation engines are used extensively in several industries, including e-commerce, supply chains, social media, media and entertainment, healthcare, travel and hospitality, and financial services, to increase revenue generation and improve user satisfaction, experience, retention, and engagement. AI techniques are increasingly used in recommender systems as they can provide high-quality recommendations compared to conventional recommendation methods.
Several AI techniques can be utilized in recommendation engines, including multilayer perceptron (MLP), autoencoder (AE), convolutional neural network (CNN), recurrent neural network (RNN), generative adversarial network (GAN), graph neural network (GNN), transfer learning (TL), active learning (AL), reinforcement learning (RL), fuzzy techniques, natural language processing (NLP), and computer vision.
MLP and AE: MLP combines the advantages of non-linear and linear modeling in a single recommendation framework. Neural collaborative filtering (NCF) based on MLPs can be used with matrix factorization to model the non-linear relationship between items and users and the linear relationship. NCF is utilized extensively in recommender systems as a general model for user-item interactions.
AutoRec integrates an AE with matrix factorization to learn the non-linear latent representations of items/users. Similarly, AutoSVD++ is primarily a hybrid method that combines a contractive AE and matrix factorization to produce item feature representations from item content.
The robustness of AutoRec can be improved using denoising techniques and by integrating side information as user-contributed tags or item content. Moreover, AE is the foundation for representation learning, which is suitable for item representation learning and user profiling in recommender systems.
CNN, RNN, GAN, and GNN: DeepCoNN can jointly model items and users through reviews by integrating two parallel neural networks. A shared layer facilitated by factorization machines connects the two CNNs.
Similarly, ConvMF integrates CNN into matrix factorization to increase the rating prediction accuracy, address the data sparsity problem, and exploit the information in user-contributed reviews. In microblogging, CNN can also be utilized for the hashtag recommendation task by introducing the attention mechanism in the hashtag selection process.
RNN is primarily employed to model and analyze the evolution of item features or user interests, as the technique is suitable for sequential data. In one study, a co-evolutionary latent feature process was proposed, and RNN was applied to model the user-item interaction temporal dynamics.
A long short-term memory (LSTM)-based model can be used to capture the user behavior dynamics to predict whether the existing user behavior should be inherited. LSTM can also be utilized in recommender systems to make in-time music recommendations.
Moreover, RNNs can be used in sequential recommender systems/session-based recommender systems where the real-time recommendation is refined based on historical sequential data. Sequential recommender systems are receiving significant attention in studies focusing on the relationship between long-term and short-term interests and integrating preference dynamics and contextual information.
GAN-based information retrieval systems are suitable for recommender systems. A GAN can be used to learn robust item/user representations from user-item interactions and knowledge graphs, tags, and images. A study introduced perturbations on the item and user embedding as an adversarial regularizer under the Bayesian personalized ranking framework.
GNNs can learn features for nodes from the graph's neighborhood information, making them suitable for recommender systems as the item-user relationships are typically represented as a bipartite graph. A highly efficient and scalable recommendation method in which the feature embedding by a GNN and random walk were incorporated was proposed in a study and deployed in Pinterest. The results demonstrated the significant potential of GNNs to improve the recommender system productivity.
Another study proposed a generalized GNN-based collaborative filtering (CF) framework with an attention-based message-passing method for information propagation. GNN is also suitable for sequential recommender systems for modeling the item sequences as a graph and is more effective than RNN as user-item interactions are considered in the sequence in GNN, while an RNN can only model one-side item information.
TL, AL, and Fuzzy Techniques: TL can extend recommendation requests from a single domain to multiple domains. All domains can benefit from mining user preferences unavailable in single-domain data by exploiting the correlation of multiple domains. Additionally, the cold-start/data sparsity problem can be addressed by exploiting several domains, as many domains possess insufficient data while many others possess rich data.
The demand for a diverse and rich recommendation and the ability to eliminate the data sparsity problem has facilitated the development of cross-domain recommender systems (CDRS). The lack of explicit feature space in CDRS is the key difference between CDRS and other TL methods.
Thus, CDRS cannot be characterized as a single type of TL method due to the practical application of several TL techniques. CDRS can offer multi-domain recommendations for online shopping retailers selling different goods while simultaneously solving the data sparsity problem.
AL can assist recommender systems to select the most representative items and deliver those items to users to rate. AL techniques can improve both the accuracy and efficiency of recommender systems. Different AL strategies, such as bootstrapping and rating impact analysis, can be integrated with common recommendation models, such as matrix factorization, decision trees (DTs), and the aspect model. Complex factors, such as the influence of items, the item attributes, and naturally acquired ratings by users, can be incorporated into the AL strategy.
Fuzzy techniques can be used for profiling and matching appropriate items in content-based recommender systems. A fuzzy user-interest drift detection approach can measure user-interest consistency using fuzzy relationships to deal with dynamic user preferences in rapidly changing big data.
AI-based Recommender Use Cases
Social Media: In social media platforms, AI-based recommendation systems can recommend content to users based on their preferences, interests, and behavior. The recommender system can analyze the user’s social media activity to identify preferences and patterns and recommend relevant content to the user.
Financial Services: In financial services, recommender systems can recommend financial products to customers that can potentially meet their financial needs based on the customers’ preferences, financial history, and behavior. These financial product recommendations can increase the revenue of financial services players and improve customer satisfaction.
Travel and Hospitality: Recommender systems can make personalized travel recommendations/suggest travel-related services, such as flights and hotels, to customers to meet their needs based on the customers’ budget, behavior, and preferences.
Supply Chain Management: In supply chain management, AI-based recommendation systems can be utilized to reduce wastage and optimize inventory. The recommender system can analyze supplier lead times, sales data, and historical inventory levels to predict demand and recommend optimal ordering and inventory management strategies.
These strategies can reduce excess inventory, improve supply chain efficiency, and minimize stockouts. Moreover, the recommendation system can also use real-time data to adjust recommendations based on supplier performance and changing demand patterns, reducing costs and environmental impact.
E-commerce: AI recommender systems are used extensively in e-commerce to suggest products of interest to customers based on their browsing history and purchase behavior, history, and preferences to increase sales and improve customer experience and satisfaction.
Healthcare: Recommender systems can suggest medications and treatments to patients based on their symptoms and medical history. These systems analyze the patient medical history data to identify trends and patterns and recommend potentially effective treatments. Such personalized treatment recommendations can reduce healthcare costs and improve patient outcomes.
Media and Entertainment: Recommender systems can make personalized content recommendations/suggest content such as TV shows, music, and movies to users based on their behavior, preferences, and viewing and listening history to improve user retention and engagement.
References and Further Reading
- Zhang, Q., Lu, J., Jin, Y. (2021). Artificial intelligence in recommender systems. Complex & Intelligent Systems, 7, 439-457. https://doi.org/10.1007/s40747-020-00212-w
- Pal, Kaushik. “How Is AI Used in Recommendation Systems?” Techopedia, 29 Mar. 2023, https://www.techopedia.com/how-is-ai-used-in-recommendation-systems
- Convergence, I. T. “Top Use Cases of AI-Based Recommendation Systems.” IT Convergence, 8 Mar. 2023, https://www.itconvergence.com/blog/top-use-cases-of-ai-based-recommendation-systems.