Search engines and recommendation engines are systems that help users find information and discover items of interest in large data spaces. A search engine enables users to retrieve relevant information in response to a query (for example, using keywords to find webpages or documents). In contrast, a recommendation engine (or recommender system) suggests items to users based on their observed behavior or preferences, even when the user has not explicitly queried for those items. Both types of engines leverage algorithms and data to filter and rank results. These technologies have become essential in the digital age, powering information retrieval and personalization on almost every major platform we use.
The importance of search and recommender systems spans across industries and applications[1]. In e-commerce, they drive product discovery and sales; in media and entertainment, they personalize content feeds; in healthcare, they assist in information retrieval and personalized care recommendations. By helping users navigate vast choices, these systems address the problem of information overload and enhance user experience. As a result, many businesses credit these engines for increased user engagement and revenue uplift, with personalized recommendations alone reported to raise revenues by 5–15% in some cases. Given their broad impact, significant research has been devoted to improving the algorithms that power search and recommendation engines, as discussed in the following sections.
Machine Learning Techniques
Collaborative Filtering
Collaborative filtering is one of the most widely used techniques in recommender systems. It operates on the principle that users with similar preferences in the past will prefer similar items in the future. There are two primary approaches: user-based filtering and item-based filtering. User-based collaborative filtering finds users (“neighbors”) similar to the target user and recommends items those neighbors liked. Item-based collaborative filtering instead finds items similar to those the target user has liked, and recommends those items to the user. Both approaches can be implemented with memory-based methods (using the entire user-item dataset to find similarities) or model-based methods (learning a predictive model, such as matrix factorization). A key advantage of collaborative filtering is that it requires no domain knowledge of item content; it can recommend complex items (movies, music, etc.) by relying solely on user interaction data. However, it faces challenges like the cold start problem (new users or items with no ratings), scalability to large datasets, and sparsity of data (most users only rate a few items). Many large-scale systems use collaborative filtering; for example, Amazon’s famous “users who bought X also bought Y” recommendations are based on item-item collaborative filtering[2].
Content-Based Filtering
Content-based filtering recommends items similar to those a user liked in the past, based on features of the items. It relies on item descriptions and user profiles. For each user, the system builds a profile of preferences by analyzing the attributes of items the user has rated positively.
For example, a content-based movie recommender might learn that a given user likes action films with a certain actor, and then recommend other movies with similar characteristics[3]. The methodology often involves computing item similarity using features such as keywords, categories, or other metadata. Techniques like TF-IDF (term frequency–inverse document frequency) and cosine similarity are commonly used to compare text-based item profiles in domains like news or books. Content-based methods are effective when item content can be meaningfully analyzed (e.g. text of an article, product attributes) and they naturally handle the new item cold start (since recommendations can be made from item features even without user ratings).
However, a well-known limitation is that they can lack diversity – the user may only be recommended very similar items to those already seen, a phenomenon called overspecialization. They also require good feature representation of items, which might be difficult for abstract items (e.g., how to quantify the qualities of a song or a piece of art). Despite these challenges, content-based recommenders are widely used in domains like news recommendation and information filtering, often in combination with other methods[3].
Deep Learning Approaches
With advances in machine learning, deep learning techniques have been increasingly applied to search and recommendation problems. Deep neural networks can model complex user-item interaction patterns and learn latent representations (embeddings) for users and items. In recommendation engines, deep learning enables modeling of nonlinear relationships and sequential user behavior. For example, neural collaborative filtering replaces traditional matrix factorization with multi-layer neural networks to learn user–item interactions [4].
Recurrent Neural Networks (RNNs) and Transformers are used for sequence-aware recommendations, capturing the order in which a user interacts with items (such as the sequence of songs a user listens to) to predict the next likely item[4]. Embeddings learned via neural networks condense users’ histories and item attributes into dense vector representations in a shared space, allowing similarity to be computed in a learned latent space. In search engines, deep learning improves result ranking and query understanding; models like BERT (Bidirectional Encoder Representations from Transformers) are employed to better interpret query context and match it with documents, significantly improving relevance for natural language queries. Overall, deep learning approaches have achieved state-of-the-art performance in many recommendation tasks by leveraging large datasets to capture subtle and complex patterns[4]. They power industry systems like YouTube’s deep neural network-based recommendation pipeline, which consists of neural networks for candidate generation and ranking to serve personalized video suggestions at scale.
Reinforcement Learning
Reinforcement learning (RL) techniques model the recommendation task as a sequential decision-making process and have gained popularity for certain recommendation and search scenarios [5]. In an RL-based recommender system, the system is an “agent” that interacts with the environment (the user and their choices) and learns a policy for recommending items that maximizes some notion of cumulative reward (e.g., long-term user engagement). Each recommendation can be seen as an action that leads to a reward signal (such as the user clicking an item, watching a video, or making a purchase).
Unlike standard one-shot recommendations, RL considers the long-term effect of recommendations – for instance, showing a diverse item now might keep a user engaged longer, even if it isn’t the most immediately relevant suggestion. Techniques such as multi-armed bandits (for context-aware or session-based recommendations) and deep Q-learning or policy gradient methods (for learning recommendation policies) have been applied in this area. An example application is in news recommendation, where an RL algorithm can continuously adapt to a user’s evolving interests to maximize the total clicks or reading time in a session. RL helps address challenges like balancing exploration (showing novel or less certain recommendations) and exploitation (showing known relevant items) to avoid filter bubbles and improve recommendation diversity.
However, applying RL in recommender systems also poses challenges, such as defining the right reward function and dealing with sparse feedback. Ongoing research is integrating RL with other techniques (e.g., using collaborative filtering as part of the state or reward mechanism) to enhance recommendation quality[5].
Hybrid Models
No single recommendation technique is optimal for all scenarios, so many real-world systems use hybrid models that combine multiple approaches. Hybrid recommender systems integrate two or more techniques (collaborative, content-based, knowledge-based, etc.) to capitalize on their complementary strengths and mitigate their weaknesses[6]. For example, a hybrid system might use content-based filtering to handle new items (cold start) by recommending items similar in content to a user’s likes, while simultaneously using collaborative filtering for items that have sufficient user feedback. The hybridization can occur in various ways: combining the outputs of different recommenders by weighted averaging, switching to a different algorithm in certain situations, or integrating multiple sources of data into one unified model. A well-known instance of a hybrid approach was the Netflix Prize solution, which ensembled hundreds of variants of collaborative filtering and other models to improve accuracy.
Even Netflix’s actual production system has historically used a mix of algorithms – combining collaborative filtering (to learn from user behavior patterns) with content-based methods (to incorporate metadata like genre or cast) and even rule-based business logic. Similarly, search engines might use hybrid techniques; for example, an e-commerce search engine could blend text-based search results with personalized recommendations on the results page. By using hybrid models, systems aim to provide more robust and accurate results, handling a wider range of scenarios than any single algorithm alone[6].
Industry Applications
E-commerce
In online retail, search and recommendation engines play a central role in connecting customers with products. An e-commerce search engine allows users to find specific items by typing queries (product names, categories, or attributes). The quality of search results can directly impact sales – if users cannot find what they want quickly, they may leave the site. On the other hand, a recommendation engine in e-commerce suggests products that users might be interested in, based on their browsing history, past purchases, or similarities to other customers. These personalized recommendations drive product discovery and cross-selling. For instance, Amazon’s recommendation system accounts for a significant portion of its sales by suggesting related or popular items (“Customers who viewed this item also viewed…”). Amazon famously implemented an item-to-item collaborative filtering algorithm that can generate recommendations in real-time and scale to millions of customers and products[7]. This capability keeps users engaged and increases the average order value by surfacing relevant items they might not have found via search alone. Beyond Amazon, virtually all major online retailers (e.g., eBay, Alibaba, Walmart) rely on a blend of search and recommender systems – search to satisfy explicit product queries, and recommendations to provide a personalized shopping experience and encourage impulse discoveries. The synergy of search and recommendation in e-commerce improves user satisfaction and loyalty, as customers are served both the items they explicitly ask for and those they are likely to want but didn’t know to ask.
Healthcare
The healthcare industry has begun adopting search and recommendation technologies to personalize patient care and information delivery. Healthcare recommendation systems can suggest personalized health interventions, wellness tips, or medical resources to patients based on their individual profiles and needs. For example, a health app might recommend diet plans or exercise programs tailored to a user’s medical history and fitness level. These systems often draw on medical domain knowledge and user data (symptoms, conditions, preferences) to make relevant suggestions. Likewise, clinicians use specialized search engines to retrieve medical literature or patient records quickly, which is critical for informed decision-making. A doctor might use a medical search engine that leverages NLP to interpret complex queries (such as searching for case studies of a rare condition). Personalized health recommenders have been developed for diverse purposes: managing chronic disease by recommending lifestyle adjustments, suggesting preventive care screenings, or even matching patients to clinical trials for which they are eligible. Privacy and accuracy are especially important in this domain – recommendations must be evidence-based and data security is paramount. According to a recent systematic review, many health recommender systems use hybrid algorithms and focus on categories like nutrition advice, general wellness information, and condition-specific management, demonstrating the breadth of applications from diet and fitness to personalized treatment information[8]. Though still an emerging field, the use of search and recommendation engines in healthcare promises to improve patient engagement and health outcomes by delivering the right information to the right person at the right time.
Media Streaming
Media streaming services (such as video and music platforms) are renowned for their sophisticated recommendation engines. These systems are key to engaging users by automatically playing or suggesting content aligned with the user’s tastes. YouTube and Netflix are classic case studies: YouTube’s recommendation engine uses deep learning to analyze a user’s watch history and the content of billions of videos, in order to suggest videos that maximize watch time and satisfaction. Netflix similarly employs a multi-faceted recommender system that takes into account your viewing history, ratings, and even time-of-day patterns to present movies and TV shows you’re likely to watch next. Spotify, a music streaming service, provides another perspective – it combines collaborative filtering (analyzing which songs are liked by users with similar tastes) with content-based analysis (e.g., audio feature analysis and natural language processing on song metadata and lyrics) to create personalized playlists and song suggestions. These platforms often integrate search as well, allowing users to explicitly look up a title or artist; the search results can be personalized or influenced by the user’s preferences. The success of streaming services is tightly coupled with their recommendation quality – for instance, Netflix reported that a large portion of viewing is driven by personalized recommendations rather than direct searches. By leveraging machine learning at scale, media streaming companies are able to keep users engaged for hours, deliver a highly personalized content feed, and surface the vast long-tail of content effectively[9].
Enterprise Solutions
Beyond consumer-facing applications, search and recommendation engines are vital in enterprise environments for knowledge management and decision support. Enterprise search systems enable employees to retrieve information from internal data sources — such as documents, emails, databases, and intranets — analogous to how a web search engine retrieves information from the internet. These systems face unique challenges: enterprise data is often unstructured and spread across many repositories, with access controls and domain-specific terminology. Effective enterprise search uses AI to index and understand content (including using natural language processing to interpret queries and documents) so that employees can quickly find relevant information, experts within the organization, or past projects related to their query. In addition, organizations deploy recommender systems internally, for example to recommend relevant documents or learning materials to staff, suggest collaborators who have expertise in a needed area, or recommend products to cross-sell in a sales context. By integrating recommendation functionality, a knowledge management portal might proactively suggest content to a user based on their role, past activity, or interests (e.g., recommending a new policy document to all managers, or suggesting troubleshooting articles to a customer support agent based on the cases they are handling). Enterprise solutions thus blend search and recommendation techniques to boost productivity and ensure that valuable knowledge is shared and utilized. However, implementing such systems can be difficult due to the complexity of enterprise data. Research has highlighted challenges like indexing data from heterogeneous sources, respecting security/privacy constraints, and tuning relevance for enterprise users. Nonetheless, progress is being made — for instance, enterprise search engines now use techniques like semantic search and question-answering (using advanced NLP) to improve results. Overall, AI-powered search and recommender systems are increasingly seen as essential tools for effective knowledge management in businesses[10].
Recent Advancements
Large Language Models (LLMs)
Large Language Models, such as BERT and GPT, have recently been integrated into search and recommendation engines to enhance their capabilities. LLMs are trained on vast amounts of text data and can understand language nuances, which makes them powerful for interpreting queries and content. In search engines, models like BERT (introduced by Google into its search ranking in 2019) enable a deeper understanding of query context and user intent, leading to more relevant search results especially for natural language queries. For example, BERT can help distinguish meanings of words based on context, so that a search for “2019 brazil traveler to USA need visa” correctly interprets the intent (whether a Brazilian traveling to the US needs a visa) and returns correct information. This represents a jump from keyword matching to true semantic understanding in search. In recommendation systems, transformer-based models (the architecture behind BERT/GPT) are used to capture sequential patterns and contextual relationships in user behavior. A notable approach is BERT4Rec, which applies a bidirectional transformer to model a user’s entire sequence of past interactions and predict the next likely item, achieving strong results in sequential recommendation tasks[11]. Similarly, generative models like GPT are being explored to produce personalized content or recommendations via conversational agents. For instance, an LLM-powered recommender could generate a natural language explanation for a movie suggestion (“You might like this film because it’s directed by Christopher Nolan, whose movies you enjoyed”) or even engage in a dialogue with the user to refine recommendations. The integration of LLMs is still evolving, but early results show improved recommendation novelty and better understanding of complex user queries/preferences. These models, however, come with challenges like high computational cost and the need to align generated outputs with factual correctness and business goals.
Retrieval-Augmented Generation (RAG)
Retrieval-Augmented Generation is a cutting-edge technique that marries the strengths of search engines with generative AI models [12]. The idea behind RAG is to improve the quality and accuracy of generated text (from models like GPT) by grounding it in retrieved knowledge. In a RAG pipeline, when a user poses a query or prompt, the system first uses a search component to fetch relevant documents or passages from a knowledge base. These retrieved results are then provided as context to a generative model, which uses them to produce a final answer or recommendation. This approach ensures that the generative model has access to up-to-date and factual information, rather than relying solely on what was stored in its parameters during training. RAG has notable applications in domains like question-answering, customer support, and personal assistants.
For example, a customer support chatbot might retrieve relevant product manuals or policy documents and then formulate a natural language answer for the customer, combining search precision with fluent generation. In the context of recommendation, RAG can be used to generate richer item descriptions or personalized newsletters by pulling in facts about items from a database before writing a recommendation paragraph. It can also support explainable recommendations, where the system generates an explanation for a recommendation by quoting retrieved evidence (e.g., “We suggest this book because it was highly rated in NY Times [12]”). By linking a language model with a search index, RAG systems continuously update their knowledge and reduce the risk of generating irrelevant or outdated information. This approach represents a significant advancement in building AI systems that are both knowledgeable and articulate, and it blurs the line between traditional search engines and AI assistants.
Multimodal AI
Traditionally, search and recommendation engines dealt mostly with textual data (queries, document text, etc.) or structured data (user ratings). Multimodal AI refers to systems that incorporate multiple types of data — text, images, audio, video, etc. — to provide more holistic search and recommendation experiences. Modern recommendation engines increasingly use multimodal signals to improve accuracy. For instance, a fashion recommendation system can analyze product images (to understand color and style) in addition to user purchase history and text reviews.
A music recommendation engine like Spotify may consider the acoustic characteristics of a track (audio signal analysis) alongside user listening patterns and text metadata, to recommend songs that “sound” similar to what the user likes. Similarly, video streaming platforms consider visual features from thumbnails or the video content itself and audio transcripts when recommending videos (for example, ensuring recommended videos have similar style or addressing similar topics). On the search side, multimodal search allows users to query by an image or voice in addition to text, and the engine must then match that query against a multimodal index. An example is Google Lens, where a user can search the web using a photograph and get results that include similar images or related information.
Incorporating multiple modalities can significantly enhance the user experience: consider a travel website that recommends destinations by analyzing both the text of reviews (for sentiment and described attractions) and the photographs of those destinations (for scenery and points of interest). Multimodal recommendation models use deep learning to fuse features from different modalities into a joint representation, handling the heterogeneity of data [13]. This advancement helps systems understand content more deeply (beyond what text alone can convey) and caters to richer user inputs and preferences. It does, however, increase system complexity and requires careful design to weight the contribution of each modality appropriately.
Evaluation Metrics
To assess the performance of search and recommendation engines, a variety of evaluation metrics are used. Common information retrieval metrics and recommender system metrics include:
● Precision: The fraction of retrieved or recommended items that are relevant. For example, if a search engine returns 10 results and 8 of them are relevant to the query, precision is 0.8. Higher precision means fewer false positives in recommendations/search results.
● Recall: The fraction of all relevant items that are retrieved or recommended. For instance, if there are 20 relevant documents in the corpus and the search engine found 10 of them, recall is 0.5. In recommendations, recall measures how many of the items a user would rate positively were actually recommended to them. There is often a trade-off between precision and recall that systems must balance based on the application [14].
● Mean Average Precision (MAP): This combines precision and recall by computing the average precision at various cut-offs for each user or query, and then averaging across all users/queries. MAP is useful for evaluating ranking quality in recommendations or search results, emphasizing that relevant items should be ranked high.
● Mean Reciprocal Rank (MRR): A metric often used in search/Q&A evaluation that looks at the rank position of the first relevant result. It is the average of the reciprocal of that rank for each query. If the first relevant result is usually at position 1, MRR is 1.0 (ideal); if it’s at position 5 on average, MRR is 0.2. MRR is sensitive to how quickly a system returns a good result.
● Click-Through Rate (CTR): In online systems, CTR measures the percentage of recommended items (or search results) that users click on. It’s a direct indicator of user engagement. For example, an e-commerce recommendation might be evaluated by what fraction of recommendation impressions led to clicks or purchases. A higher CTR suggests the recommendations were attractive and relevant to users.
● User Satisfaction and Feedback: Beyond automated metrics, user studies and A/B tests are employed to gauge user satisfaction. This can include explicit ratings from users (after consuming a recommendation, did the user rate it positively?), dwell time (how long the user engaged with the content), or survey responses about their experience. User-centric evaluation is crucial because a system optimized for precision/recall might still fail if it doesn’t actually make users happier with the platform [14].
It’s important to use multiple metrics when evaluating these systems, as each metric captures different aspects of performance. For instance, a recommendation algorithm might have high precision (the user likes most suggested items) but low recall (it misses many items the user would have liked if known). In practice, online systems also continuously evaluate performance through A/B testing, monitoring business KPIs (e.g., conversion rate, revenue lift), and collecting qualitative feedback, ensuring that the search or recommendation engine is meeting its objectives.
Challenges and Future Directions
Despite substantial progress, search and recommendation engines face ongoing challenges and open research questions. Some of the key challenges and future directions include:
Cold Start Problem
This arises when new users or new items enter the system with no historical data. Collaborative filtering algorithms struggle in such cases since they rely on past interactions. Future systems are exploring solutions like better user onboarding (asking preferences initially), integrating content-based features to recommend new items, and using cross-domain data to mitigate cold start issues (for example, using a user’s social media data to inform recommendations when they join a new platform) [15].
Fairness and Bias
Recommender systems and search algorithms can inadvertently amplify biases present in data or create “filter bubbles” that narrow a user’s exposure to information. Ensuring fairness (e.g., recommendations are not discriminating or systematically excluding certain groups or content) is a growing concern. Researchers are developing fairness-aware algorithms that balance relevance with fairness criteria, and techniques to detect and mitigate biases in training data. This also extends to the supplier side in recommendations (for instance, fair exposure of content creators on a platform) [15].
Transparency and Explainability
As these systems become more complex (especially with deep learning and LLMs), their decisions can be opaque. There is a push for explainable AI in recommendations and search – providing users with understandable reasons for results. For example, displaying “Why am I seeing this?” alongside a recommendation can increase user trust. Explanations can be generated by highlighting which past behaviors led to a recommendation or what attributes match the user’s profile. Transparent systems can help users trust and effectively use the engine, and help developers debug and improve algorithms.
Privacy Preservation
Search and recommendation engines often leverage personal data, raising privacy concerns. Stricter data protection regulations and user awareness mean future systems must incorporate privacy-preserving techniques. Approaches like federated learning (where the model is trained across user devices without centralizing raw data) and differential privacy (adding noise to data or models to protect individual information) are being explored to continue improving personalization while respecting user privacy. Achieving a balance between personalization and privacy will be an ongoing challenge.
Multi-objective Optimization
Real-world recommendation and search systems rarely optimize a single metric; they must consider multiple objectives simultaneously. For example, an e-commerce recommender might aim to maximize user satisfaction and conversion rate while also maintaining diversity in shown products and ensuring fairness to vendors. These objectives can conflict (a highly relevant result set might lack diversity; promoting diversity might reduce immediate relevance). Multi-objective optimization and multi-stakeholder recommendation are emerging research areas, where algorithms explicitly account for various goals and constraints. This could involve dynamic algorithms that adjust what is shown not just based on relevance to the user, but also business rules, ethical considerations, or long-term impacts on the user base [15].
In summary, the future of search and recommendation engines will involve making them more adaptable, fair, transparent, and secure. The incorporation of richer context (e.g., location, social context), real-time adaptation to user feedback, and continual learning from streaming data are also active areas of development. By addressing these challenges, next-generation systems will be better positioned to serve both user needs and societal expectations.
Conclusion
Search and recommendation engines have evolved remarkably over the past few decades, becoming indispensable across a range of digital services. Early recommender systems in the mid-1990s were relatively simple, often relying on manual data and straightforward collaborative filtering techniques, and web search engines primarily used keyword matching and basic ranking algorithms. Today, these systems employ sophisticated machine learning techniques – from deep neural networks to reinforcement learning and large language models – to deliver highly relevant and personalized results. We have seen how modern recommenders in domains like e-commerce and streaming media blend multiple algorithms and data sources to serve suggestions that feel intuitive and tailor-made for each user. Likewise, search engines have incorporated semantic understanding and even conversational abilities, far surpassing their original capabilities. This evolution has been driven by both technological advancements and the growing demands of users for better experiences.
Looking ahead, the line between search and recommendation is blurring: search results are becoming more personalized, and recommendation feeds are increasingly capable of handling explicit queries. The integration of generative AI (like RAG systems) hints at a future where search engines might not just find information but also synthesize and explain it in human-like dialogue. Meanwhile, recommender systems are learning to be more context-aware, fair, and transparent. The core goal remains the same as it was at the outset – to help people make sense of large information spaces – but the means to achieve it continue to advance. From the first scholarly discussions of recommender systems in the 1990s to the AI-driven platforms of today, the progress has been profound[16] . In conclusion, search and recommendation engines will continue to converge and improve, leveraging the latest in machine learning to better serve users. As they do so, ongoing research and responsible innovation will be key to ensuring these systems are not only intelligent and effective, but also aligned with human values and needs.
References
[1] Lu, J., Wu, D., Mao, M., Wang, W., & Zhang, G. (2015). Recommender system application developments: A survey. Decision Support Systems, 74, 12–32. https://doi.org/10.1016/j.dss.2015.03.008
[2] Su, X., & Khoshgoftaar, T. M. (2009). A survey of collaborative filtering techniques. Advances in Artificial Intelligence, 2009, 1–19. https://doi.org/10.1155/2009/421425
[3] Lops, P., Gemmis, M. D., & Semeraro, G. (2011). Content-based recommender systems: State of the art and trends. In F. Ricci, L. Rokach, B. Shapira, & P. B. Kantor (Eds.), Recommender Systems Handbook (pp. 73–105). Springer. https://doi.org/10.1007/978-0-387-85820-3_3
[4] Zhang, S., Yao, L., Sun, A., & Tay, Y. (2019). Deep learning based recommender system: A survey and new perspectives. ACM Computing Surveys, 52(1), 1–38. https://doi.org/10.1145/3285029
[5] Zhao, X., Xia, Y., Tang, J., & Zhang, H. (2023). A survey on reinforcement learning for recommender systems. IEEE Transactions on Neural Networks and Learning Systems. https://doi.org/10.1109/TNNLS.2023.3241993
[6] Burke, R. (2002). Hybrid recommender systems: Survey and experiments. User Modeling and User-Adapted Interaction, 12, 331–370. https://doi.org/10.1023/A:1021240730564
[7] Linden, G., Smith, B., & York, J. (2003). Amazon.com recommendations: Item-to-item collaborative filtering. IEEE Internet Computing, 7(1), 76–80. https://doi.org/10.1109/MIC.2003.1167344
[8] Kowatsch, T., & Maass, W. (2010). Health recommender systems: Concepts, requirements, technical basics and challenges. In Proceedings of the 2010 International Conference on Electronic Healthcare (pp. 359–366). Springer. https://doi.org/10.1007/978-3-642-23635-8_45
[9] Gomez-Uribe, C. A., & Hunt, N. (2016). The Netflix recommender system: Algorithms, business value, and innovation. ACM Transactions on Management Information Systems, 6(4), 1–19. https://doi.org/10.1145/2843948
[10] Hawking, D. (2004). Challenges in enterprise search. In Proceedings of the 15th Australasian Database Conference (pp. 15–24). http://crpit.com/confpapers/CRPITV27Hawking.pdf
[11] Sun, F., Liu, J., Wu, J., Pei, C., Lin, X., Ou, W., & Jiang, P. (2019). BERT4Rec: Sequential recommendation via bidirectional encoder representations from Transformers. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (pp. 1441–1450). https://doi.org/10.1145/3357384.3357895
[12] Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., … & Riedel, S. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, 33, 9459–9474. https://proceedings.neurips.cc/paper/2020/file/6b493230205f780e1bc26945df7481e5-Paper.pdf
[13] Zhu, Y., Wang, X., Wu, L., & Li, X. (2022). A comprehensive survey on multimodal recommender systems: Taxonomy, evaluation, and future directions. IEEE Transactions on Multimedia, 24, 1401–1419. https://doi.org/10.1109/TMM.2022.3155974
[14] Herlocker, J. L., Konstan, J. A., Terveen, L. G., & Riedl, J. T. (2004). Evaluating collaborative filtering recommender systems. ACM Transactions on Information Systems, 22(1), 5–53. https://doi.org/10.1145/963770.963772
[15] Koren, Y., Bell, R., & Volinsky, C. (2009). Matrix factorization techniques for recommender systems. Computer, 42(8), 30–37. https://doi.org/10.1109/MC.2009.263[16] Resnick, P., & Varian, H. R. (1997). Recommender systems. Communications of the ACM, 40(3), 56–58. https://doi.org/10.1145/245108.245121
