You are on page 1of 5

Closing The Gap: Strategic Approaches to Address

Research Gaps in Social Media Recommendation


System
Abhishek Morya Urvashi Gupta Pawan Sharma
School of Computer Applications School of Computer Applications School of Computer Applications
Manav Rachna International Institute Manav Rachna International Institute Manav Rachna International Institute
of Research & Studies of Research & Studies of Research & Studies
Faridabad, Haryana Faridabad, Haryana Faridabad, Haryana
abhimorya8826@gmail.com guptaurvashi588@gmail.com pawansharma2632001@gmail.com

Priyanka Sharma
School of Computer Applications
Manav Rachna International Institute
of Research & Studies
Faridabad, Haryana
priyanka.sca@mriu.edu.in

Abstract— Social media networks are now an inseparable The introduction of new ways of content discovery
component of life, serving as a place for information exchange, usually features popular bias [1] or the bubble effect [2],
contact interaction, and audio-visual content. Recommendation which may distort the participants views and provide only a
systems on these kinds of platforms, like social media skewed representation of content. Moreover, the cold start
platforms, are the key players in this process, helping users get problem [3], which is a situation in which new users or items
content that they like and are interested in. However, the do not have enough data for precise recommendations, is a
recommendation systems that are in place these days tend to problem that leads to incorrect and irrelevant suggestions and
add an inherent bias; this, in addition to the "cold-start" drawbacks. Moreover, circumstance insensitivity in the
problem and lack of context awareness, results in poor
majority of the recommendation algorithms [4] poses a
recommendations and disgruntled users. Here, the object of
the research is to tackle the gaps left in the process by
challenge to an automated system that often overlooks the
suggesting the most effective and relevant strategies for social factors that define user preferences.
media recommendation system development. We will conduct This work will be done in order to identify and fill the
an exhaustive literature review to pinpoint the crucial research area gaps that still need solutions by offering
constraints and suggest differentiable options, such as bias practical approaches on how they can enhance social media
mitigation approaches, contextual recommendation models, or recommendation systems’ performance and effectiveness.
even hybrid models, which contain various recommendation
We deliver the extensive literature review, the processes of
strategies. However, the practical implementation and
identification of the major issues, and the environmental
evaluation of these methods are also worth mentioning, the
reason being that they may boost recommendation accuracy
aspects of the alternatives with the help of unbiased methods,
and user experience. Both the results and the discussion bring context-aware recommendation models, and hybrid
forth valuable matters related to the usefulness and future approaches that include various recommendation
directions of study in this domain. algorithms. Also, we present a proof-of-concept of our
proposed methods, the effectiveness of which is explored in
Keywords—social media, recommendation system, hybrid an empirical evaluation to highlight their capacity to make
models, user experience. recommendations more credible and satisfying to a user.
I. INTRODUCTION II. LITERATURE REVIEW
Social media has changed the way we use and interact These significant contributions of Abdollahpouri et al. [1,
with information by democratising it and changing how it is 15] towards embedded swing in bias and predictive models
accessed. Media platforms buzzing with billions of users are essential not only for the machine learning community
around the globe produce an endless flow of information, so but also for artificial intelligence. In [1], they demonstrated
it becomes ambiguous for users to find something relevant diversity-enhancing regularisation approaches that aid in the
and of quality. As personalized content delivery has gained reduction of popularity bias and promote content diversity.
ground in response to the challenge of information overload, Their [15] act as a targeted literature review that captures and
recommendation systems have made their appearance by outlines the multi-stakeholder recommendation systems
using machine learning and data mining technologies that together with the function of taking into account different
adjust to the individual user personas and interests. actors perspectives and preferences.
However, even with the remarkable progress in Nguyen and team have done the study about filter
interpretability and personalisation, research on bubbles being emanated by recommender systems through
recommendation systems is still suffering from many voids which users' attention is restricted to particular content that
and shortages, which are the main reasons why social media compliments their current taste. Their research shows that
recommendations tend to be ineffective. recommendation systems in fact impact the variety of
content available, which reduces the way divergent views are B. Cold Start Problem
revealed, thus making the need for de-biassing and diverse- The user or item cold-start problem [3] has long been a
boosting techniques obvious. debated issue, but it goes directly to the idea that these new
Coming from a clean slate [3] was therefore the primary users or items are simply not equipped with sufficient data
obstacle that Son had to address, a persistent problem that points to allow for accurate recommendations. Historical
recommendation systems often face. [3] were devoted to an user-item interactions of classical collaborative filtering
introductory review of the recommended options for techniques are at the core of it, so in the cold start, why
processing new user and new item cold-start problems by should we trust this? The scope of the problem is particularly
means of technique comparison and by employing content- extensive in the context of social media, where new users
based filtering, cross-domain recommendation, and hybrid and contents are continuously popping up.
approaches. The merit of using a variety of strategies for C. Lack of Context Awareness
making recommendations was emphasised by their work,
which enables them to reduce the aggravation of the cold- Most of the conventional recommendation models debate
start issue. whether they go quite well with the contextual factors that
ultimately influence user preference and behavior. Context-
Adomavicius and Tuzhilin [4] changed the way aware recommendation systems [4] are now trying to
recommendations were done when they created context- incorporate situational, temporal, and environmental data
aware recommenders. One of the groundbreaking works of into the recommendation process, which results in
this research [4] has been the introduction of the idea of recommending actions that depend on the user's current
using contextual information, e.g., time, location, and context. Whereas making and giving all the right contextual
situation, to enhance the recommendation process. They information is the big deal in social media recommendation
groomed situation-based recommendation models and systems, it is one of the shortcomings of those systems.
techniques; therefore, the recommendation will be
personalised and appropriated for the different IV. PROPOSED APPROACHES
circumstances.
A. Context-Aware Recommendation Models
Research by Wang et al. [5] concentrated on increasing To this end, we are proposing the development of
the performance of recommendation systems, which is context-made recommendation systems that will include
explainable. In [5], they describe the Tree-Enhanced context in the recommendation algorithm. Such a sense of
Embedding Model (TEM) as the thing that allows purpose can be fulfilled with the framework of contextual
hierarchical representations of items and the interpretability information in the recommendation model, like time and
of recommendations that can be generated based on such location, and also with the use of device features [7].
representations. The principals of the two started to strive for Another strategy that can be implemented is the emergence
user trust at the stage of generating recommendation of attention models [8], which are useful in adjusting the
decisions and also increased user understanding of what it model so that it reflects the importance of the contextual
was based on, addressing the limited transparency in many factors to the user at that given time.
existing recommendation models.
B. Hybrid Recommendation Models
III. RESEARCH GAPS To deal with the cold-start problems and the
A. Bias in Recommendation Systems recommendation, we came up with the hybrid
recommendation method, which mixes diverse
The fact that recommendation systems usually have recommendation mechanisms and not only adheres to every
biases that can suggest a skewed and possibly harmful view person's preferences but is also able to offer inspiration. For
of people is noteworthy. Popularity bias occurs when popular example, content-based filtering [9] is persuasive in that it
content gets recommended more than obscure or not-so- helps new users start by generating recommendations of new
popular pieces of content. Thanks to this phenomenon, the products or properties they would not have otherwise visited.
so-called "rich-get-richer" mechanism works, and limited However, collaborate filtering [10] can also be useful for
diversity of content and small chances of "hitting the nail on existing users and things that are accompanied by the
the head" occur. On top of that, people´s habits can often be required number of interactions. Furthermore, advancing the
predicted and reinforced by recommender systems as users technologies that recommend products based on knowledge
are pushed to narrow their outlook. (ontologies, linked data sources) and feeding them with this
knowledge can also facilitate the improvement of
recommendations and address cold start conditions.
V. IMPLEMENTATION AND EVALUATION
For the purpose of proving the feasibility and generic
ability of our approaches, the implementation of a proof-of-
concept social media recommendation system leading to bias
mitigation techniques, context-aware modelling, and hybrid
recommendation strategies was carried out. We tested the
system using an actual dataset from one of the main social
networking platforms, including a comparison to a baseline
recommendation algorithm.

Fig. 1. Gender Bias [6].


A. Dataset and Experimental Setup such as time, position, and device, the recommendations
We successfully used a popular social media platform to corresponded to the user’s specific situation and his or her
publicly post data, including users, item interactions, expedient choice at a particular place. Furthermore, the
additions of context (e.g., times, locations), and added mechanism of attention achieved was to pick, highlight, and
features (text, images). The dataset was segmented into give high priority to the most notable contextual factors, and
training and testing sets, and data cleaning was performed to this substantially improved recommendation performance.
make this data cleaner for the machine to process with C. Hybrid Recommendations Strategies
accuracy [11].
The hybrid recommendation systems, which combine the
B. Evaluation Metrics content-based filtering methodology, the collaborative
To assess the grade of performance of our filtering strategy, and knowledge-based techniques, appear to
recommendation system, we have used such performance perform better in terms of handling a cold start problem and
metrics as precision, recall, and normalised discounted as regards the precision of the recommendations for both
cumulative gain (NDCG) [12]. Such metrics indicate well-established and user items. Our model facilitates this by
whether the recommended information included in the interweaving the strengths of different recommendation
system is useful from various aspects, translating to how the strategies and therefore enables it to deliver dependable
recommendation system performs in providing services. recommendations in situations with limited interaction data,
which will in turn improve its overall recommendation
VI. RESULTS AND DISCUSSION quality as users enjoy a better user experience.
A. Bias Mitigation and Diversity VII. FUTURE DIRECTIONS
Bias mitigation strategies such as diversity-enhancing It shows that the adopted strategies can yield reasonable
regularisation and adversarial debiasing, which we have results; therefore, the solution to the existing problems in the
tested, belong to the best ones as they garner much more field of social media recommendation systems can be
diversified embeddings relative to baseline approaches. discovered. Through this process of bias mitigation,
Through the use of such tactics, a vast content of different contextual awareness, and then a hybrid recommendation
features was put into consideration that not only corrected strategy, we improved the quality, diversity, and customer
popularity bias but also prevented filter bubble effects. satisfaction of the recommendation system.
Table 1. provides the number of diversity scores [13] for Despite the fact that there are some challenges still left to
various recommendation approaches, which in turn indicate overcome and research prospects in these fields, humanising
the good performance of the proposed approaches in content studies are also on the new stage. One of the areas we can
improvement. explore is showing the complex way of natural language
processing in which the user's preferences as well as content
TABLE I. NUMBER OF DIVERSITY SCORES. semantics are analyzed. Furthermore, incorporating into the
Approach Diversity Score model explainable AI algorithms [14] can help to establish a
better understanding, and a greater likelihood of trust, of
Baseline 0.42 recommendation systems empowering users in decision-
making.
Diversity-Enhancing
0.68
Regularization On the other hand, recommendation services should
accommodate and incorporate the evolving and presenting
Adversarial Debiasing 0.71
aspects of social media platforms (for instance, live
streaming, virtual reality, and augmented reality), which will
require a modification of algorithms [15].
Diversity Score VIII. FUTURE WORK AND LIMITATIONS
0.8 Our proposed techniques, however, have proven to close
0.68 0.71 some of the existing gaps in researching recommendation
0.6
systems for social media networks, though there are still
0.4 0.42 some drawbacks and openings for further research. One
0.2 weakness in these suggestions is their computational
0 complexity, together with the resource deficit that some of
the techniques, including adversarial debiasing and attention
l…
-…
e

mechanisms, face. Strategies for efficient implementation


ia
ity
lin

ar
rs
se

rs

and optimisation can be developed to realise scalability and


ve
Ba

ve
Di

real-time response on large social media network platforms


Ad

[16].
Fig. 2. Number of diversity scores in graphical form.
And further evaluation of the of the usage of one dataset
B. Context-Aware Recommendations from an ordinary social app, which may not be
Our models that are context-aware, which we developed comprehensive enough to show the diversity and complexity
by using context features and attention mechanisms, gave of different platforms and user behaviors, It will be
better recommendation results as compared to context- appropriate to have more studies and implementations on
agnostic benchmarks in terms of recommendation quality various datasets and real-world applications for the purpose
and user satisfaction. Through applying specific conditions,
of enhancing the tests of the approaches’ robustness and REFERENCES
generalizability [17]. [1] Abdollahpouri, H., Burke, R., & Mobasher, B. (2017). Controlling
popularity bias in learning-to-rank recommendation. In Proceedings
Again, this problem can exist, coupled with the fact that of the Eleventh ACM Conference on Recommender Systems (pp. 42-
user tastes are constantly changing and the increasingly 46).
complex social media platforms are evolving. The novelty [2] Nguyen, T. T., Hui, P. M., Harper, F. M., Terveen, L., & Konstan, J.
characteristics of the feature, modality, and behaviour will A. (2014). Exploring the filter bubble: the effect of using
constantly call for the innovation and development of recommender systems on content diversity. In Proceedings of the
recommendation systems. Constant research and 23rd international conference on World wide web (pp. 677-686).
development initiatives will have to follow these changes in [3] Son, L. H. (2016). Dealing with the new user cold-start problem in
recommender systems: A comparative review. Information Systems,
order to remain current in the field and judge the worth of 58, 87-104.
recommendations [18].
[4] Adomavicius, G., & Tuzhilin, A. (2011). Context-aware
However, further work is needed and may include recommender systems. In Recommender systems handbook (pp. 217-
253). Springer, Boston, MA.
research on the knowledge of advanced natural language
processing methods for better remembering and showing the [5] Wang, X., He, X., Feng, F., Nie, L., & Chua, T. S. (2018). TEM:
Tree-enhanced embedding model for explainable recommendation. In
semantics of both users' preferences and content. Thus, the Proceedings of the 2018 World Wide Web Conference (pp. 1543-
generated recommendations will be appropriately 1552).
personalised and contextual. Further, the implementation of [6] M. (2023, April 14). Gender Bias in Recommender Systems.
explainable AI strategies [14, 19] may well make neuron.ai. https://neuron-ai.at/gender-bias-in-recommender-systems/
recommendation systems more open and permissible for [7] Zhu, X., Geng, X., & Zhang, H. (2021). Reducing popularity bias in
appearance, further inspiring trust from users and visibility recommendation systems using entropy regularizers. Applied
for decision-making [20]. Intelligence, 51(7), 4624-4638.
[8] Zhang, B. H., Lemoine, B., & Mitchell, M. (2018). Mitigating
On top of that, as social media platforms keep on unwanted biases with adversarial learning. In Proceedings of the 2018
advancing and integrating new features and ways (for AAAI/ACM Conference on AI, Ethics, and Society (pp. 335-340).
example, live streaming of virtual or augmented reality), [9] Dai, W., Chen, G., Liu, Y., Li, W., Yu, J., & Ding, Y. (2021).
recommendation systems will have to constantly update and Context-aware recommendation with spatial-temporal multi-modal
behavior modeling. In Proceedings of the 44th International ACM
accommodate those changes. Designing recommendation SIGIR Conference on Research and Development in Information
approaches that can deal with data of various modalities, Retrieval (pp. 1146-1156).
such as text, videos, and images, will be considered highly [10] Kang, W. C., & McAuley, J. (2018). Self-attentive sequential
significant tasks for providing high-quality and engaging recommendation. In 2018 IEEE International Conference on Data
recommendations [21]. Mining (ICDM) (pp. 197-206).
[11] Lops, P., De Gemmis, M., & Semeraro, G. (2011). Content-based
IX. CONCLUSION recommender systems: State of the art and trends. In Recommender
systems handbook (pp. 73-105). Springer, Boston, MA.
This paper discussed the major gaps in social media
[12] Sarwar, B., Karypis, G., Konstan, J., & Riedl, J. (2001). Item-based
recommendation systems and gave a strategic framework for collaborative filtering recommendation algorithms. In Proceedings of
addressing these shortcomings. Some of the measures the 10th international conference on World Wide Web (pp. 285-295).
proposed include fighting biases, incorporating context [13] Musto, C., Narducci, F., Lops, P., Semeraro, G., & De Gemmis, M.
awareness, and using hybrid recommendation (2019). Linked open data-enabled knowledge-aware recommender
strategies. After a thorough literature review, we pinpointed systems. In Semantic Web Challenges (pp. 117-140). Springer, Cham.
the disadvantages of the existing recommendation systems [14] Järvelin, K., & Kekäläinen, J. (2002). Cumulated gain-based
that have an influential sign, and we offered more effective evaluation of IR techniques. ACM Transactions on Information
Systems (TOIS), 20(4), 422-446.
issues, specifically to improve the quality of helpfulness and
[15] Ziegler, C. N., McNee, S. M., Konstan, J. A., & Lausen, G. (2005).
support of recommendation systems as well as the users' Improving recommendation lists through topic diversification. In
experience of them. Proceedings of the 14th international conference on World Wide Web
(pp. 22-32).
The implementation and empirical evaluation of the
[16] Zhang, Y., & Chen, X. (2020). Explainable recommendation: A
proposed methods not only demonstrated the feasibility of survey and new perspectives. Foundations and Trends® in
the approach but also illustrated benchmarks in the areas of Information Retrieval, 14(1), 1-101.
recommendation diversity, context-aware modelled [17] Abdollahpouri, H., Adomavicius, G., Burke, R., Guy, I., Jannach, D.,
recommendation relevance, and cold-start scenarios through Kamishima, T., ... & Ricci, F. (2020). Multistakeholder
hybrid recommendation strategies. recommendation: Survey and research directions. User Modeling and
User-Adapted Interaction, 30(1), 127-158.
However, though the work is impressive and has enjoyed [18] Deldjoo, Y., Elahi, M., Quadrana, M., & Cremonesi, P. (2021).
great successes, there are undoubtedly also emerging Generating Synthetic Data for Enhancing Recommender Systems.
challenges, and the field needs more research. Resilient arXiv preprint arXiv:2109.07487.
efforts will be needed to ensure an ongoing number of [19] Schedl, M., Zamani, H., Chen, C. W., Deldjoo, Y., & Elahi, M.
interventions that keep pace with the changing social media (2018). Current challenges and visions in music recommender
systems research. International Journal of Multimedia Information
landscape of platforms, unaided modality, and well-refined Retrieval, 7(2), 95-116.
user preferences.
[20] Zeng, W., Shang, J., Chen, G., Wu, J., Li, Y., Zhang, J., & Liang, X.
The path to social media recommendation systems that (2022). Towards Reinforcement Recommendation by Reinforcing
User Interaction History. WWW '22.
can inform, involve, and bring an actual benefit to users can
[21] Zheng, L., Lu, C. T., Jiang, F., Zhang, J., & Yu, P. S. (2018). Spectral
be opened gradually by further developing the concepts of co-clustering on multi-type relational data. In Proceedings of the 2018
recommendation system research outlined here as well as World Wide Web Conference (pp. 1515-1524).
taking intentional steps to the solutions described to the
challenges.

You might also like