Analysis of Developments and Challenges in Dealing with

Recommender System Cold-Start Issue

Yue Zhong

Faculty of humanities, Zhujiang College, South China Agricultural University, Guangzhou, 510900, China

Keywords: Cold Start Problem, Meta Learning, Knowledge Graph, Large Language Model.

Abstract: The cold-start problem, caused by a lack of historical interaction data, remains a significant challenge for

recommender systems. This study explores three strategies to address this issue: knowledge graphs, meta-

learning, and large language models. The meta-learning method, Dual enhanced Meta-learning with Adaptive

Task Scheduler (DMATS), improves embedding accuracy and task adaptation through autonomous learning.

The knowledge graph method, Knowledge-Enhanced Graph Learning (KEGL), enhances recommendation

quality using collaborative embeddings and knowledge-enhanced attention. The large language model method,

Automated Dis-entangled Sequential recommendation (AutoDisenSeq-LLM), optimizes recommendations

by leveraging text understanding. These methods were tested on datasets like MovieLens, Amazon, and Yelp.

The meta-learning method demonstrated strong generalization, the knowledge graph method tackled data

sparsity, and the language model method showed potential but needed further improvement. Challenges

include high computational costs, lack of standardized evaluation metrics, and dataset issues. Future research

should focus on novel techniques, datasets, and metrics.

1 INTRODUCTION

Nowadays, a large amount of information fills

people's lives, which makes people's lives more

convenient, but at the same time brings a lot of

redundant information, resulting in users in the state

of information overload. Recommender Systems (RS)

have emerged at this time, RS can help users filter

information based on the user's own preferences,

recommending appropriate information to the user to

solve the problem of information overload.

Nonetheless, there are still issues with RS, and one of

them is the Cold Start Problem (CSP), which is the

phenomena whereby the performance of

recommendations drastically declines due to a lack of

historical interaction data for new users or new objects

(Qian et al., 2024). It is primarily divided into three

categories: user cold start (Panda et al., 2022), item

cold start (Yuan et al., 2023) and system cold start

(Wu et al., 2024). These problems are especially

prominent in e-commerce, short video, news

recommendation and other scenarios. This research

can optimize the user experience in RS, meet the

practical application requirements of e-commerce

Corresponding author

platforms (allowing cold-start items to gain exposure

and bring transaction growth directly), provide

solutions to the core challenges of the current RS and

the solution to the cold-start problem can be migrated

to other dynamic scenarios such as social media.

The three most popular traditional approaches to

recommendation algorithms prior to the complete

adoption of RS in the internet world were hybrid

recommendation algorithms, content filtering-based

recommendation algorithms, and collaborative

filtering-based recommendation algorithms (Panda et

al., 2022). These three approaches are inexpensive and

use straightforward algorithms. However, as RS have

evolved, the algorithms for making recommendations

have been modified iteratively. The inability of these

three conventional approaches to process information

while dealing with CSP has become a challenge for

RS. A number of approaches have been developed in

recent years to address the issue of data sparsity in

CSP, including meta-learning, knowledge mapping,

etc., to tackle CSP with elevation metrics. All of these

methods have more or less solved the problems of low

recommendation quality, low effect, and low diversity

brought by the CSP. However, these methods also

Zhong, Y.

Analysis of Developments and Challenges in Dealing with Recommender System Cold-Start Issue.

DOI: 10.5220/0013990300004916

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Conference on Public Relations and Media Communication (PRMC 2025), pages 289-295

ISBN: 978-989-758-778-8

289

have their drawbacks includes: Scenario

generalization ability, Computational efficiency, Data

dependency, etc. To solve the problem of cross-

domain migration and generalization ability, using

Large Language Models (LLMs) generating user

profiles with text can be the way to improve it (Zhang

et al., 2025).

This study's primary goals are to review and

summarize the state of the CSP research. The

following objectives are the focus of this paper: First,

a summary of the common approaches to solving the

CSP is given. A thorough technical and

methodological examination of CSP solving will

follow this introduction. Second, this study looks at

the evaluation metrics and datasets that are frequently

utilized for the CSP. This analysis helps assess the

current progress in addressing cold-start issues and

evaluates the effectiveness of different approaches in

practical applications. Third, this paper critically

examines the limitations of existing methods, datasets,

and evaluation metrics. By identifying the weaknesses

in current research, areas for improvement are

highlighted. Finally, this paper discusses potential

solutions to the current problems in the CSP and

suggesting future research directions. This paper aims

to advance the development and enhancement of RS

by suggesting new research directions and possible

techniques.

2 METHODOLOGY

This paper summarized the existing methods on

solving CSP mainly including: meta-learning,

knowledge mapping, machine learning, deep learning,

graphical neural networks, LLMs, and so on. In this

chapter, meta-learning, knowledge graphs, and LLMs

are selected as representative methods to be

introduced in terms of principles as well as

motivation after an overview of existing methods, and

in the next chapter, the performance and limitations

of the three methods are compared with, and the

limitations of the current evaluation metrics and

datasets are also analyzed, so as to provide

suggestions and directions for the construction of new

datasets and new evaluation methods in the future.

The processing is shown in the Figure 1.

2.1 Overview

Previous research has developed new models and

techniques to address CSP and increase the accuracy

of RS in this context. These emerging methods have

their own unique perspectives and performances in

dealing with CSP: Meta-learning: meta-learning

methods, as one of the key research method directions

at present. Its core lies in reducing the amount of data

samples needed for learning, which leads to better

performance and faster adaptation to new tasks. This

is advantageous for enhancing recommendation

models' capacity for generalization (WU et al., 2024).

Knowledge Graph: Knowledge Graph method

includes connection-based methods, embedding-

based methods, and propagation-based methods. It

centers on combining user information and various

elements to build a huge knowledge network graph,

thus effectively solving the problem of scarce data and

complex computation (Zhang et al., 2024). LLMs: The

rise of LLMs as the most emerging technology is a

great breakthrough in natural language processing

methods. Using LLMs to understand the context, pre-

training LLMs to model and represent cold-start users

can significantly enhance the recommendation

performance of RS in cold-start situations (Zhang et

al., 2025). The next part of this paper will introduce

these three methods in more detail and describe how

they work and what motivates them.

Alt Text for the figure: A flowchart illustrating the process

of summarizing existing methods, analyzing meta-learning,

knowledge graphs and LLMs, comparing their limitations,

and proposing future datasets and evaluation metrics

improvements.

Figure 1. Overview of methodology (Picture credit :

Original).

PRMC 2025 - International Conference on Public Relations and Media Communication

290

2.1.1 Meta-Learning

Meta-learning, as an important research direction in

the current field of deep learning, is also known as

learn to learn. Traditional machine learning methods

for task-specific training require large amounts of

data support, so in a new domain with little or no data

(i.e., cold start), recommendation performance is often

poor. Compared with traditional machine learning,

meta-learning provides a new learning method, both

through the joint training of multiple tasks to learn

prior knowledge, in the face of new tasks, you can use

the previously learned prior knowledge to quickly

learn, this method significantly enhances the model's

adaptability across diverse tasks, particularly boosting

its generalization ability in RS (WU et al., 2024).

Broadly categorized, meta-learning techniques

can be divided into three main approaches:

optimization-based methods, model-based methods,

and metric-based methods. Metric-based methods and

model-based methods are mainly applied to

classification tasks, while optimization-based

methods perform better in the face of a wider range of

task assignments. Optimization-based methods are

also the most widely used methods in RS, as the Model

Agnostic Meta-Learning (MAML) framework is the

basis of many existing meta-learning methods, which

is also an optimization-based method (Finn et al.,

2017).

2.1.2 Knowledge Graph

In essence, a knowledge network that records the links

between elements is called a knowledge graph. By

structuring the words of the relationships between

entities, and assigning the meaning of information to

each point and each edge in the network, the entire

knowledge graph contains a large amount of

information, so that it can go on to capture potentially

possible connections between individuals, and get the

result of efficient storage and utilization of knowledge

and thus enhance the effectiveness of

recommendations.

Knowledge graphs can also be differentiated in

three types: hybrid approaches, path-based methods,

and embedding-based methods. In item cold start, the

main problem faced by the knowledge graph method

is that the embedding generated in data scarcity is

usually unreliable, leading to the reduction of

recommendation performance, so in item cold start,

the path-based method is often used as an auxiliary

method to raise the quality of recommendations

(Zhang et al., 2025).The knowledge graphs

construction is often based on extracting and

integrating knowledge in different formats from

multiple databases, so that knowledge graphs can be

used in RS to analyze the relationship between the user

and the item through the rich semantic information

they contain, and more accurately predict the user's

interests and needs, thus solving the data sparsity

problem and also giving recommendation results

interpretability, which allows the system to explain the

reason for the recommendation (Zhang et al., 2024).

2.1.3 Large Language Models

LLMs are deep learning-trained generative artificial

intelligences, with excellent text comprehension and

text generation capabilities, and can complete

complex conversations. In recent years, LLMs have

garnered significant attention in the field of RS due to

their effectiveness in natural language processing.

Numerous related models have been developed,

achieving notable results. Leveraging their powerful

text processing capabilities, LLMs are frequently

employed in RS to address cold-start challenges,

including small-sample and zero-sample scenarios.

There are two categories of research on LLMs in

RS: LLMs as RS and LLMs as knowledge enhancers.

LLMs as RS: In this type of research, LLMs are used

directly to generate recommendation results. The

textual descriptions of users and items allow the

LLMs to understand user preferences and item

characteristics, thus generating high-quality

recommendations despite the lack of data.

LLMs as knowledge enhancers: In this type of

research, LLMs are used to enhance the knowledge

representation of RS. The RS acquires richer

knowledge information by pre-training LLMs, which

helps to optimize and enhance the recommender

model and enhance the RS's performance while

dealing with sparse data. (Zhang et al., 2025)

2.2 Datasets and Evaluation Metrics

Description

The evaluation metrics and datasets are important

components of RS research and development, and

they have a significant influence on how

recommender models are constructed and trained. The

datasets used in the 72 papers screened in the 2019-

2023 RS and the evaluation metrics are examined in

SLRRS. Datasets frequently utilized in recommender

systems (RS) research span a wide range of domains

and include well-known examples such as the

MovieLens series (100K, 1M, and 10M), Netflix, and

Yahoo Music for media recommendations. Other

commonly used datasets include FilmTrust, Epinions,

and BookCrossing, which focus on trust-based or

Analysis of Developments and Challenges in Dealing with Recommender System Cold-Start Issue

291

book-related recommendations, as well as

MovieTweetings and Yelp, which cater to social

media and business reviews, respectively.

Additionally, platforms like Ciao and Amazon

provide datasets that are widely adopted for e-

commerce recommendation studies. These datasets

collectively offer diverse scenarios for evaluating RS

models. The evaluation of recommender systems (RS)

relies on a variety of metrics designed to assess

different aspects of performance. Commonly used

metrics include Root Mean Square Error (RMSE) and

Mean Absolute Error (MAE) for measuring prediction

accuracy, as well as Normalized Discounted

Cumulative Gain (NDCG), Precision, Recall, and F1

value for ranking and relevance evaluation. The

majority of evaluation metrics referenced in the

literature for this thesis are based on these established

metrics (Saifudin et al., 2024).

3 RESULTS AND DISCUSSION

This paper analyzes the principles and the effects of

meta-learning, knowledge graph and LLMs in solving

CSP. The principles of the three methods in solving

the CSP have been analyzed in the previous section,

and the next section will summarize the effectiveness

of the methods based on the experimental results of

the related literature:

3.1 Results Analysis

The representative method selected for meta-learning

in this paper is Dual enhanced Meta-learning with

Adaptive Task Scheduler (DMATS) (He et al., 2025)

created by D He et al. The datasets selected for this

method are DBook, MovieLens and Yelp The

evaluation metrics are MAE, RMSE and

NDCG.Table1 shows the performance of DMATS on

the datasets DBook, MovieLens and Yelp in the user's

cold-start problem, as shown in Table 1.

Table 1. Performance of DMATS on three datasets.

Evaluation

index

DBoo

MovieLens Yelp

MAE 0.5653 0.8283 0.8515

RMSE 0.6461 0.9490 1.0004

nDCG@5 0.8807 0.7465 0.8618

Though DMATS did not perform as well as the

other two (user-item cold start, item cold start) in the

user cold start, its performance was still excellent,

reaching either optimality or sub-optimality in the

comparison with the other baseline methods.

Alt Text for the figure: KEGL performs better than baseline methods in Precision and Recall on AB and Yelp.

Figure 2. Performance of KEGL on two datasets.

PRMC 2025 - International Conference on Public Relations and Media Communication

292

Alt Text for the figure: KEGL performs better than baseline methods in HR and NDCG on AB and Yelp.

Figure 3. Performance of KEGL on two datasets.

The representative method selected for

knowledge graphs in this paper is Knowledge-

Enhanced Graph Learning (KEGL) (Zhang et al.,

2025). The datasets selected for this method are

Amazon-book and Yelp2018. the selected evaluation

criteria are Precision, Recall, HR, NDCG. Because

the original article does not show the specific data,

this paper only quotes its comparison picture with the

selected baseline method (Figures 2 and 3). Figure 2

and Figure 3 clearly show the superiority of KEGL

compared with its selected benchmark method in HR,

NDCG, Precision and Recall in Amazon-book and

Yelp2018 in the case of new item complete cold-start.

The representative method selected for LLMs in

this paper is Automated Disentangled Sequential

recommendation with large language models

(AutoDisenSeq-LLM) (Wang et al., 2025). The

datasets selected for this method are MovieLens-1m

(ML-1m), Amazon Beauty (AB), Amazon Game

(AG), and Steam. The selected evaluation criteria are

HR and NDCG. Table 2 shows the recommendation

performance of this method on the four datasets.

Table 2. performance of AutoDisenSeq-LLM on four

datasets.

Evaluation

index

ML-

AB AG Steam

10 0.7125 0.3995 0.6634 0.6605

DCG

10 0.4770 0.2601 0.4277 0.4176

3.2 Discussion

As one of the mainstream cold-start problem solving

methods in recent years, meta-learning methods

perform well in dealing with cold-start problems.

Taking DMATS as an example, the autonomous

embedding learning mechanism and adaptive task

scheduler of DMATS helps the model to learn better

node representations and improve the embedding

accuracy, Consequently, the data sparsity issue is

resolved and the long-tail issue is effectively handled,

so the recommendation performance is improved.

However, the meta-learning approach represented by

DMATS has its drawbacks. First, its meta-learning

Analysis of Developments and Challenges in Dealing with Recommender System Cold-Start Issue

293

framework is complex, which leads to the complexity

of model training and optimization, that is, it requires

a lot of time and computational resources. Second,

although DMATS considers the effectiveness of tasks

in task scheduling, it still assumes that there is some

correlation between different tasks, which may not be

fully valid in practical applications, especially in

scenarios where user preferences are very diverse,

which means the lack of scenario generalization

ability.

Knowledge graph methods are also widely used

in cold-start recommendation, taking KEGL as an

example. KEGL proposes a collaborative-enhanced

guaranteed embedding generator to guarantee the

quality of embedding and constructs a knowledge-

enhanced gated attention aggregator to adaptively

control the weights of guaranteed embedding and

neighbor embedding, which effectively improves the

recommendation quality and solves the items CSP.

However, even so, this method still fails to avoid the

common problems faced by knowledge graph

methods. First, its effectiveness is highly dependent

on the quality and completeness of the knowledge

graph, and when the quality of the knowledge graph

itself is not high and complete enough, the

recommendation effect will be reduced. Second, the

introduction of two additional operators also

increases the computational complexity, which is

already not low, and raises the time cost of model

training. Finally, although KEGL considers cold-start

neighbors, the lack of interaction information in a

completely cold-start scenario also reduces the

recommendation efficiency.

The LLMs is discussed at the end. On the one

hand, compared with other methods, the LLMs is

nascent, take AutoDisenSeq-LLM as an example, the

performance of facing the CSP is not good enough

compared with other methods. But on the other hand,

because the youth of LLMs brings more possibilities

for this method, LLMs is likely to become the trend

of CSP research in the future. The advantages of

LLMs are obvious, with its excellent text

comprehension in AutoDisenSeq-LLM, the

recommendation list is sorted and optimized, which

improves the accuracy of the recommendation well.

Equally obvious are the drawbacks, which is that the

model is too complex and requires a lot of

computational resources and time, which is prevalent

in current methods.

Current evaluation metrics for cold-start

problems, such as MAE, RMSE, Precision, Recall,

and NDCG, provide a multidimensional assessment

of method performance. However, the lack of

standardized metrics makes it difficult to make a

comprehensive comparison of methods. A more

standardized and comprehensive assessment

framework is needed to address this problem. In terms

of datasets, while public datasets such as MovieLens,

Amazon, and Yelp support cold-start research, they

tend to suffer from low timeliness, slow updates, and

single-domain limitations. These issues hinder the

assessment of method generalizability and cross-

domain capabilities. In addition, some datasets

contain sensitive user information, raising privacy

concerns in an era when data security is increasingly

important. Future datasets should prioritize timeliness,

multi-domain coverage, and privacy protection to

better support the development and evaluation of

cold-start recommendation methods.

4 CONCLUSION

This paper focuses on the current state of the CSP in

RS and analyzes the principles, performance, and

limitations of three advanced approaches: meta-

learning, knowledge graphs, and LLMs. Specifically,

DMATS improves embedding accuracy and

automation, KEGL enhances recommendation

quality through high-quality embeddings, and

AutoDisenSeq-LLM leverages the text

comprehension ability of LLMs to optimize

recommendation lists, thereby improving accuracy.

Experiments with datasets such as Amazon,

MovieLens, and Yelp demonstrate good performance;

however, challenges such as weak scene

generalization, low computational efficiency, and

issues related to datasets timeliness and privacy still

persist. Future research will focus on integrating

LLMs with other advanced methods to leverage their

strengths, improving algorithm efficiency while

reducing computational costs. Additionally, efforts

will be made to develop privacy-preserving datasets

to enhance cross-domain generalization and real-

world applicability. These studies aim to improve the

performance further and generalizability of RS in

dynamic contexts.

REFERENCES

Finn, C., Abbeel, P., & Levine, S. 2017. Model-agnostic

meta-learning for fast adaptation of deep networks. In

International conference on machine learning, 1126-

1135.

He, D., Cui, J., Wang, X., Song, G., Huang, Y., & Wu, L.

2025. Dual Enhanced Meta-learning with Adaptive

PRMC 2025 - International Conference on Public Relations and Media Communication

294

Task Scheduler for Cold-Start Recommendation. IEEE

Transactions on Knowledge and Data Engineering.

Panda, D. K., & Ray, S. 2022. Approaches and algorithms

to mitigate cold start problems in recommender

systems: a systematic literature review. Journal of

Intelligent Information Systems, 59(2), 341-366.

Qian, M., Wei, X., Yi, Q. 2024. Survey on Solving Cold

Start Problem in Recommendation Systems. Journal of

Frontiers of Computer Science & Technology, 18(5).

Saifudin, I., & Widiyaningtyas, T. 2024. Systematic

literature review on recommender system: Approach,

problem, evaluation techniques, datasets. IEEE Access,

12, 19827-19847.

Wang, X., Chen, H., Pan, Z., Zhou, Y., Guan, C., Sun, L.,

& Zhu, W. 2025. Automated disentangled sequential

recommendation with large language models. ACM

Transactions on Information Systems, 43(2), 1-29.

WU, G. D., LIU, X. X., BI, H. J., FAN, W. C., & TU, L. J.

2024. Review of personalized recommendation

research based on meta-learning. Computer

Engineering & Science, 46(02), 338.

Wu, X., Zhou, H., Shi, Y., Yao, W., Huang, X., & Liu, N.

2024. Could small language models serve as

recommenders? towards data-centric cold-start

recommendation. In Proceedings of the ACM Web

Conference, 3566-3575.

Yuan, H., & Hernandez, A. A. 2023. User cold start

problem in recommendation systems: A systematic

review. IEEE access, 11, 136958-136977.

Zhang, J. C., Zain, A. M., Zhou, K. Q., Chen, X., & Zhang,

R. M. 2024. A review of recommender systems based

on knowledge graph embedding. Expert Systems with

Applications, 123876.

Zhang, W., Bei, Y., Yang, L., Zou, H. P., Zhou, P., Liu, A.,

... & Yu, P. S. 2025. Cold-Start Recommendation

towards the Era of Large Language Models (LLMs): A

Comprehensive Survey and Roadmap. arXiv preprint

arXiv:2501.01945.

Zhang, Z., Zhu, Y., Dong, M., Ota, K., Zhang, Y., & Ren,

Y. 2025. Embedding Guarantor: Knowledge-Enhanced

Graph Learning for New Item Cold-Start

Recommendation. IEEE Transactions on Emerging

Topics in Computational Intelligence.

Analysis of Developments and Challenges in Dealing with Recommender System Cold-Start Issue

295