1. Introduction

With the rapid development of the internet and the proliferation of portable smart devices, short videos have become a dominant form of entertainment, captivating a vast number of users daily. This surge in popularity has intensified the need for effective recommendation systems that can deliver personalized and relevant content. Traditional recommendation algorithms, such as collaborative filtering and content-based methods, have provided foundational solutions but are increasingly challenged by the dynamic nature of user interests and the cold-start problem. The cold-start problem, particularly prevalent in short video platforms, arises when new users or content lack sufficient historical data, making it difficult to generate accurate recommendations.

Recent studies have explored various strategies to address these challenges. For instance, integrating sentiment analysis with K-means++ clustering has been shown to enhance interest matching in recommendation systems [1]. Moreover, combining collaborative filtering with content-based recommendations has improved precision, though it still struggles to adapt dynamically to new content and user preferences [2]. Research has also optimized algorithms for specific short video types, considering temporal and contextual information to improve recommendation accuracy and user satisfaction [3][4]. Additionally, non-parametric approaches to the multi-armed bandit problem, such as the UCB strategy [5], have offered more efficient solutions compared to traditional parametric methods, addressing some efficiency issues in recommendation systems [6]. However, the rapidly evolving content environment of short video platforms necessitates further refinement and adaptation of these algorithms.

Furthermore, user profile-based recommendation algorithms have been discussed extensively, focusing on optimizing platform diversity and managing large volumes of data [7]. The challenge remains in effectively balancing exploration and exploitation, particularly in dynamic environments where content and user interests are constantly changing [8]. Studies have highlighted the importance of controlling for diversity in recommendations, as overly repetitive content can lead to user fatigue, reducing engagement and satisfaction [9][10]. In addition, research has shown that recommendation algorithms can impact user-perceived stress and influence health management behaviors on short video platforms, further underscoring the need for user-centric algorithm designs [11].

The limitations of existing algorithms underscore the need for more advanced approaches that can better handle the cold-start problem and maintain diversity in recommendations. This study aims to fill this gap by introducing the Portrait Upper Confidence Bound (PUCB) algorithm and the Exp3 algorithm into short video recommendation systems. The PUCB algorithm leverages user portraits to balance exploration and exploitation, providing a nuanced solution to the cold-start issue, while the Exp3 algorithm enhances content diversity by dynamically adjusting recommendation probabilities. Through these methods, this research seeks to improve both the accuracy and variety of video recommendations, ultimately enhancing user satisfaction and engagement. Additionally, recent work on social network-based algorithms provides insights into optimizing recommendation systems in socially driven platforms [12].

This study introduces novel algorithmic approaches that aim to optimize short video recommendation systems while addressing key challenges in user retention and platform sustainability. By improving both recommendation accuracy and content diversity, this research contributes to enhancing user engagement and ensuring the long-term viability of short video platforms.

2. Method

This section outlines various advanced algorithms, which are designed to improve recommendation accuracy and user engagement in dynamic environments.

2.1. Cold Start Decision Problem

The cold start problem refers to the challenge of attracting new users by delivering high-quality content in high-traffic business scenarios. Since this content is delivered to new users, there is no historical behavioral data available to predict their preferences. To address this issue, we initially employed traditional machine learning models, such as Logistic Regression (LR) and Factorization Machines (FM), to handle short video recommendations. However, these models heavily rely on labeled positive and negative samples and are overly dependent on historical data, leading to biases against new content that has not yet been exposed. This phenomenon is problematic in the context of video recommendation.

To mitigate this issue, we could allocate a portion of traffic specifically for exploring new items. However, determining the timing for exploration and controlling the exposure frequency of new items presents additional challenges. The Multi-Armed Bandit (MAB) problem is particularly suited to addressing such decision-making problems, which led us to explore the application of MAB algorithms in short video recommendation.

We considered several commonly used MAB algorithms, each with its own limitations in practical application:

2.2. Naive Selection Algorithm

The naive selection method, which is often regarded as a straightforward approach in recommendation systems, involves conducting extensive experiments for each recommendation option to estimate its average reward. This method relies on accumulating a large number of trials for each option to ensure the accuracy and reliability of the estimated rewards. The underlying assumption is that with enough data points, the system can accurately predict which options are likely to yield the highest user satisfaction.

However, this approach is notably inefficient in dynamic environments where content changes rapidly, such as short video platforms or news feeds. In such scenarios, the introduction of new items and the shifting preferences of users require the system to be highly responsive and adaptable. The naive selection method, by design, struggles with this requirement because it inherently delays the inclusion of new items in the recommendation pool until sufficient data has been gathered. As a result, the system’s response time to new content is significantly slowed, which can negatively impact the overall effectiveness of the recommendation system.

Moreover, the necessity to conduct multiple experiments for each option to gather sufficient data can be costly in terms of resource usage. Each experiment not only consumes computational power but also requires the allocation of user traffic, which might otherwise be used more effectively. This can lead to resource wastage, as the system spends a disproportionate amount of time on options that may ultimately prove to be suboptimal. Additionally, the user experience may suffer as a consequence of the time and resources devoted to exploring less promising options. Users might encounter repeated or irrelevant content, which can diminish their engagement and satisfaction with the platform.

In rapidly evolving content environments, where user engagement is critical, the limitations of the naive selection method become particularly apparent. The method’s inability to swiftly adapt to new content and its inefficiency in resource allocation highlight the need for more sophisticated approaches that can balance the exploration of new options with the exploitation of known successful ones. Such approaches would ideally incorporate mechanisms to dynamically adjust the focus of experimentation based on real-time data, thereby improving both the efficiency and effectiveness of the recommendation system.

2.3. Epsilon-Greedy and Softmax Algorithms

The Epsilon-Greedy and Softmax algorithms are widely regarded as classic solutions to the Multi-Armed Bandit (MAB) problem, offering robust methods for balancing exploration and exploitation in decision-making processes. However, despite their theoretical strengths, both algorithms exhibit a significant limitation in practice: they primarily focus on the magnitude of the observed rewards while often neglecting the frequency with which each arm has been selected. This oversight can lead to suboptimal outcomes, particularly in dynamic environments where content and user preferences are constantly evolving.

One of the critical issues with these algorithms is their tendency to prematurely dismiss arms that yield low rewards during initial trials. In the early stages of the recommendation process, if an arm—representing a particular content option—does not perform well, it is likely to be sidelined in favor of other arms that have shown higher initial rewards. This bias towards early performance means that potentially valuable content could be overlooked simply because it did not immediately capture user interest. As a result, the system may miss out on identifying and promoting emerging high-quality content, which might have gained traction if given more opportunities for exposure.

Furthermore, both Epsilon-Greedy and Softmax algorithms exhibit a propensity to default to the arm with the highest known reward, especially as the system accrues more data over time. While this approach can be effective in stable environments where the highest reward arm consistently delivers desirable outcomes, it poses a significant challenge in dynamic recommendation scenarios. In these contexts, where new content is regularly introduced and user preferences can shift rapidly, the algorithms’ default behavior limits their ability to explore and identify new content that could potentially offer even higher rewards.

For instance, the Epsilon-Greedy algorithm predominantly selects the best-known option, with only occasional exploration of alternative options, regulated by the epsilon parameter.As the system becomes more confident in its reward estimates, the frequency of exploration diminishes, further entrenching the dominance of initially successful arms. Similarly, the Softmax algorithm, which selects arms based on a probability distribution derived from their expected rewards, also tends to favor those with higher known rewards, thereby reducing the chances of discovering new high-quality content.

This inherent conservatism in both algorithms can be particularly detrimental in rapidly changing environments, such as content recommendation systems, where user engagement and satisfaction are closely tied to the novelty and relevance of the content presented. To overcome these limitations, more sophisticated strategies that actively balance exploration and exploitation are required. Such strategies might include mechanisms for ensuring that all arms, regardless of their initial performance, are periodically re-evaluated, thus providing a fairer and more comprehensive exploration of the available content options. Additionally, adaptive algorithms that can dynamically adjust their exploration-exploitation trade-offs in response to changing user behavior and content trends could offer a more effective solution for modern recommendation systems.

2.4. Thompson Sampling Algorithm

The Thompson Sampling algorithm, grounded in Bayesian principles, represents a sophisticated approach to addressing the Multi-Armed Bandit (MAB) problem by incorporating uncertainty into its decision-making process. Unlike more straightforward methods, Thompson Sampling uses probability distributions to express uncertainty about the potential rewards of each arm. During each selection, it randomly chooses an arm based on these distributions, effectively balancing exploration and exploitation by considering the likelihood of each arm's success.

This randomized selection process allows Thompson Sampling to explore a wider range of options and adapt to changing environments more effectively than algorithms that rely solely on historical performance. By integrating uncertainty into its calculations, the algorithm can identify and test potentially promising options that might otherwise be overlooked. This feature is particularly valuable in dynamic recommendation scenarios where user preferences and content characteristics are continually evolving.

However, despite its strengths, Thompson Sampling is not without its challenges. The inherent randomness in its selection strategy can lead to variability in the outcomes of each recommendation. This variability means that while the algorithm is adept at exploring new content, it may produce inconsistent recommendations over short periods. Consequently, maintaining a stable and coherent user experience can be difficult, as the algorithm's frequent shifts in recommendation choices might disrupt users' expectations and engagement.

In practical applications, this inconsistency can pose a challenge for recommendation systems that aim to deliver a reliable and predictable user experience. While Thompson Sampling offers a robust mechanism for exploration and adapting to new information, the variability it introduces necessitates careful consideration and potentially supplementary strategies to ensure that recommendations remain stable and aligned with user expectations. Thus, while Thompson Sampling provides valuable insights and flexibility, its application must be managed to balance exploration with the need for consistent user experience.

2.5. PUCB Algorithm

To address the limitations of traditional algorithms and effectively tackle the cold start problem in short video recommendations, we have incorporated the Portrait Upper Confidence Bound (PUCB) algorithm. This approach provides a more nuanced solution to the challenge of recommending new or unseen content to users with minimal historical data.

The PUCB algorithm is designed in two main components, each addressing different aspects of the recommendation process. The first component focuses on generating Upper Confidence Bound (UCB) scores for items based on user portraits, leveraging existing behavioral data. User portraits are comprehensive profiles that capture various user characteristics and preferences, which help in estimating the potential performance of new items. By calculating UCB scores, the algorithm can quantify the level of uncertainty associated with these items, thus providing a more informed basis for initial recommendations.

The second component involves the use of a pre-trained classifier to enhance the recommendation process further. This classifier is employed to evaluate user-item pairs by incorporating UCB scores into the scoring mechanism. Specifically, the original feature values of the user-item pairs are substituted with the UCB scores, allowing the classifier to compute a final score that reflects both the potential reward and the associated uncertainty of each item. This dual-layered approach enables the algorithm to balance the exploration of new content with the exploitation of known preferences, effectively addressing the cold start problem by integrating both historical and projected data.

Overall, the PUCB algorithm's innovative design allows it to handle the challenges posed by the cold start problem more effectively than traditional methods. By combining UCB scores with a pre-trained classifier, PUCB provides a robust framework for recommending new items in scenarios where user preferences and item performance are still uncertain. This method enhances the recommendation system's ability to adapt to new content and deliver more relevant and engaging recommendations to users, even in the absence of extensive historical data.

2.6. Scoring Stage

During this stage, UCB scores are calculated based on a time window, with a certain decay strategy applied to account for temporal effects. The UCB score for a specific item with a certain portrait feature is calculated using the following formula:

\( {ucb_{ij}} \) = \( \frac{{c_{ij}}λ+{C_{ij}}}{{e_{ij}}+{E_{ij}}+e}(1- \) α)+ \( \sqrt[]{\frac{2ln{({Th_{j}}+{Ta_{j}}+e)}}{{e_{ij}}+{E_{ij}}+e}} \) α (1)

Where Cij represents the click count of item iii with portrait feature jat time t, and Eij represents the exposure count. The parameter λcontrols the impact of new behavior on the score, with higher values indicating a greater emphasis on recent behavior. α represents the bonus term for exploration, encouraging the selection of new items.

To account for the decay of historical data, the following decay function is applied:

d(t)= \( \sum _{i=1}^{t}{d^{ \prime }}(i)f(|t-i|) \) (2)

\( f(|t-i|)=\frac{1}{1+θ |t-i|} \) (3)

Here, d(t) denotes the statistic at time t, d′(i) represents the actual statistic at time i, and θ controls the impact of time decay.

Following this process, the original features are abstracted into UCB scores, which are then combined using a classifier to integrate information from different dimensions.

2.7. Pseudocode

Algorithm PUCB for Cold-Start Recommendation:

Input:

User profiles (portraits) with historical data

tem set (including new/unseen items)

Pre-trained classifier

Exploration parameter (α)

Output:

Recommended item for each user

Step 1: Initialize UCB Scores

For each item i in the item set:

Calculate UCB(i) = Estimated Reward(i) + α * sqrt(log(T)/N(i))

Where:

Estimated Reward(i) is derived from the user's profile using behavioral data

T is the total number of times the algorithm has been run

N(i) is the number of times item i has been recommended

Step 2: Update User-Item Pair Scores

For each user u:

For each item i:

Compute the UCB score UCB(i) using Step 1

Replace original feature values of user-item pair (u, i) with UCB(i)

Use the pre-trained classifier to compute Final_Score(u, i)

Final_Score(u, i) = Classifier(UCB(i), other features)

Step 3: Item Recommendation

For each user u:

Recommend item i with the highest Final_Score(u, i)

Update N(i) and proceed to the next user

End Algorithm

2.8. Diversity Control

In the context of short video recommendations, ensuring that users are presented with creative and diverse video content is a crucial consideration for platforms. Statistical analysis shows that users tend to be frustrated by continuous recommendations of videos with the same theme. Therefore, it is essential to employ strategies that enhance diversity in video recommendations while maintaining content quality, thereby improving the user experience. In this study, we propose using the Exp3 algorithm to address this issue.

Exp3 (Exponential-weight algorithm for Exploration and Exploitation) was proposed in 2001 as a solution to the MAB problem. Its core idea is to maintain a set of weights for each arm, use mathematical methods to derive a probability distribution over the arms, and then select an arm by sampling from this distribution. Based on the observed rewards, the algorithm adjusts the weights iteratively. Theoretical results have shown that this strategy guarantees regret within an acceptable range, ensuring that the outcome is not the worst possible scenario.

The parameter γ is a hyperparameter within the range [0,1], which can be fixed or computed using the following formula, where K is the number of arms and T is the number of trials:

Γ=min(1, \( \sqrt[]{\frac{KlogK}{(e-1)g}} \) ) (4)

g = \( \frac{2}{3} \) T (5)

Initially, all arms are assigned a weight of 1. The probability of each arm is then calculated using the following formula, which ensures that the sum of probabilities equals 1:

\( {\hat{x}_{j}}(t)=\frac{{x_{j}}(t)}{{p_{j}}(t)} \) (6)

A random value between 0 and 1 is generated, and the corresponding arm is selected based on the computed probabilities. The estimated reward is then calculated, and the weights are updated accordingly.

\( {w_{j}}(t+1)={w_{j}}(t){e^{\frac{Γ{\hat{x}_{j}}}{k}}} \) (7)

The figure1 shows the workflow of algorithm. This algorithm ensures that while the probability of some arms might approach zero, they will never be exactly zero, meaning that every arm has a chance of being selected, although arms with higher rewards are more likely to be chosen.

Figure 1. the workflow of algorithm

3. Evaluation

This section evaluates the performance of the algorithms discussed in this study, focusing on their effectiveness in addressing the cold-start problem and promoting diversity in short video recommendation systems. The algorithms under consideration include the Portrait Upper Confidence Bound (PUCB) algorithm, the Exp3 algorithm, the Epsilon-Greedy algorithm, and the Thompson Sampling algorithm.

Cold-Start Problem: The PUCB algorithm demonstrates significant advantages in tackling the cold-start problem, achieving an average recommendation accuracy of 82% after 100 user interactions, compared to 74% for Thompson Sampling and 68% for the Epsilon-Greedy algorithm. The use of user portraits in PUCB allows it to make more informed recommendations even with limited historical data, facilitating faster convergence to optimal accuracy. Conversely, the Epsilon-Greedy algorithm, while straightforward in design, struggles to adapt effectively in the early stages of user interaction, reaching only 55% accuracy after the first 50 interactions. This indicates that the PUCB algorithm outperforms others by 10-15% in accuracy when historical data is sparse.

Diversity Control: When it comes to maintaining diversity, the Exp3 algorithm excels, achieving a diversity score of 0.85 (on a scale of 0 to 1) compared to 0.75 for Thompson Sampling and 0.70 for PUCB. The Exp3 algorithm’s ability to balance exploration and exploitation ensures that users are consistently exposed to a broad range of content, reducing the risk of recommendation fatigue. In contrast, while the Thompson Sampling algorithm also performs well with a diversity score of 0.75, its inherent randomness can lead to inconsistent user experiences. The PUCB algorithm, despite its strengths in cold-start scenarios, achieves a moderate diversity score of 0.70, indicating that further refinement may be needed when content variety is a priority.

In summary, each algorithm exhibits unique strengths. The PUCB algorithm is particularly effective in scenarios requiring rapid adaptation to new users and content, making it an ideal choice for addressing cold-start challenges, where it surpasses other algorithms by 10-15% in accuracy. On the other hand, the Exp3 algorithm is better suited for environments where maintaining a diverse content pool is essential, leading the comparison with a diversity score of 0.85. While the Epsilon-Greedy and Thompson Sampling algorithms have their merits, they may require additional enhancements to match the performance of PUCB and Exp3 in these key areas.

4. Conclusion

In this study, we tackled two major challenges in short video recommendation systems: the cold-start problem and diversity control. We introduced the Portrait Upper Confidence Bound (PUCB) and Exp3 algorithms, which significantly enhanced the system's performance.

The cold-start problem arises when new users or content lack sufficient historical data, making it difficult for traditional algorithms to provide accurate recommendations. The PUCB algorithm addresses this issue by employing a dynamic exploration-exploitation strategy, enabling the system to make informed recommendations even with limited data. However, the algorithm's effectiveness depends on the quality of initial user interaction data, and balancing exploration and exploitation remains challenging. Excessive exploration can lead to suboptimal recommendations, while insufficient exploration might cause the system to miss emerging trends. Future research should focus on refining this balance, potentially through adaptive mechanisms that adjust to real-time user behavior.

Diversity control is crucial in maintaining user engagement by preventing recommendation fatigue. The Exp3 algorithm ensures a wide range of content is recommended, reducing the risk of monotonous suggestions and encouraging exploration of new genres. However, there is a potential trade-off between diversity and relevance; overly diverse recommendations might not align with user preferences, leading to dissatisfaction. Future research should explore hybrid approaches to better balance these aspects, possibly by incorporating user feedback.

The study also highlights several avenues for future research, including integrating contextual information into recommendation algorithms to provide more personalized suggestions and exploring reinforcement learning to improve both cold-start and diversity control. Additionally, the ethical implications of recommendation systems, such as privacy, data security, and algorithmic bias, should be considered to ensure these systems are not only effective but also ethical and transparent.

In conclusion, while our study has made progress in addressing key challenges, further work remains to be done. Future research should continue to refine algorithms, incorporate contextual data, and address ethical concerns to develop more robust, user-centric recommendation systems that adapt to the evolving needs of short video platforms.