Dynamic User Preferences Optimization in Time-aware Recommendations Ainaz Ebrahimi1 , Zheying Zhang1 and Kostas Stefanidis1 1 Tampere University, Finland Abstract Recommender systems play a vital role in mitigating information overload by predicting user preferences. While traditional algorithms like collaborative filtering and content-based filtering have demonstrated their effectiveness, they often struggle to adapt to the dynamic nature of user preferences over time. This study addresses these limitations by enhancing the Time Correlation Coefficient (TCC) model with time-aware techniques, providing a more sophisticated understanding of the temporal shifts in user interests. We propose four advanced methodologies: Content-based Similarity, Time-based Decay, Cuckoo Search Optimization, and Decay Model Selection, each designed to improve recommendation accuracy by integrating dynamic, time-sensitive elements into the recommendation process. Our experiments reveal significant improvements in recommendation precision, demonstrating the advantages of these methodologies over the baseline TCC model in various performance metrics. The results emphasize the effectiveness of these dynamic strategies in personalizing user experiences, with a balanced approach to both accuracy and computational efficiency. This work lays a solid foundation for future research in recommendation technologies, offering practical insights and applications that can be extended across diverse domains. By enhancing recommender systems with a deeper understanding of temporal user behavior, we aim to improve the overall user experience in digital environments. Keywords Dynamic Attenuation Coefficient, Time-aware Recommendations 1. Introduction with sparse data [13]. Hybrid recommendation systems have emerged to ad- Recommendation systems are an integral part of many on- dress these issues, integrating the strengths of both content- line platforms, designed to deliver personalized suggestions based and collaborative filtering to provide more accurate to users across various sectors [1]. These systems analyze and diverse recommendations [14]. Leading companies such large volumes of user and item data to predict user pref- as Amazon and Netflix increasingly adopt these hybrid mod- erences, aiming to recommend items that users are likely els, delivering more comprehensive and personalized recom- to enjoy, even without prior exposure to similar products mendations by analyzing both content attributes and user or services [2]. Prominent examples such as Amazon and behavior [14, 15]. However, these systems typically assume Netflix have led the way, utilizing recommendation systems static user preferences, failing to capture the temporal evo- to offer tailored suggestions based on user behavior and lution of user interests, which naturally shift over time due preferences. to personal growth, changing tastes, and external influences There are two primary approaches used in recommen- [12, 16]. To address these shortcomings, time-aware rec- dation systems: content-based filtering and collaborative ommendation systems have been developed, incorporating filtering [3]. Content-based filtering recommends items by temporal data to enhance the accuracy and relevance of evaluating the attributes of items a user has previously en- predictions [17, 18, 19]. Recent advancements in this field gaged with, whereas collaborative filtering provides recom- have focused on integrating temporal factors into recom- mendations by leveraging the preferences of similar users. mendation processes. One such study introduced the Time Both methods have their advantages and are frequently com- Correlation Coefficient (TCC) model, which combines a time bined to enhance recommendation accuracy and reliability. correlation coefficient with optimized K-means clustering to The content-based filtering approach excels in recommend- improve recommendation accuracy [20]. Despite these ad- ing items like articles or news by focusing on the properties vancements, existing time-aware models, particularly those of the items themselves [4, 5]. It employs algorithms such using the TCC [20], often struggle to fully capture the evolv- as the Vector Space Model and probabilistic models, includ- ing nature of user preferences. These models tend to over- ing the Naive Bayes Classifier, Decision Trees, and Neural look item similarity and the complex temporal dynamics of Networks, to analyze item similarities and generate relevant user behavior, leading to less precise recommendations. recommendations [6, 7]. The collaborative filtering tech- To overcome these limitations, this paper focuses on en- nique uses user-item interaction data to predict preferences hancing the recently proposed TCC model, building upon based on the choices of similar users [8, 9, 10, 11]. Although its foundation to better address the evolving nature of user both content-based and collaborative filtering have proven preferences over time. In this regard, we propose several effective, they have inherent limitations. Content-based techniques, all of which share the same overarching goal of systems often lack diversity, as they tend to recommend improving the TCC model. Each technique offers a distinct items similar to those the user has already consumed [12]. approach, providing valuable advantages in different sce- Collaborative filtering, while addressing this issue, faces narios, but the ultimate objective remains to enhance the challenges with scalability, especially in large user bases performance and adaptability of the TCC model. This leads to several important questions: Published in the Proceedings of the Workshops of the EDBT/ICDT 2025 Joint Conference (March 25-28, 2025), Barcelona, Spain • How can the accuracy of time-aware recommenda- Envelope-Open ainaz.ebrahimi@tuni.fi (A. Ebrahimi); zheying.zhang@tuni.fi tion systems, especially those using the TCC, be (Z. Zhang); konstantinos.stefanidis@tuni.fi (K. Stefanidis) Orcid 0000-0002-6205-4210 (Z. Zhang); 0000-0003-1317-8062 improved to better capture the evolving nature of (K. Stefanidis) user preferences? © 2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). • What modifications can be made to the TCC for- CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings mula to more effectively integrate item similarity significant advancement is Collaborative Topic Regression and temporal dynamics? (CTR), which integrates content-based features with collab- • What innovative techniques can be developed to orative filtering through Latent Dirichlet Allocation (LDA), incorporate temporal context and improve the accu- incorporating temporal dynamics to track changes in user racy and relevance of recommendations? preferences over time [25]. Matrix factorization techniques have also been combined with neural networks to capture This paper presents the following contributions to address temporal patterns more effectively. For instance, Collabora- these challenges: tive Deep Learning (CDL), a hierarchical Bayesian model, merges deep representation learning for content with collab- • Enhancement of the TCC algorithm by incorporat- orative filtering for ratings. This model effectively manages ing item similarity scores, improving recommenda- sparse data while capturing the temporal evolution of user tion accuracy. preferences [26]. • Development of four innovative algorithms designed Beyond matrix factorization, time-aware collaborative to adapt to the evolving nature of user preferences, filtering approaches have been extensively explored. Time- enabling the detection of shifts in interests over time. Weighted Collaborative Filtering (T-UCF) applies an expo- These algorithms enhance the ability to deliver per- nential decay formula to older data, giving more weight to sonalized and highly accurate recommendations. recent interactions [27]. This approach ensures that more re- • Extensive experiments conducted on three datasets cent user behaviors have a greater impact on the recommen- (two from Amazon and one from MovieLens), dations, improving accuracy in scenarios where user pref- demonstrating the effectiveness of the proposed erences evolve rapidly. Additionally, temporal clustering model in improving the accuracy and relevance of models, such as the multiclass co-grouping (MCoC) model recommendations. presented in [28], further enhance recommendation preci- The structure of the paper is as follows: In Section 2, sion by segmenting users and items into subgroups based we provide a review of recent developments in time-aware on temporal patterns. Another notable example is Bayesian recommendation systems and collaborative filtering models. Probabilistic Tensor Factorization (BPTF), which models the Section 3 presents the proposed algorithms based on the user- item interaction as a three-dimensional tensor (user, Time Correlation Coefficient and its improvements. Section item, time). This allows the model to capture the evolution 4 details the experimental setup, including the datasets and of both user preferences and item characteristics over time. evaluation metrics employed, while Section 5 presents the The BPTF model, introduced by Xiong et al. [29], is par- results accompanied by a comprehensive analysis. Finally, ticularly useful for handling large, sparse datasets, such as Section 6 provides the concluding remarks of the paper. those found in movie recommendation scenarios [30]. In another study, Ahmadian et al. [31] proposed the Recommender System with Temporal Reliability and Con- 2. Related Works fidence (RSTRC), which integrates temporal factors into reliability and confidence measurements. This system dif- Over the past decades, research in recommendation systems fers from previous work by incorporating time into both has steadily evolved, progressing from traditional methods the confidence scores and reliability assessments of user like Collaborative Filtering (CF) and Content-Based Filtering profiles, thereby improving recommendation precision. Fur- (CBF) to more sophisticated models that address dynamic thermore, FSTS, a novel search technique incorporating both changes in user preferences. Traditional recommendation time-sensitive parameters and stability variables, has been systems, though effective, face limitations when it comes to evaluated on the MovieLens dataset. The algorithm demon- accounting for temporal aspects of user-item interactions strated improvements in coverage, popularity, recall, and [1, 21]. precision, although it struggled with the dynamic changes A key challenge that traditional recommendation sys- in time-sensitive factors [32]. Cui et al. [20] developed a tems face is their inability to account for the evolution of model specifically for Internet of Things (IoT) environments, user preferences over time. Time- aware recommendation combining a Time Correlation Coefficient with a refined systems (TARS) seek to remedy this by explicitly incorpo- K-means clustering algorithm. By leveraging temporal dy- rating temporal factors into their models [22, 23]. TARS namics, their model demonstrated a 5.2 % improvement in leverage the fact that user preferences are not static and recommendation accuracy on datasets such as MovieLens change over time, improving the relevance of recommen- and Douban. This highlights the increasing importance of dations by modeling time-based patterns of user behavior. temporal factors in domains where user preferences are One of the earliest and most notable approaches in this field heavily time dependent. However, it encountered limita- is Time-SVD++, introduced by Koren [22]. This method tions in capturing time-dependent user preferences, which extends the matrix factorization technique by incorporating our study aims to address. time-dependent factors for both users and items, allowing Our research integrates temporal dynamics to model the the model to account for the gradual changes in user pref- evolution of user preference behaviors more effectively. Our erences. The Time-SVD++ model proved its effectiveness primary focus is on enhancing the Time Correlation Coef- during the Netflix Prize competition, where it outperformed ficient (TCC) by incorporating item similarity scores, al- many traditional collaborative filtering techniques by mod- lowing the model to account for both temporal variations eling user behavior over time. and item-specific relationships. Furthermore, we introduce With considering Temporal Dynamics in Matrix Fac- innovative algorithms designed to determine a personal- torization Building new matrix factorization models have ized interest-shifting parameter for each user, enabling the emerged. Collaborative Evolution (CE) is one such model system to dynamically adapt to changes in each user pref- that captures temporal changes by introducing a time- erences over time. These advancements collectively aim to dependent factor into matrix factorization [24]. Another improve the precision and relevance of the recommenda- tions provided by the TCC model. Correlation Coefficient (TCC) formula [20]: 1 Δ𝑡 2 tcc𝑖 = 1 − (1 − exp (− 2 )) (1) 3. Methodology √2𝜋𝜎 2𝜎 In this paper, we address key limitations in current time- where: aware recommendation system, particularly the Time Cor- • Δ𝑡 = 𝑡𝑖 − 𝑡1 is the time difference between the 𝑖-th relation Coefficient (TCC) model. The TCC model is one and most recent interaction. of the foundational approach for time-aware recommenda- • 𝜎 is the attenuation coefficient. tions. It employs a coefficient (TCC) to adjust all ratings, incorporating the influence of time on user interests. , but it Algorithm steps are in below: faces two critical limitations: (1) it applies a uniform, static 1. Time Differences Calculation: Compute the time attenuation coefficient to all user ratings without adapting differences (Δ𝑡) between user interactions. to user behavior or context, and (2) it does not consider item 2. Determine 𝜎: Use a single static attenuation coeffi- similarity, which is essential for capturing relationships be- cient determined through trial and error. tween items in a user’s preference profile. 3. Calculate TCC: Apply the formula above to com- To overcome these limitations, we propose a set of four pute TCC values for all user ratings, weighting them methodologies, each designed to supplement and improve accordingly. the TCC model. While all these methods share the ultimate 4. Generate Recommendations: Utilize the adjusted goal of enhancing the accuracy and performance of TCC, ratings to produce personalized recommendations. they address different aspects of its improvement. Three methodologies focus on determining the attenuation coeffi- Limitations: While effective, this method has certain cient dynamically, making it more adaptive and personal- limitations: ized, while one of them addresses the lack of item similarity 1. Static Attenuation Coefficient: The use of a single in TCC. 𝜎 value does not account for individual user behav- The Proposed Methods: iors or dynamically changing preferences. 1. Content-Based Similarity (Addressing Item Similar- 2. Lack of Item Similarity Consideration: TCC ity): Incorporates item similarity into TCC using does not incorporate item similarity, which is a crit- NLP techniques to analyze item content, ensuring ical factor in refining recommendations. older ratings for similar items remain relevant. To overcome these challenges, this study proposes en- 2. Time-Based Decay (Addressing Dynamic Attenuation hancements to the TCC model, which will be explained in Coefficient): Introduces a time-sensitive coefficient the following. to model the diminishing relevance of older ratings, adapting to temporal dynamics. 3. Decay Model Selection (Addressing Dynamic Atten- 3.2. Content-Based Similarity Model uation Coefficient): Dynamically selects the most This approach balances historical and recent user interac- suitable decay model based on user behavior and tions to maintain the influence of past evaluations, thus dataset characteristics. enhancing personalization. The TCC cannot be calculated 4. Cuckoo Search Optimization (Addressing Dynamic accurately without considering item similarity. Even older Attenuation Coefficient): Optimizes the attenuation ratings can be valuable if the item is highly similar to the coefficient using metaheuristic techniques to adapt most recent item the user has rated. Proposed enhancements to evolving user preferences. to the TCC involve using NLP techniques to calculate item similarity. Specifically, descriptive attributes of products are To validate the proposed approaches, we conduct experi- analyzed to determine similarity between the most recently ments on real-world datasets, including Amazon and Movie- rated item and others. Algorithm steps are in below: Lens. The results demonstrate that these methods effec- tively improve the TCC model’s accuracy and performance 1. Descriptive Feature Analysis: Analyze product in diverse recommendation scenarios. features (e.g., brand, material) using NLP techniques In the following sections, we first discuss the limitations such as TF-IDF. of the existing TCC model in Section 3.1. From Sections 3.2 2. Similarity Score Calculation: Compute the cosine to 3.5, we provide detailed descriptions of each proposed similarity between the most recently rated product method, highlighting their specific contributions to refining 𝑓recent and previously rated products 𝑓𝑗 : the TCC model. ⟨𝑓recent , 𝑓𝑗 ⟩ 𝑠(𝑓recent , 𝑓𝑗 ) = (2) ‖𝑓recent ‖‖𝑓𝑗 ‖ 3.1. Time Correlation Coefficient (TCC) The Time Correlation Coefficient Collaborative Filtering 3. Threshold and Rating Check: (TCCF) is a recommendation approach that enhances accu- • If the similarity score 𝑠(𝑓recent , 𝑓𝑗 ) is above racy by integrating temporal dynamics into the recommen- a predefined threshold (based on domain dation process. Traditional collaborative filtering methods knowledge and datasets) and the rating 𝑟𝑗 ≥ 4, often assume static user preferences, overlooking the phe- set: nomenon of ”interest drift,” where user preferences evolve TCC = 1 (3) over time. TCCF mitigates this issue by assigning greater • Otherwise, calculate the TCC based on its for- weight to more recent interactions, as reflected in the Time mula (Equation 1) and then multiply it by all ratings 𝑟𝑗 for each user 4. Generate Recommendations: 4. Return Decay Parameters: Provide both 𝑏exp and • Use the adjusted TCC values to generate rec- 𝑏gauss for further analysis. ommendations. 5. Calculate TCC: Integrate the decay parameters into the Time Correlation Coefficient (TCC) formula to The proposed approach improves recommendation preci- compute the final coefficient. sion by considering item similarity and recent ratings, dy- 6. Generate Recommendations: Use the decay pa- namically adjusting the TCC to reflect current user interests. rameters to refine user ratings and improve engage- It enhances efficiency through the use of NLP techniques for ment strategies. similarity calculations, delivering more precise and relevant recommendations. The decay functions are defined as follows: 𝑦exp = 𝑎 ⋅ 𝑒 −𝑏exp 𝑥 (5) 3.3. Time-Based Decay Method This method, along with the two subsequent approaches, 𝑥−𝑏gauss 2 −( ) dynamically calculates a personalized attenuation coeffi- 𝑦gauss = 𝑎 ⋅ 𝑒 𝑐 (6) cient (𝜎) for each user, in contrast with the baseline model where: where 𝜎 was static and determined via trial and error. By tailoring 𝜎 to user behavior, this method captures temporal • 𝑎 is the initial value or amplitude, determining the patterns more effectively, improving the personalization starting point of the decay curve. and accuracy of recommendations. The approach uses an • 𝑏exp and 𝑏gauss are the decay rate parameters, indi- exponential decay model to account for the diminishing cating how quickly the decay occurs over time. influence of past interactions over time. Algorithm steps • 𝑐 is the standard deviation in the Gaussian model, are in below: controlling the width of the decay curve and indicat- 1. Time Differences Calculation: Compute the time ing how spread out the decay is around the mean. differences (time-diff𝑖 ) between consecutive interac- Decay functions provide valuable insights into user be- tions using normalized timestamps. havior by revealing how quickly user satisfaction diminishes 2. Determine Decay Factor: Set a constant decay over time. For instance, a rapid decline (exponential) may factor (here is considered 0.5 as a balance) to control indicate the need for immediate follow-ups, while a gradual the rate at which past interactions lose relevance. decline (Gaussian) suggests sustained engagement efforts. 3. Exponential Decay: Use the formula: Additionally, decay parameters enable forecasting trends, allowing organizations to proactively address potential de- 𝜎𝑖 = exp(−decay-factor ⋅ time-diff𝑖 ) (4) clines in user satisfaction. This dynamically computes the decay coefficient (𝜎) for each interaction. 3.5. The Cuckoo Search Optimization 4. Calculate TCC: Integrate the computed 𝜎 into the Technique TCC formula to adjust ratings, ensuring recommen- In this section, we explain the Cuckoo Search optimization, dations reflect the temporal relevance of user inter- a metaheuristic algorithm inspired by the brood parasitism actions. behavior of cuckoo birds. We then describe the steps in- 5. Generate Recommendations: Use the adjusted volved in using Cuckoo Search to determine the optimal TCC values to generate recommendations. sigma (𝜎) value for each user, which is critical for improving This method enhances temporal sensitivity by adapting recommendation accuracy. to shifts in user preferences, using a smooth exponential decay function for realistic modeling. 3.5.1. Cuckoo Search Algorithm The Cuckoo Search algorithm is a metaheuristic optimiza- 3.4. Decay Model Selection Method tion technique that excels in solving complex optimization This approach employs both exponential and Gaussian de- problems by balancing exploration and exploitation. It uses cay functions to analyze how user ratings evolve over time. random nest selection and Lévy flights for exploration, in- By utilizing a dual-model technique, it identifies whether spired by the cuckoo bird’s brood parasitism behavior. The ratings decline rapidly (exponential) or steadily (Gaussian), algorithm operates under the following key rules [20]: offering insights into user engagement and satisfaction. The • Random egg-laying in nests. method aims to dynamically adapt to temporal changes in • Retaining nests with the best eggs (solutions). user preferences, enhancing the accuracy of time-aware • Generating new nests via Lévy flights if a nest is recommendations. Algorithm steps are in below: discovered with a probability 𝑝. 1. Check for Sufficient Data: Ensure at least two data points are available for model fitting. The position update for a cuckoo is defined as: 2. Fit Exponential Decay Model: Use the formula (𝑡 ′ ) (𝑡) (𝑡) (𝑡) y𝑖 = y𝑖 + 𝛽 ⋅ 0.01 ⋅ 𝑀 ⋅ (y𝑖 − q𝑔 ) ⋅ 𝑠 (7) provided in Equation 5 to fit the data and extract the decay parameter 𝑏exp . where: 3. Fit Gaussian Decay Model: Use the formula pro- (𝑡 ′ ) (𝑡) • y𝑖 and y𝑖 represent the positions of the 𝑖-th vided in Equation 6 to fit the data and extract the cuckoo at times 𝑡 ′ and 𝑡, respectively. decay parameter 𝑏gauss . • 𝛽 > 0 is the scaling factor for the step size. • 𝑀 is a randomized value from the Lévy distribution. for training recommendation algorithms. These datasets • 𝑠 is a random variable generated using a normal cover a wide range of products and media, enabling the (Gaussian) distribution, 𝑁 (0, 1). evaluation of methods across diverse contexts. Both Ama- zon and MovieLens datasets support collaborative filtering For local search (when a cuckoo’s nest is observed), the techniques, leveraging user-item interactions to identify position update is defined as: patterns. However, the results can vary across different (𝑡+1) (𝑡 ′ ) (𝑡 ′ ) (𝑡 ′ ) datasets due to the temporal dynamics and data character- y𝑖 = y𝑖 + rand ⋅ (y𝑘 − y𝑗 ) (8) istics. Additionally, the diversity in rating patterns, time where: intervals, and the inclusion of temporal features like time differences or timestamps can influence how effectively the (𝑡 ′ ) (𝑡 ′ ) • y𝑘 and y𝑗 represent randomly selected positions model adapts to the dataset’s structure. Thus, the alignment of other cuckoos at time 𝑡 ′ . between the dataset’s temporal properties and the method’s • rand is a uniformly generated random variable that design significantly impacts performance. guides the exploration process. 4.2. Evaluation Metrics 3.5.2. Optimization Steps Using Cuckoo Search To validate the proposed approaches, we employed several The Cuckoo Search optimization technique is used to de- comparison metrics, including Recall, Precision, F-measure, termine the optimal sigma (𝜎) value for each user, which is and Mean Absolute Error (MAE), alongside time complexity critical for improving the precision and recall of recommen- analysis and memory usage. These metrics are crucial for dations. Algorithm steps are in below: assessing the performance and effectiveness of recommen- 1. Data Preprocessing: Load and preprocess the dation systems. They introduced shortly in below: dataset to extract user ID, item ID, rating, and times- 1. Recall (R): Recall measures the ability of the rec- tamp. ommendation system to identify relevant items. It 2. Population Initialization: Create a diverse set of is calculated as the ratio of the number of recom- initial sigma (𝜎) values within a specified range. mended items that are also favorite items for the user 3. Lévy Flight: Update sigma values using Lévy flight to the total number of favorite items. The formula steps (Equation 7) to balance exploration and ex- for recall is given by: ploitation. 4. Fitness Function: Evaluate the performance of |𝐴(𝑖) ∩ 𝐵(𝑖)| 𝑅= (10) each sigma value using a fitness function based on |𝐵(𝑖)| precision and recall metrics: Here, 𝐴(𝑖) is the number of recommended items to 𝑓 (𝜎) = precision(𝜎) + recall(𝜎) (9) the target user 𝑢, and 𝐵(𝑖) is the number of favorite items for user 𝑢. 5. Calculate TCC: Integrate the optimized sigma val- 2. Precision (P): Precision evaluates the relevance of ues into the Time Correlation Coefficient (TCC) for- the recommended items. It is defined as the ratio of mula to compute the final coefficient. the number of recommended items that are actually 6. Generate Recommendations: Use the optimized relevant to the total number of recommended items. TCC values to generate personalized recommenda- The formula for precision is given by: tions for each user. |𝐴(𝑖) ∩ 𝐵(𝑖)| This optimization method fine-tunes the sigma (𝜎) value 𝑃= (11) |𝐴(𝑖)| for each user, enhancing precision and recall, and thereby boosting overall recommendation accuracy. By adapting 3. F-measure: The F-measure combines precision and to evolving user interests, the system remains resilient to recall into a single score, providing a balanced view changes in behavior and preferences. The personalized of the model’s performance. It is particularly useful sigma values ensure that recommendations are tailored to in imbalanced class scenarios. The formula for F- individual users. measure is: 2⋅𝑃 ⋅𝑅 𝐹-𝑚𝑒𝑎𝑠𝑢𝑟𝑒 = (12) 𝑃 +𝑅 4. Experimental Evaluation where 𝑃 is precision and 𝑅 is recall. 4.1. Datasets 4. Mean Absolute Error (MAE): MAE measures the average absolute difference between predicted and In our evaluation, we utilize three datasets to assess the actual ratings, providing a straightforward assess- proposed methods: the Amazon Phones and Accessories ment of recommendation quality. A smaller MAE dataset with 20.8M ratings, 1.3M items, and 11.6M users, indicates better recommendation quality. The MAE the Amazon Video Games dataset containing 4.6M ratings, is calculated as: 137.2K items, and 2.8M users which the timespan for these datasets is from 1996 to 2023[33], and the MovieLens dataset 𝑁 ∑𝑖=1 |𝑥𝑖 − 𝑦𝑖 | [34], which includes 20000263 ratings across 27278 movies 𝑀𝐴𝐸 = (13) 𝐴 which collect from 1995 to 2015. These diverse datasets offer a robust evaluation platform for the proposed tech- Here 𝑥𝑖 is the predicted user’s score, 𝑦𝑖 is the actual niques. They contain realworld user interactions, including user’s score, and 𝐴 is the recommended items to the ratings, users, and timestamps, providing authentic data intended user. Table 1 MAE Comparison Across Datasets Method MovieLens MAE Cellphone MAE Videogame MAE Average MAE TCC 0.8258 1.2402 1.3472 1.1377 Content-Based 0.6864 1.1744 1.2968 1.0525 Time-Based 0.8236 1.1255 1.3161 1.0884 Cuckoo Search 0.8964 1.0925 1.2952 1.0947 Decay Model 0.8204 1.2222 1.3666 1.1364 Table 2 Precision Comparison Across Datasets Method MovieLens Precision Cellphone Precision Videogame Precision Average Precision TCC 0.7604 0.7473 0.7941 0.7673 Content-Based 0.8729 0.7859 0.8283 0.8290 Time-Based 0.7622 0.7764 0.8016 0.7801 Cuckoo Search 0.7778 0.8502 0.8252 0.8177 Decay Model 0.7613 0.7567 0.8016 0.7732 5. Time Complexity: Time complexity measures the • Cuckoo Search achieves the best F-measure (0.6460), efficiency of an algorithm by analyzing how its run- improving over TCC by 22.25%. time grows as the input size increases. It is typi- cally expressed using Big-O notation to represent • Content-Based improves F-measure by 8.42%, while the upper bound of an algorithm’s growth rate. The Time-Based shows an improvement of 3.66%. general formula for time complexity is: • Cuckoo Search achieves the best recall (0.7402), im- 𝑇 (𝑛) = 𝑂(𝑓 (𝑛)) (14) proving over TCC by 10.13%. Here 𝑇 (𝑛) represents the time complexity as a func- • Content-Based and Time-Based also slightly outper- tion of the input size 𝑛. 𝑂(𝑓 (𝑛)) describes the growth form TCCF with 2.99% and 1.49% increases in Recall, rate of the algorithm (e.g., 𝑂(1), 𝑂(𝑛), 𝑂(𝑛2 ), etc.). respectively. 6. Memory Usage (Space Complexity): Memory • Content-Based achieves the highest average preci- usage refers to the amount of storage an algorithm sion (0.8290), improving over TCC by 8.04%. requires during its execution. This includes both fixed storage (e.g., constants, program code) and • Cuckoo Search also significantly improves, showing variable storage that depends on the input size. The a 6.58% increase in precision over TCC. general formula for memory usage is: 𝑆(𝑝) = 𝐴 + 𝑆(𝐼 ) (15) 5.2. Comparison based on Time Complexity and memory usage Here 𝐴 is the fixed part, such as constants or pro- gram code. 𝑆(𝐼 ) is the variable part, which depends As you can see in Table 5: on the input size 𝐼 (e.g., recursion stack, dynamic • Content-Based: More computationally efficient than memory allocation). TCC due to its focus on vector dimensions (𝑑) rather than trial steps (𝑗). In terms of memory usage is 5. Experimental Result slightly higher than TCC due to the inclusion of vector dimensions, but still manageable for smaller 5.1. comparison based on MAE, Precission, 𝑑. Recall,and F-measure • Time-Based: Simpler and faster than TCC, focusing The results of the comparison between the basic variant only on user-record pairs without trial steps. More of the TCC model and the proposed models, based on ten efficient than TCC, in memory usage, with linear recommended items, are presented for different metrics. scaling based on the number of records (𝑚). Based on given result in Table 1, 2, 3, and 4: • Cuckoo Search: Provides optimization capabilities • Content-Based achieves the lowest MAE on average for 𝜎, though computationally expensive in terms of (1.0525), outperforming TCC by 7.48% on average. iterations (𝑖) and population size (𝑝).Extremely low in memory usage, making it suitable for memory- • Time-Based also slightly improves over TCC with a constrained environments. 4.33% reduction in MAE. • Decay Model: Faster than TCC due to its focus on • Cuckoo Search does not outperform TCC for MAE decay fitting (𝑓), which is typically small.its memory consistently but is comparable. usage is moderate, scaling linearly with records (𝑚) and filtered data (𝑘). Table 3 Recall Comparison Across Datasets Method MovieLens Recall Cellphone Recall Videogame Recall Average Recall TCC 0.3119 0.9009 0.8038 0.6722 Content-Based 0.3329 0.9148 0.8291 0.6923 Time-Based 0.3273 0.9060 0.8133 0.6822 Cuckoo Search 0.4963 0.9018 0.8226 0.7402 Decay Model 0.3242 0.9066 0.8097 0.6802 Table 4 F-measure Comparison Across Datasets Method MovieLens F-Measure Cellphone F-Measure Videogame F-Measure Average F-Measure TCC 0.2865 0.6708 0.6277 0.5283 Content-Based 0.3122 0.7188 0.6870 0.5727 Time-Based 0.3007 0.6988 0.6434 0.5476 Cuckoo Search 0.4703 0.7694 0.6983 0.6460 Decay Model 0.2984 0.6811 0.6426 0.5407 5.3. Overall Comparison and Key insight • Final selection: Best Overall: Content-Based for its balance of per- • Content-Based: formance, computational efficiency, and memory Performance: Achieves the lowest MAE (7.48% lower usage. than TCC) and the highest Precision (8.04% higher), Best for Performance: Cuckoo Search for superior making it the most accurate method for prediction recall and F-Measure. tasks. Best for Simplicity: Time-Based for scalability and Time Complexity 𝑂(𝑛 ⋅ 𝑚 ⋅ 𝑑), efficient for large-scale low resource requirements. systems. Memory Usage: Slightly higher than TCC (𝑂(16𝑑 + 40𝑚)), manageable for smaller 𝑑. 6. Conclusion Best Use Case:Accuracy-driven tasks with balanced computational and memory requirements. In conclusion, this study has advanced the understanding of time-aware recommendation systems by addressing the • Cuckoo Search: limitations of existing algorithm, Time Correlation Coeffi- Performance: Achieves the highest F-Measure cient (TCC). The proposed methodologies—Content-based, (22.25% higher) and Recall (10.13% higher) than TCC, Cuckoo Search, Decay Model, and Time-Based—highlight making it ideal for applications where both accuracy the need to balance performance with computational effi- and recall are critical. ciency. Future research should focus on the generalizability Time Complexity: 𝑂(𝑖 ⋅ 𝑝 ⋅ 𝑛 ⋅ 𝑚), computationally of these methods across diverse datasets and the integration expensive. of real-world user feedback, ensuring that recommendation Memory Usage: Extremely low (𝑂(8 ⋅ (𝑝 + 5))), ideal systems continue to evolve and effectively meet users’ dy- for memory-constrained environments. namic preferences. Ultimately, these advancements aim to Best Use Case: Performance-critical applications enhance user experience and foster greater satisfaction and with sufficient computational resources. loyalty. • Time-Based: Performance: Modest improvements in F-Measure References (3.66%) and MAE (4.33%) over TCC. Time Complexity: 𝑂(𝑛 ⋅ 𝑚), the simplest and most [1] A. Y.-A. Chen, D. McLeod, Collaborative filtering for computationally efficient method. information recommendation systems, in: Encyclope- Memory Usage: Linear (𝑂(24𝑚)), highly scalable. dia of E-Commerce, E-Government, and Mobile Com- Best Use Case: Real-time or resource-constrained merce, IGI Global, 2006, pp. 118–123. environments. [2] Y. Deldjoo, Z. He, J. McAuley, A. Korikov, S. Sanner, Table 5 Complexity and Memory Usage Comparison Across Datasets Method Time Complexity Memory Usage TCC (Baseline) 𝑂(𝑘 ⋅ 𝑚 ⋅ 𝑗), 𝑛 = users, 𝑚 = user records, 𝑗 = trial steps 𝑂(𝑘 ⋅ 𝑚), 𝑛 = users, 𝑚 = user records Content-Based 𝑂(𝑘 ⋅ 𝑚 ⋅ 𝑑), 𝑑 = vector dimensions, 𝑚 = user records 𝑂(𝑘 ⋅ 𝑑 + 𝑚), 𝑑 = vector dimensions, 𝑚 = user records Time-Based 𝑂(𝑛 ⋅ 𝑚), 𝑚 = user records 𝑂(2𝑚), 𝑚 = user records Cuckoo Search 𝑂(𝑘 ⋅ 𝑚 ⋅ 𝑖), 𝑖 = iterations, 𝑝 = population size 𝑂(𝑘 ⋅ (𝑝 + 𝑓 )), 𝑝 = population size Decay Model 𝑂(𝑘 ⋅ 𝑓 ⋅ 𝑚), 𝑓 = decay fitting, 𝑚 = user records 𝑂(𝑘 ⋅ 1 + 𝑘), 𝑚 = user records, 𝑘 = filtered data A. Ramisa, R. Vidal, M. Sathiamoorthy, A. Kasirzadeh, dation systems: Principles, methods and evaluation, S. Milano, A review of modern recommender systems Egyptian informatics journal 16 (2015) 261–273. using generative models (gen-recsys), in: Proceedings [20] Z. Cui, X. Xu, X. Fei, X. Cai, Y. Cao, W. Zhang, J. Chen, of the 30th ACM SIGKDD Conference on Knowledge Personalized recommendation system based on collab- Discovery and Data Mining, 2024, pp. 6448–6458. orative filtering for iot scenarios, IEEE Transactions [3] F. Sayyed, R. Argiddi, S. Apte, Generating recommen- on Services Computing 13 (2020) 685–695. dations for stock market using collaborative filtering, [21] R. Cai, R. Lu, W. Chen, Z. Hao, Counterfactual contex- Int. J. Comput. Eng. Sci 3 (2013) 46–49. tual bandit for recommendation under delayed feed- [4] J. Bobadilla, F. Ortega, A. Hernando, A. Gutiérrez, Rec- back, Neural Computing and Applications (2024) 1–15. ommender systems survey, Knowledge-based systems [22] Y. Koren, Collaborative filtering with temporal dy- 46 (2013) 109–132. namics, in: Proceedings of the 15th ACM SIGKDD [5] R. Burke, Hybrid recommender systems: Survey and international conference on Knowledge discovery and experiments, User modeling and user-adapted inter- data mining, 2009, pp. 447–456. action 12 (2002) 331–370. [23] I. Al-Hadi, N. M. Sharef, M. N. Sulaiman, N. Mustapha, [6] N. Friedman, D. Geiger, M. Goldszmidt, Bayesian net- Review of the temporal recommendation system with work classifiers, Machine learning 29 (1997) 131–163. matrix factorization, Int. J. Innov. Comput. Inf. Control [7] E. Bartocci, E. A. Gol, I. Haghighi, C. Belta, A formal 13 (2017) 1579–1594. methods approach to pattern recognition and synthe- [24] Y. Wan, Y. Chen, C. Yan, An integrated time-aware sis in reaction diffusion networks, IEEE Transactions collaborative filtering algorithm, in: Knowledge Man- on Control of Network Systems 5 (2016) 308–320. agement in Organizations: 15th International Confer- [8] P. Lops, M. De Gemmis, G. Semeraro, Content-based ence, KMO 2021, Kaohsiung, Taiwan, July 20-22, 2021, recommender systems: State of the art and trends, Proceedings 15, Springer, 2021, pp. 369–379. Recommender systems handbook (2011) 73–105. [25] A. Hamzehei, R. K. Wong, D. Koutra, F. Chen, Collabo- [9] R. Borges, K. Stefanidis, On measuring popularity bias rative topic regression for predicting topic-based social in collaborative filtering data, in: Proceedings of the influence, Machine Learning 108 (2019) 1831–1850. Workshops of the EDBT/ICDT 2020 Joint Conference, [26] H. Wang, N. Wang, D.-Y. Yeung, Collaborative deep Copenhagen, Denmark, March 30, 2020, volume 2578 learning for recommender systems, in: Proceedings of CEUR Workshop Proceedings, CEUR-WS.org, 2020. of the 21th ACM SIGKDD international conference [10] R. Borges, K. Stefanidis, Feature-blind fairness in col- on knowledge discovery and data mining, 2015, pp. laborative filtering recommender systems, Knowl. Inf. 1235–1244. Syst. 64 (2022) 943–962. [27] H. Su, X. Lin, B. Yan, H. Zheng, The collaborative [11] K. Stefanidis, E. Ntoutsi, H. Kondylakis, Y. Velegrakis, filtering algorithm with time weight based on mapre- Social-based collaborative filtering, in: R. Alhajj, J. G. duce, in: Big Data Computing and Communications: Rokne (Eds.), Encyclopedia of Social Network Analysis First International Conference, BigCom 2015, Taiyuan, and Mining, 2nd Edition, Springer, 2018. China, August 1-3, 2015, Proceedings 1, Springer, 2015, [12] G. Adomavicius, A. Tuzhilin, Toward the next gen- pp. 386–395. eration of recommender systems: A survey of the [28] J. Bu, X. Shen, B. Xu, C. Chen, X. He, D. Cai, Im- state-of-the-art and possible extensions, IEEE trans- proving collaborative recommendation via user-item actions on knowledge and data engineering 17 (2005) subgroups, IEEE Transactions on Knowledge and Data 734–749. Engineering 28 (2016) 2363–2375. [13] D. H. Stern, R. Herbrich, T. Graepel, Matchbox: large [29] L. Xiong, X. Chen, T.-K. Huang, J. Schneider, J. G. Car- scale online bayesian recommendations, in: Proceed- bonell, Temporal collaborative filtering with bayesian ings of the 18th international conference on World probabilistic tensor factorization, in: Proceedings of wide web, 2009, pp. 111–120. the 2010 SIAM international conference on data min- [14] J. B. Schafer, D. Frankowski, J. Herlocker, S. Sen, Col- ing, SIAM, 2010, pp. 211–222. laborative filtering recommender systems, in: The [30] T. Xue, B. Jin, B. Li, W. Wang, Q. Zhang, S. Tian, A adaptive web: methods and strategies of web person- spatio-temporal recommender system for on-demand alization, Springer, 2007, pp. 291–324. cinemas, in: Proceedings of the 28th ACM Inter- [15] C.-N. Ziegler, S. M. McNee, J. A. Konstan, G. Lausen, national Conference on Information and Knowledge Improving recommendation lists through topic diver- Management, 2019, pp. 1553–1562. sification, in: Proceedings of the 14th international [31] S. Ahmadian, N. Joorabloo, M. Jalili, M. Ahmadian, conference on World Wide Web, 2005, pp. 22–32. Alleviating data sparsity problem in time-aware rec- [16] C. C. Aggarwal, C. C. Aggarwal, Time-and location- ommender systems using a reliable rating profile en- sensitive recommender systems, Recommender Sys- richment approach, Expert Systems with Applications tems: The Textbook (2016) 283–308. 187 (2022) 115849. [17] K. Stefanidis, I. Ntoutsi, K. Nørvåg, H. Kriegel, A frame- [32] S. Pang, S. Yu, G. Li, S. Qiao, M. Wang, Time-sensitive work for time-aware recommendations, in: DEXA, collaborative filtering algorithm with feature stability, volume 7447 of Lecture Notes in Computer Science, Computing and Informatics 39 (2020) 141–155. Springer, 2012, pp. 329–344. [33] Y. Hou, J. Li, Z. He, A. Yan, X. Chen, J. McAuley, Bridg- [18] K. Stefanidis, E. Ntoutsi, M. Petropoulos, K. Nørvåg, ing language and items for retrieval and recommenda- H. Kriegel, A framework for modeling, computing tion, arXiv preprint arXiv:2403.03952 (2024). and presenting time-aware recommendations, Trans. [34] F. M. Harper, J. A. Konstan, The movielens datasets: Large Scale Data Knowl. Centered Syst. 10 (2013) History and context, Acm transactions on interactive 146–172. intelligent systems (tiis) 5 (2015) 1–19. [19] F. O. Isinkaye, Y. O. Folajimi, B. A. Ojokoh, Recommen-