BinRec: addressing data sparsity and cold-start challenges in recommender systems with biclustering

Applied Intelligence (2025) 55:830 https://doi.org/10.1007/s10489-025-06725-6 1 Introduction Recommender systems (RSs) were introduced in the 90 s to support users in a given decision making situation [1]. These software tools, from the user’s point of view, reduce the search space when accessing a product catalog and process the opinions of other users. For many companies dedicated to selling products of any kind, providing multimedia content, or simply being a database of movie rankings, the RSs facilitate the user’s selection and/or purchase process while also focusing attention on a different set of less wellknown and successful products, increasing their chances of sale. The use of RSs has become essential in many applications, including social networks [2], online education [3], and Big Data ecosystem [4, 5]. Although there are various types of RSs, the most widely used approach is User Based Collaborative Filtering (UBCF) [6]. UBCF systems process a database of user ratings (’likes,"star ratings,’ or numerical data) on any kind of item (selling products, books, movies, etc.) to generate item Francisco Gómez-Vela, Aurelio Lopez-Fernandez, Miguel GarcíaTorres and Federico Divina contributed equally to this work. Domingo Rodríguez-Baena [email protected] Francisco Gómez-Vela [email protected] Aurelio Lopez-Fernandez [email protected] Miguel García-Torres [email protected] Federico Divina [email protected] 1 Computer Science, Universidad Pablo de Olavide, Cta. Utrera km 1, Seville ES-41013, Sevilla, Spain 2 Department of Computer Languages and Systems, Universidad de Sevilla, Seville ES-41004, Sevilla, Spain 3 Data Science and Big Data Lab, Universidad Pablo de Olavide, Cta. Utrera km 1, Sevilla ES-41013, Sevilla, Spain Abstract Recommender Systems help users in making decision in different fields such as purchases or what movies to watch. UserBased Collaborative Filtering (UBCF) approach is one of the most commonly used techniques for developing these software tools. It is based on the idea that users who have previously shared similar tastes will almost certainly share similar tastes in the future. As a result, determining the nearest users to the one for whom recommendations are sought (active user) is critical. However, the massive growth of online commercial data has made this task especially difficult. As a result, Biclustering techniques have been used in recent years to perform a local search for the nearest users in subgroups of users with similar rating behaviour under a subgroup of items (biclusters), rather than searching the entire rating database. Nevertheless, due to the large size of these databases, the number of biclusters generated can be extremely high, making their processing very complex. In this paper we propose BinRec, a novel UBCF approach based on Biclustering. BinRec simplifies the search for neighbouring users by determining which ones are nearest to the active user based on the number of biclusters shared by the users. Experimental results show that BinRec outperforms other state-of-the-art recommender systems, with a remarkable improvement in environments with high data sparsity. The flexibility and scalability of the method position it as an efficient alternative for common collaborative filtering problems such as sparsity or cold-start. Keywords Recomender systems · Biclustering · Collaborative filtering · Data mining Accepted: 13 June 2025 © The Author(s) 2025 BinRec: addressing data sparsity and cold-start challenges in recommender systems with biclustering DomingoRodríguez-Baena1· FranciscoGómez-Vela1· AurelioLopez-Fernandez2· MiguelGarcía-Torres3· FedericoDivina3 1 3 D. Rodríguez-Baena et al. recommendations for a specific user, answering the question: What is the most popular among people who share my tastes? As a result, they are based on the idea that users who have previously shared similar tastes will almost certainly share similar tastes in the future. To do so, a subset of users (nearest users) who are similar to the user for whom recommendations are sought (active user) is chosen, and their ratings are combined to predict unseen items. However, UBCF techniques include important issues that are still a challenge for researchers, such as sparsity, performance, and the cold-start problem [7, 8]. When rating databases are large and include a significant number of items, there are few ratings available for each user. This makes it harder when looking for nearest users because there is less information available to look for similar behaviour in terms of rating (sparsity problem). Furthermore, in these cases, UBCF approaches perform poorly in terms of scalability [9, 10]. Moreover, adding a new user to the ratings database may be problematic because this user has an empty or very small rating profile, and as a result, the system cannot accurately calculate the similarities between such cold-start users and others [11]. Thus, since the first UBCF approach was proposed by Resnick et al. [12], many papers have been published in an attempt to improve prediction results and solve the major issues associated with these techniques [13]. One of the most important tasks of UBCF processing is determining the nearest user(s) to the active one by calculating similarity measures between them. Given the huge amounts of data available, this search could be complex and time-consuming [14]. To deal with the diversity of user profiles, reduce dimensionality and making easier to find the nearest users, Clustering techniques have been used either directly or as a preprocessing stage in UBCF systems [15]. These techniques generate disjoint groups of users by taking into account the ratings of all the items. However, in real-world scenarios, there is usually a strong correlation between the preferences of subsets of users on subsets of items [16]. Thus, during the last years, UBCFs based on Biclustering have been proposed to reduce the consequences of sparsity issues in the input dataset [17]. Biclustering techniques group subsets of elements that have similarities under a subset of attributes and have been used successfully in many areas of study [18]. So, instead of performing a global search to find the nearest users to the active one, Biclustering can be used to narrow the search space. The rating database biclusters, which are subsets of users with similar rating behaviour under a subset of items, represent dense areas where the nearest users can be found in a local search. The generation of biclusters takes place in a phase prior to the recommendation, called the offline phase. Following that, and on demand, the nearest users to the active user are searched in the biclusters to generate a specific recommendation (online phase) [19]. However, due to the large number of biclusters that can be extracted from the input rating datasets, processing them to determine the similarity measures when looking for the nearest users can be challenging. Furthermore, there are several important aspects associated with the use of Biclustering. For example, which Biclustering technique applies, if the same set of biclusters can be used for every active user, how to update the biclusters when new data is added to the rating dataset or how to afford the cold-start problem. This work presents BinRec, a novel UBCF approach that uses biclusters to make the search for the nearest users more efficient in sparse databases, while also optimizing the computational complexity of such a search through a simple data structure. This structure can also be used to address the cold-start problem, which is common in rating databases. BinRec is flexible, adapting to any Biclustering technique. However, the use of the BiBit Biclustering algorithm [20] is proposed, as it allows the incremental generation of new biclusters as new users are added to the rating database. This feature, combined with the aforementioned efficiency, makes BinRec well suited for managing large-scale databases with a high number of biclusters. The main contributions of this work can be summarized as follows: ●Introduction of BinRec, a novel collaborative filtering approach that leverages Biclustering techniques to group users with similar rating preferences, thereby reducing the search space for nearest neighbours in sparse rating databases. ●BinRec reduces the computation time required to identify the nearest users to a given one based on biclusters, using a simple data structure that stores the number of biclusters shared between users, while also achieving highly accurate predictions of user recommendations. ●The cold start problem can also be addressed using the same information stored in the previously mentioned data structure. ●BinRec adapts to any Biclustering technique. However, the use of BiBit Biclustering algorithm is proposed due to its performance and the possibility of developing an incremental biclusters processing when new users are added to the rating database. All these contributions make BinRec an efficient and accurate approach that can be applied to large rating databases and a high number of biclusters generated from them, while also addressing the issues of sparsity and cold-start. To demonstrate this, the study includes experimentation and a comparison with other benchmark collaborative 1 3 830 Page 2 of 17 BinRec: addressing data sparsity and cold-start challenges in recommender systems with biclustering filtering approaches and a bicluster-based recommendation technique. The rest of the paper is organized as follows: Section 2 reviews the state of the art in FC-based recommender systems and Biclustering. In Section 3, the BinRec approach is described in detail, followed by a series of experiments and analysis of results in Section 4. Finally, Section 5 presents conclusions and possible future lines of research. 2 Background As it has been commented in the previous Section, several UBCF techniques based on Biclustering have been proposed in the last years [21]. Following, they are analyzed from different points of view. In first place, the generation of the biclusters: whether the Biclustering technique is applied to the entire dataset or only a portion of it, which Biclustering technique is used, and so on. Second, how the generated biclusters are processed to find the nearest users. Finally, how the cold-start problem is handled is an important factor to consider. In terms of bicluster generation, most approaches apply Biclustering techniques to the entire rating dataset, regardless of who is the active user [17, 22–33]. In these cases, all of the generated biclusters are used to predict any active user’s recommendations. However, if a new user and their ratings are added to the input dataset, the Biclustering technique must be applied again to include this new user. Other research works only generate biclusters that include the active user. The generation of the biclusters is thus less computationally expensive, but it must be repeated for each active user. For example, in the works [34, 35], a hierarchical Biclustering method is applied to generate biclusters in different layers consisting of co-rated items among the active one and other users. In the first layer, a bicluster is created for every item rated by the active user, including also the rest of users that have rated the same item. In the second layer, pairs of biclusters from the first layer are combined, so the subset of items increases while the subset of users decreases in the new biclusters. The process ends in a layer in which the biclusters contains the maximum number of items with the minimum number of users. Some research works are very flexible when it comes to selecting which Biclustering technique to use [23, 30–33, 36], allowing you to use whatever technique you want. Obviously, both the final results and the performance may vary depending on the option chosen. In the six previously cited works, the Biclustering algorithms selected to conduct the experimentation are Bimax [37], QUBIC [38] or XMotif [39]. In many other cases, specific Biclustering algorithms have been designed to group together subsets of users and subsets of items. For example, BIC-aiNET, proposed in [22] and used again by Desai et al. [26], is an adaptation of the Artificial Immune Network model applied to Clustering techniques [40]. In the work of Elnabarawy et al. [27], the Biclustering method BARTMAP is used [41], although the authors don‘t clarify if it is possible to use any other technique. After obtaining the biclusters, the following step is to look for the target biclusters where the nearest users will be chosen. In general, to select the target bicluster or biclusters it is necessary to calculate similarity measures between all the biclusters and the active user. This can be a costly task considering the large number of biclusters that can be generated. Some works simplify this task, just processing only the biclusters that contain the active user. For example, in [22] a residue is calculated only for those biclusters that include this user. The bicluster with the smallest residue and the item to be rated is chosen. Then, the recommendation is generated directly as an average of the ratings of this item in the bicluster. The biclusters to which the active user belongs are determined in first place in the work of Yoldar et.al. [33]. Then, he top-N recommendations are then chosen based on the joint decision of these biclusters. Another interesting example could be the work presented by Huan-huan et al. [42] where a scalable recommender system based on biclustering and moth flame optimization algorithm is proposed. Another recent study, [17] proposes a novel Biclustering method based on modified fuzzy adaptive resonance theory. This paper proposes a new measure to reflect the similarity between users that considers the effect of the number of common elements of the users. In most cases, however, a similarity calculation process is applied to all biclusters. In some occasions, this process is applied to search for the nearest users but also for the nearest items to the one to be rated. In [24], several nearest items to the item to be rated are selected from the biclusters, calculating a similarity measure in two phases. The same process is carried out for the nearest users to the active one. Then, the prediction is obtained by means of a formula that combines the previously calculated similarities. The nearest K biclusters to the active user and the target item are obtained in [30], computing the CjacMD (CosineJaccardMean Measure of Divergence) similarity measure. In the work of Symeonidis et al. [23] and Sing et al. [32], a similarity based on the number of common items is calculated to find the K nearest biclusters to the active user. Then, a ranking of the N top items is extracted from these biclusters, based on the appearance frequency of each item. In other works, the smallest bicluster for the active user, that is, the bicluster that contains this user and the greatest number of items, is selected [25, 28]. To accomplish this, all biclusters must be processed via intersection. In a second phase, the 1 3 Page 3 of 17 830 D. Rodríguez-Baena et al. The nearest users are calculated from two different sources of information in the work of Faryad et al. [52]. In a first step, a Cosine function is used to calculate the similarity between users based on the user-item ratings dataset. Secondly, demographic similarities between users are obtained through a weighted average of their demographic data. The final similarities between users are determined using a linear combination of both ratings and demographic similarities. However, in many cases, the problem of sparsity affects not only the data describing user behaviour (ratings), but also the data describing the users (e.g. personal information). To address this issue, a novel Inductive Heterogeneous Graph Neural Network (IHGNN) model is presented [53]. This model converts new users, items, and associated multimodal information into a Modality-aware Heterogeneous Graph (M-HG), which preserves the rich and heterogeneous relationship information among them. The approach presented in this paper, BinRec, reduces the complexity of processing biclusters to find the nearest users to the active one, improving performance on large databases with high sparsity while maintaining a high level of accuracy in its recommendations. To manage the coldstart issue, it is not necessary to process additional information about users, but rather the information on ratings and relationships between users obtained through biclusters is used. Also, BinRec is flexible to using any Biclustering technique, though a specific one is proposed to partially solve the problem of generating the entire set of biclusters again when new data are added to the input dataset. Finally, prior to applying the Biclustering technique, a preprocessing task is used to fine-tune the precision of the recommendations. 3 Materials and methods In this section, the new approach proposed by authors, BinRec, is fully explained. The starting hypotheses are the following: ●In a bicluster, a sub-group of users share a similar opinion about a sub-group of items, so it is a very interesting source of knowledge to generate new recommendations. ●The users with whom the same tastes are shared are those with whom a greater number of biclusters are shared. The proposal presented in this work is divided into two phases: offline and online. The offline phase comes before the recommendation and consists mainly of processing the biclusters obtained from the rating database. During the online phase, the recommendation is carried out. Furthermore, the offline and online phases are separated into two candidate set of items for a recommendation is extracted from the bicluster neighbourhood of the smallest bicluster. The Mean Absolute Difference measure is calculated for all the biclusters in [29, 43] to find the nearest bicluster that has a strong partial similarity with the preferences of an active user. Recent works also apply specific algorithms to compute the neighbourhood of the active user, such as Memetic algorithms in [34] or the Moth Flame Optimization algorithm in [35]. In [44], the biclusters are processed and mapped in a square grid to represent different states in a Markov decision problem. To do so, biclusters are sorted by their Scaling Mean Squared Residue (SMSR) values and then merged to fit and place them in the square matrix. In conclusion, as observed in the reviewed works, reducing the search space through the generation of biclusters involves a massive processing effort to find the nearest neighbours. This task can be computationally demanding, considering the large number of biclusters that can be obtained from increasingly extensive rating databases. Recent advancements in recommender systems extend beyond biclustering, particularly utilising graph-based methodologies to enhance performance in sparse or intricate user-item contexts. These approaches represent interactions as bipartite or heterogeneous graphs. For instance, LiteGSR [45] refines social graphs using PageRank-based centrality, reducing aggregation overhead in GNN-based recommendations. KGCFRec [46] integrates knowledge graphs with collaborative filtering through a dual-channel GNN and adaptive attention, outperforming baselines on multiple datasets. Similarly, the Graph Convolutional Recommendation System with Bilateral Attention [47] enhances performance by combining user-item and knowledge graphs via attention mechanisms. Beyond graph-based methods, Turbo-CF [48] offers fast recommendation through matrix decomposition-free filtering, while PolyCF [49] uses spectral graph filters to improve collaborative filtering accuracy and scalability. Finally, the users cold-start problem appears when the existing rating data about a user are insufficient, so the system cannot generate efficient recommendations. It happens whenever there are users who have not rated any item (new users) or who have rated a very few items (cold users) [50]. To address this issue, most of the proposed solutions find different ways to associate users not using rating data but other types of information. For example, in [26], the input dataset is clustered as a preprocessing step based on users’ age and location, not on their ratings. Other works, such as [51], use both explicit (ratings) and implicit data (browsing history, purchase and click activity, etc.) about users to generate recommendations. To do so, authors combine a Probabilistic Matrix Factorization based on rating data and a Bayesian Personalized Ranking based on explicit data. 1 3 830 Page 4 of 17 BinRec: addressing data sparsity and cold-start challenges in recommender systems with biclustering encoding, if user u recommends an item i, then (u, i) is equal to 1; otherwise, it is equal to 0. A detailed illustration of the offline phase is provided in Fig. 2 by means of an example. As it can be observed, the offline phase is divided into three steps: the binarization of the rating matrix, the generation of the biclusters and the creation of the Biclusters matrix. The input data of the example, A, is a 4x10 matrix in which 4 users rate 10 items using scores from 1 to 5. The A(u, i)=−1 value represents that item i has not been rated yet by user u. As it has been said before, the binarization of the rating matrix is the first step. This step is necessary in order to apply the BiBit biclustering algorithm, which operates on binary databases. The binarization process—i.e., transforming the data into ones and zeros—requires a threshold that determines whether an element will be converted to 1 or 0. Ratings databases are usually based on a list of discrete values, for example, ratings from 1 to 5 stars. Thus, the threshold will be a specific value within that list. For instance, if the threshold is set to 3, it means that all ratings equal to or greater than 3 will be converted to the value 1, and the remaining values to 0. The resulting database highlights those ratings equal to or greater than 3, which will be considered positive recommendations. So, if the possible score values of the ratings range from x to y, then for each value Sv ∈[x, y] , a new binary matrix B_Sv can be generated, in which B_Sv(u, i)=1 if in the original dataset A(u, i)≥Sv and it means that item i is recommended by user u; B_Sv(u, i)=0 otherwise. In the example of Fig. 2, Sv =2 , so a new binary matrix is created ( B_2 ). As it will be explained in Section 4, this preprocessing offers the possibility of influencing the sensitivity of the recommender system. stages (see Fig. 1). Following, every phase, along with its stages, is described. 3.1 Offline phase The goal of the offline phase is to use the local behaviour patterns, regarding user recommendations, extracted by the biclusters to generate a matrix, M, that reflects the relationships between different users based on the similarity of their ratings. This matrix will contain, for each user, the number of biclusters shared with others and will be used to quickly and easily identify the nearest users to the active one. When two users appear in the same bicluster, it means that they have recommended the same subset of items. Therefore, if they share many biclusters, it indicates a higher level of affinity. As we have seen in the state-of-the-art review, the search for nearest neighbours involves applying complex measures to all the biclusters generated for each recommendation, resulting in a high computational load. With this new proposal, a single data structure will be generated, the matrix M, which, without complex calculations, will contain the necessary information to determine the similarity between users for all the new recommendations. As it can be observed in Fig. 1, Biclustering is applied to the ratings dataset in the offline phase to generate biclusters, that is, subsets of users with similar rating behaviour under subsets of items. Although any Biclustering technique can be used in our proposal, in this work, binary Biclustering approaches have been proposed. This kind of technique adapts perfectly to the classification of recommendations based on labels: not recommended and recommended. In addition, as will be seen below, the input data can be preprocessed in various ways due to their binarization. So, when working with users and items to be rated in a binary Fig. 1 This is the general schema of the proposal. It is divided into two different phases: offline and online. Biclustering is applied to the rating dataset in the offline phase to generate a set of biclusters. A bicluster is defined as a subgroup of users with similar rating behaviour under a subgroup of items. Next, these biclusters are used to define the biclusters matrix, M, that will be used in the online phase during the selection of the nearest users. In the online phase, the nearest user selection and recommendation process take place 1 3 Page 5 of 17 830 D. Rodríguez-Baena et al. of bits generated as a result of the AND operation between these rows. The potential bicluster is then derived from this pair of rows and the columns in the pattern that are equal to 1. Finally, new rows are added to the bicluster if they are compatible with that pattern. So, if a new user is added to the rating dataset, for example U5 (see Fig. 3), it will not be necessary to apply BiBit to the whole new binary matrix. Instead, Bibit will be partially applied, just creating new patterns with the row U5 and the rest of rows (new bicluster Bic6) and checking if that row is compatible with The procedure for generating biclusters is the second step. In the example, five biclusters have been extracted from the binary matrix. The Bibit Biclustering algorithm was used to accomplish this [20]. Several reasons are given: First and foremost, Bibit performs great with binary matrices of all sizes and shapes [54]. Besides, when a new user is introduced to the ratings dataset, Bibit offers the option of developing an incremental biclusters processing rather than having to process the entire input dataset again. BiBit generates a pattern for each pair of rows, which is a group Fig. 3 In the example, a new user U5 has been added to the input rating dataset A (enhanced in red). The binary Biclustering algorithm BiBit is applied only this new row. Thus, Bibit generates a new pattern with that row and every one of the previous rows and from these patterns a new bicluster is created (Bicluster Bic6). At the same time, Bibit checks if the new row is compatible with the patterns obtained in previous executions, modifying the existing biclusters if necessary (changes in red in biclusters Bic1, Bic2 and Bic5) Fig. 2 In the first step of the offline phase, the input ratings dataset, A, is binarized using a threshold value ( Sv =2 in the example). Then, a new binary version of A is created, B2 . In the second step, a binary Biclustering algorithm (BiBit) is applied to B2 , extracting a group of biclusters. For example, the Bicluster Bic1 is composed by two users, U1 and U2, both recommending the items I4 and I10. Finally, in the third step, a users square matrix, M, is created from the biclusters generated 1 3 830 Page 6 of 17 BinRec: addressing data sparsity and cold-start challenges in recommender systems with biclustering shares a concrete number of biclusters with UJ, represented by the weight of that edge. In our example, U2 is connected with the rest of users and the sum of its edges represents the highest value, 5. Therefore, U2 will be chosen to represent the rating dataset’s overall opinion and his/her recommendations will be assign to the new users. 3.2 Online phase The online phase is divided into two steps. The active user’s nearest users are determined in the first step. The recommendation process is then carried out based on the nearest users. Following, the examples of Figs. 5 and 6 are used to clearly explain the online phase. BinRec, unlike most Biclustering-based recommendation proposals (see Section 2), employs a simple measure based on the M matrix rather than calculating complex similarity measures that entail processing all of the bicluster elements. Instead, it is assumed that a user who shares a large number of biclusters with another user can determine that his or her opinion is more important than other users’ opinions. Let’s suppose that the active user is U2. To find its nearest users, the column # of M, which contains the average of the number of biclusters that a specific user shares with the rest, is used (see Fig. 5). To calculate that average, only those users with whom every user shares bicluters are taken into account. So, the nearest users of U2 are those with whom U2 shares a number of biclusters greater or equal than that average (1.6), that is, U3 (he/she shares 2 biclusters with U2) and U4 (he/she shares 2 biclusters with U2). As it can be observed, the user U1 is discarded because he/she only shares 1 bicluster with U2. Next, the potentially recommendable items are selected, that is, those that the active user had not rated up to that moment: I1, I2, I6 and I8. To know which of these items will be recommended, the patterns created in previous executions (modifications in Bic1, Bic2 and Bic5). Next, in the third step, a Biclusters matrix M is created. M is a square matrix and its dimension is the number of users in the database. The value of M(I, J) reflects the number of biclusters shared by users UI and UJ, whereas the main diagonal represents the number of biclusters in which every user UX occurs (see Fig. 2). This matrix fulfils two important functionalities (see Section 3.2): M will be used in the online phase during the nearest user selection process and M also be used to determine the user who appears in the largest number of biclusters. This user will be used in the online phase to address the cold-start issue [55]. The hypothesis is that the user who appears in the highest number of biclusters is the one that brings together the greatest number of similar tastes with the rest of the users. In Fig. 4, a graph generated from M is shown. In this graph, every node is a user and every edge, from UI to UJ, determines that UI Fig. 5 The first stage of online phase is represented in this image, in which the nearest users to the active user, U2, are determined using the Biclusters Matrix M. M stores, for every user, the number of biclusters shared with the rest of user. The last column, # , stores the average of the number of biclusters shared for every user. In the case of U2, the user shares biclusters with U1 (1 bicluster), U3 (2 biclusters) and U4 (2 biclusters). Since the nearest users are those with whom a user shares a number of biclusters equal to or greater than the average (1.6 in the case of U2), the U3 and U4 users are selected. In the input dataset A, the potentially recommendable items from U2 are highlighted in red Fig. 4 This figure shows a graph generated from matrix M. Every node is a user and every edge, from UI to UJ, determines that UI shares a concrete number of biclusters with UJ, represented by the weight of that edge. U2, as it can be observed, is the user which appears in the largest number of biclusters 1 3 Page 7 of 17 830 D. Rodríguez-Baena et al. in common with a certain user, the more valuable their opinion will be in generating recommendations. Furthermore, if the Biclustering technique is integrated into the BinRec framework, this data structure can be generated in parallel with the biclusters. Besides that, the recommendation process for a specific user can be carried out independently, resulting in significant scalability. As a result, because the search for nearest users is executed during the online phase (on demand), the purpose of this proposal is to reduce processing complexity. Finally, the cold-start problem is solved by taking advantage of the information stored in the biclusters matrix M. In the following section, the way BinRec behaves under different conditions is analysed. Besides, a comparison between the new approach and existing UBCF methods is carried out. 3.3 Experimentation Following, the methodology used in the experimentation, along with the description of the datasets and measures used, are introduced. 3.3.1 Experiments workflow Every experiment follows the general schema shown in Fig. 7 and is divided into 4 different phases. The input consists of a rating dataset A and a threshold value for binarization, Sv, chosen from the dataset’s range of possible scores, with extreme values discarded. In our case, because all of the datasets used in the experimentation have scores ranging from 1 to 5, Sv can be 2, 3, or 4 (1 and 5 are discarded). Following the selection of Sv, the training and test datasets are generated. To do so, a subset of the input dataset is chosen the second step is carried out (see Fig. 6). It is an iterative process in which a new rating value, Final Rec, is generated for each item selected in the previous step by calculating the average of the rating values from the nearest users. The rating values equal to −1 are ignored. Finally, to decide if an item is recommended or not, the new rating value must be equal or greater to the score value used in the offline phase to transform the input matrix A into a binary matrix, B_Sv ( Sv =2 in our example). The new rating values for active user U2 are generated in the example of Fig. 6. In the case of item I1, the average of the opinions of the nearest users, A(U3,I1) = 5 and A(U4,I1) = 1 , is used to generate a new rating value: Final Rec = (5 + 1)/2=3>=Sv . As this new value is greater than Sv =2 , the item I1 is considered as recommended. On the contrary, in the case of I2, the user U3 has not rated yet this item, A(U3,I2) = −1 , so its opinion is not taken into account. Then, the new rating value is Final Rec = A(U4,I2)/1=1/1=1< Sv , so I2 is not recommended. Finally, the cold-start issue is addressed. When a new user is added to a recommendation system, commonly it has not items rated yet. This is called a cold-start problem [51] and it implies that this new user will not be part of any bicluster. In these cases, the solution proposed by this work is to use the recommendations from the most connected user, that is, the user that appears in the largest number of biclusters (see Section 2). In our example, U2 is the most connected user, so it is selected as representative of the opinions of the rest of users. As a conclusion, the use of the bicluster matrix, M, implies avoiding the calculation of a complex similarity measure between the active user and all the biclusters generated. The idea behind it is that the more biclusters you have Fig. 6 The second step of the online phase is represented in this image. For every item not rated yet by the active user U2, a new rating value, Final Rec, is generated by the average of the rating values of the nearest users (U3 and U4). The items not rated are ignored. If the new rating value is greater or equal to Sv, the item is recommended 1 3 830 Page 8 of 17 BinRec: addressing data sparsity and cold-start challenges in recommender systems with biclustering 6040 users about 39523 movies (1 million ratings). The CiaoDVD dataset has 278,483 ratings on 99,746 items provided by 7,375 users. Also, these datasets have been used in the experimentation because of their high sparsity level. Sparsity refers to the phenomenon that occurs when the number of users and items in a ratings database is very large, but the number of ratings given by users to items is very low. Therefore, a ratings database with a sparsity level of X% only includes that percentage of all possible ratings. Specifically, Fig. 8 shows the level of sparsity for the three datasets. It can be seen that CiaoDVD has the highest value (99,9%). These levels of sparsity make them suitable for measuring the performance of the proposals presented in the context of recommendation systems. 3.3.3 Performance metrics A recommender system’s performance is typically assessed using two types of metrics: label-based or value-based [58]. In classification, label-based measures are used to determine the precision with which a class is assigned. In our case, there are two classes of items: recommended by a user and not recommended by a user. Thus, like in similar works [35], Precision and Recall measures are used in this paper. To compute these metrics, a confusion matrix must be created [59], in which the four different cases are considered: ●TP (true positive): The algorithm correctly recommends an item I that is actually relevant to the user U. and stored in the format User - Item - Value (Test dataset, T). The criteria for that selection are as follows: the elements are chosen at random, with the idea that the final selection must be balanced, with half based on items recommended by users (rating value equals or exceeds Sv) and the other half based on items not recommended by users (rating value less than or equal to Sv). Then, the rating values of T are replaced in the input dataset with the value −1 and next A is binarized using Sv as a threshold, resulting in the training dataset A′ . To determine the number of elements in the test dataset, the 20% of the total recommendations present in the original database ha been used. Following, the BiBit binary Biclustering algorithm is applied to the training dataset A′ ( see justification in Section 2). In phase 3, new rating values for the pairs User - Item stored in T are predicted. Finally, in phase 4, these new rating values are compared to those stored in T, and several performance measures are obtained to assess the quality of the new recommendations. 3.3.2 Datasets description The datasets used in the experimentation are the MovieLens 100 K and 1M datasets [56] and the CiaoDVD dataset [57]. These three datasets are considered the standard datasets in evaluating the recommendation techniques. They include the user’s ratings for movies using the 5-point rating scale; that is, the score value 5 is highly liked, and the score value 1 is most disliked. The dataset MovieLens_100K is composed of ratings from 943 users about 1682 movies (100,000 ratings). The dataset MovieLens_1M includes the opinion of Fig. 7 This is the general schema followed by all the experiments. It is divided into four different phases. The input consists of the input ratings dataset, A, and the threshold value for binarization, Sv. The training ( A′ ) and test (T) datasets are generated in the first phase. Then, biclusters are generated and recommendations are predicted based on the test dataset using the novel BinRec methodology. Finally, several performance measures are calculated in phase 4 1 3 Page 9 of 17 830 D. Rodríguez-Baena et al. 4. Khadija A, Almohsen H (2015) Recommender systems in light of big data. Int J Electric Comput Eng 5 5. Zhang Q, Lu J, Jin Y (2021) Artificial intelligence in recommender systems. Complex Intell Syst 7:439–457 6. Fkih F (2022) Similarity measures for collaborative filteringbased recommender systems: Review and experimental comparison. J King Saud Univer-Comput Inf Sci 34:7645–7669 7. Saifudin I (2024) & Widiyaningtyas, T. Approach, problem, evaluation techniques, datasets. IEEE Access, Systematic literature review on recommender system 8. Nanthini M, Pradeep Mohan Kumar K (2022) Cold start and data sparsity problems in recommender system: A concise review, 107–118 (Springer) 9. Ismai S, Özlem N, Özgür U (2013) Cluster searching strategies for collaborative recommendation systems. Inf Process Manage 49:688–697 10. Heidari N, Moradi P, Koochari A (2022) An attention-based deep learning method for solving the cold-start and sparsity issues of recommender systems. Knowl-Based Syst 256:109835 11. Tahmasebi F, Meghdadi M, Ahmadian S (2021) A hybrid recommendation system based on profile expansion technique to alleviate cold start problem. Multimed Tools Appl 80:2339–2354 12. Resnick P, Iacovou N, Suchak M, Bergstrom P, Riedl J (1994) Grouplens: An open architecture for collaborative filtering on netnews, 175–186 13. Ko H, Lee S, Park Y, Choi A (2022) A survey of recommendation systems: recommendation models, techniques, and application fields. Electronics 11:141 14. Shen J, Zhou T, Chen L (2020) Collaborative filtering-based recommendation system for big data. Int J Comput Sci Eng 21:219–225 15. Papadakis H, Papagrigoriou A, Panagiotakis C, Kosmas E, Fragopoulou P (2022) Collaborative filtering recommender systems taxonomy. Knowl Inf Syst 64:35–74 16. Sim K, Gopalkrishnan V, Zimek A e a (2013) A survey on enhanced subspace clustering. Data Min Knowl Disc 26:332–397 17. Sun J, Zhang Y (2022) Recommendation system with biclustering. Big Data Mining Anal 5:282–293 18. Madeira S, Oliveira A (2004) Biclustering algorithms for biological data analysis: a survey. IEEE/ACM Trans Comput Biology Bioinf 1:24–45 19. Castanho E N, Aidos H, Madeira S C (2024) Biclustering data analysis: a comprehensive survey. Brief Bioinf 25:bbae342 20. Rodriguez-Baena DS, Perez-Pulido A, Aguilar-Ruiz JS (2011) A biclustering algorithm for extracting bit-patterns from binary datasets. Bioinformatics 27:2738–2745 21. Silva G, Madeira C, Rui S (2024) A comprehensive survey on biclustering-based collaborative filtering. ACM Comput Surv 56. h t t p s : / / d o i . o r g / 1 0 . 1 1 4 5 / 3 6 7 4 7 2 3 22. de Castro P, de Franca F, Ferreira H, Von Zuben F (2007) Applying biclustering to perform collaborative filtering, 421–426 23. Symeonidis P, Nanopoulos A, Papadopoulos A, Manolopoulos Y (2008) Nearest-biclusters collaborative filtering based on constant and coherent values. Inf Retrieval 11 24. Zhang D et al (2014) Cold-start recommendation using bi-clustering and fusion for large-scale social recommender systems. IEEE Trans Emerg Topics Comput 2:239–250 25. Alqadah F, Reddy C, Hu J, Alqadah H (2015) Biclustering neighborhood-based collaborative filtering method for top-n recommender systems. Knowl Inf Syst 44:475–491 26. Desai T, et al (2016) An enterprise-friendly book recommendation system for very sparse data, 211–215 27. Elnabarawy I, Wunsch D, Abdelbar A (2016) Biclustering artmap collaborative filtering recommender system, 2986–2991 deep learning and rational techniques for improving personalization, although these methods tend to have a higher demand on computational resources. In this context, the use of biclusters to reduce the search space and facilitate recommendation in sparse environments is shown to be scalable. In addition, there is a growing trend towards the creation of hybrid RSs that combine the benefits of Biclustering with matrix factoring and deep learning models, which could offer a balance between accuracy and efficiency in future recommender systems. Author contributions Domingo S. Rodríguez-Baena was responsible for the conception and design of the study. Material preparation, data collection and analysis were performed by Domingo S. RodríguezBaena. All authors wrote the first draft of the manuscript and commented on earlier versions of the manuscript. All authors read and approved the final manuscript. Funding Funding for open access publishing: Universidad Pablo de Olavide/CBUA. Data availability All data used is available at the following url: h t t p s : / / g r o u p l e n s . o r g / d a t a s e t s / m o v i e l e n s / and h t t p s : / / w w w . c s e . m s u . e d u / t a n g j i l i / d a t a s e t c o d e / t r u s t s t u d y . h t m BinRec source code is available at h t t p s : / / g i t h u b . c o m / d s r o d b a e / B i n R e c. Declarations Ethical and informed consent for data used The authors have no conflicts of interest or ethics. The data used in the document are public and accessible for research purposes. Conflicts of interest The authors declare that they have no conflicts of interest. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit h t t p : / / c r e a t i v e c o m m o n s . o r g / l i c e n s e s / b y / 4 . 0 /. References 1. Sharma R, Singh R (2016) Evolution of recommender systems from ancient times to modern era: A survey. Indian J Sci Technol 9 2. Davoodi E, Kianmehr K, Afsharchi M (2013) A semantic social network-based expert recommender system. Appl Intell 39:1–13 3. Ahmad HK, Qi C, Wu Z, Muhammad BA (2023) Abine-crs: course recommender system in online education using attributed bipartite network embedding. Appl Intell 53:4665–4684 1 3 830 Page 16 of 17 BinRec: addressing data sparsity and cold-start challenges in recommender systems with biclustering 49. Qin Y, Ju W, Gu Y, Zhang M (2025) Polycf: Towards optimal spectral graph filters for collaborative filtering. ACM Trans Inf Syst (TOIS) 43:1–25 50. Roy D, Dutta M (2022) A systematic review and research perspective on recommender systems. J Big Data 9 51. Feng J, Xia Z, Feng X, Peng J (2021) Rbpr: A hybrid model for the new user cold start problem in recommender systems. KnowlBased Syst 214 52. Faryad T, Majid M, Sajad A (2021) A hybrid recommendation system based on profile expansion technique to alleviate cold start problem. Multimed Tools Appl 80:2339–2354 53. Cai D, Qian S, Fang Q, Hu J, Xu C (2023) User cold-start recommendation via inductive heterogeneous graph neural network. ACM Trans Inf Syst 41 54. Lopez-Fernandez A, Rodriguez-Baena D, Gomez-Vela F, Divina F, Garcia-Torres M (2021) A multi-gpu biclustering algorithm for binary datasets. J Parallel Distrib Comput 147:209–219 55. Sciarrone F, Yadav U, Duhan N, Bhatia K (2020) Dealing with pure new user cold-start problem in recommendation system based on linked open data and social network features. Mobile Inf Syst 2020 56. Harper F, Konstan J (2015) The movielens datasets: History and context. ACM Trans Interact Intell Syst 5 57. Guo G, Zhang J, Thalmann D, Yorke-Smith N (2014) Etaf: An extended trust antecedents framework for trust prediction, 540–547 58. Desai M, Ansari N (2023) An innovative method to increase agricultural productivity using machine learning-based crop recommendation systems, IEEE, 645–651 59. Hasnain M et al (2020) Evaluating trust prediction and confusion matrix measures for web services ranking. Ieee Access 8:90847–90861 60. Kuanr M, Mohapatra P (2021) Assessment methods for evaluation of recommender systems: a survey. Found Comput Dec Sci 46:393–421 61. Hug N (2020) Surprise: A python library for recommender systems. J Open Source Softw 5:2174 62. Ukey N et al (2023) Survey on exact knn queries over highdimensional data space. Sensors 23:629 63. Jain K, Jindal R (2023) Sampling and noise filtering methods for recommender systems: A literature review. Eng Appl Artif Intell 122:106129 64. Yan Y, Moreau C, Wang Z, Fan W, Fu C (2024) Transforming movie recommendations with advanced machine learning: A study of nmf, svd, and k-means clustering, IEEE, 178–181. 65. Airen S, Agrawal J (2023) Movie recommender system using parameter tuning of user and movie neighbourhood via co-clustering. Procedia Comput Sci 218:1176–1183 Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. 28. Chandralekha M, Saranya K, Sudha-Sadasivam G (2016) Biclustering based collaborative filtering algorithm for personalized web service recommendation. Int J Comput Appl 142:18–24 29. Virnala S, Vivekanandan K (2018) A novel biclustering with mean absolute difference similarity measure for collaborative filtering recommender system. Int J Pure Appl Math 118 30. Kant S, Mahara T (2018) Nearest biclusters collaborative filtering framework with fusion. J Comput Sci 25:204–212 31. Choi S, et al (2018) Reinforcement learning based recommender system using biclustering technique 32. Singh M, Mehrotra M (2018) Impact of biclustering on the performance of biclustering based collaborative filtering. Expert Syst Appl 113:443–456 33. Yoldar M, Özcan U (2019) Collaborative targeting: Biclusteringbased online ad recommendation. Electron Commerce Res Appl 35 34. Bansal S, Baliyan N (2020) Bi-mars: A bi-clustering based memetic algorithm for recommender systems. Appl Soft Comput 97 35. Wu H, Ke G, Wang Y, Chang Y (2022) Prediction on recommender system based on bi-clustering and moth flame optimization. Appl Soft Comput 120 36. Kataria S, Batra U (2023) Implementation of nearest co-cluster collaborative filtering using a novel similarity index. Indian J Sci Technol 16:2204–2216 37. Prelić A, et al (2006) A systematic comparison and evaluation of biclustering methods for gene expression data. Bioinformatics 1 38. Li G, Ma Q, Tang H, Paterson A, Xu Y (2009) QUBIC: a qualitative biclustering algorithm for analyses of gene expression data. Nucleic Acids Res 37 39. Murali T, Kasif S (2003) Extracting conserved gene expression motifs from gene expression data 8:77–88 40. Abbass H A, Sarker R A, Newton C S (2001) ainet: An artificial immune network for data analysis, 231–260 41. Rui X, Donald C (2011) Bartmap: A viable structure for biclustering. Neural Netw 24:709–716 42. Wu H-H, Ke G, Wang Y, Chang Y-T (2022) Prediction on recommender system based on bi-clustering and moth flame optimization. Appl Soft Comput 120:108626 43. Saravanan S, Britto A, Prabin S (2023) Novel k means biclustering fusion based collaborative recommender system. Cloud Data Sci 124:607–616. h t t p s : / / d o i . o r g / 1 0 . 4 0 2 8 / p - 5 1 0 v f d 44. Arta I, et al (2024) A reinforcement learning recommender system using bi-clustering and markov decision process. Expert Syst Appl 237 45. Ga S, Cho PH, Moon GE, Jung S (2025) Efficient gnn-based social recommender systems through social graph refinement. The J Supercomput 81:1–24 46. Peng J et al (2024) Kgcfrec: Improving collaborative filtering recommendation with knowledge graph. Electronics 13:1927 47. Yang H, Yang C (2024) Graph convolutional recommendation system based on bilateral attention mechanism. J Eng 48. Park J, Shin Y, Shin W (2024) Turbo-cf: Matrix decompositionfree graph filtering for fast recommendation, 1234–1243 (ACM) 1 3 Page 17 of 17 830