Recommendation systems help users navigate vast amounts of data, with bundle recommendation systems enhancing personalization and customized experience by grouping related items. However, many existing methods overemphasize relevance, leading to repetitive suggestions and user fatigue. This paper introduces two novel bundling methods-Bundle Partition and Bundle Function-designed to balance both diversity and relevance. These methods were evaluated using Amazon datasets on the Appliances, All_Beauty, and Luxury_Beauty categories. Results show a significant increase in diversity, as measured by Intra-List Diversity (ILD), while maintaining high relevance through average ratings. Furthermore, the novelty, assessed via Mean Inverse User Frequency (MIUF), indicates that these methods ofer a fresh and relevant experience. These findings emphasize the importance of diversity in enhancing user engagement.
In many recommendation contexts, particularly in online
shopping and travel package suggestions, users often
prefer to purchase a collection of items rather than a single
product. Therefore, recommending a set of related items
collectively, rather than individually, is more efective. This
strategy, known as bundle recommendation, involves
suggesting groups of complementary items to enhance
decisionmaking, align with real-world buying behavior, and boost
both satisfaction and sales [
A significant advancement in recommendation system
development is the incorporation of diversification into the
recommendation process [
This paper introduces a hybrid bundle recommendation approach that balances relevance and diversity by integrating collaborative and content-based filtering. It predicts user preferences through collaborative filtering and refines recommendations using item features. The approach includes two diversity-aware bundling methods: Bundle Partition, which selects diverse items aligned with user interests; and Bundle Function, which ensures both user relevance and variation among items. Utilizing NLP techniques to calculate item similarities, this method enhances recommendation quality by reducing redundancy.
The proposed methods are evaluated using real-world
datasets from Amazon’s Appliances, All_Beauty, and
Luxury_Beauty, whose extensive metadata, including product
descriptions, categories, and user ratings, enabled advanced
natural language processing (NLP) analysis [
Overall, the main contributions of this work are : (i) a hybrid model that balances relevance and diversity in bundle recommendations using collaborative filtering and contentguages and Analytical Processing of Big Data, co-located with EDBT/ICDT
CEUR
ceur-ws.org based techniques; (ii) the use of NLP to analyze item features, providing more content-rich and diverse bundle recommendations compared to existing user-centric or budget-focused models; (iii) an evaluation of the proposed bundling methods using ILD and MIUF metrics to illustrate the impact of diversity and novelty on user engagement.
Bundle Recommendations. In e-commerce, users often
purchase multiple items, making bundle recommendations
essential for suggesting sets of products rather than
individual ones [
This article presents techniques for generating diverse and relevant bundles, evaluated using metrics like ILD and MIUF. Unlike approaches that balance relevance and diversity with budget constraints—potentially compromising eficiency or diversity—this method optimizes novelty and user satisfaction while reducing computational costs. It achieves this through dynamic similarity-based bundling, randomized partitions, and strategic item selection.
In this paper, we address the challenge of creating product bundles that balance diversity and relevance based on user preferences, aiming to reduce redundancy and enhance the user experience. we propose a hybrid approach to ensure both relevance and diversity while keeping relevant items within each bundle. The method combines collaborative ifltering using SVD to identify user interests and a contentbased approach to select suitable items for each bundle. Determining User Preferences. First, we identify user preferences to ensure bundles align with individual interests. The SVD algorithm is used to generate personalized recommendations based on interaction data, and the top-rated item is selected as the ”target” item for bundling. Bundling involves selecting items that not only align with user preferences and but also add value through diversity. After choosing the target item, additional items are selected based on distinct features to ensure variety and relatedness, aiming to create a well-rounded bundle that avoids redundancy and enhances user satisfaction. For example, if a user’s toprated item is a smartphone, the bundle may include a phone case, screen protector, or wireless earbuds. These items are selected based on the user’s interest in technology (from the SVD-based analysis) and are diverse enough to ofer a broader experience. This prevents repetition and ensures each item adds value in a diferent way. The process uses content-based filtering to assess item features, maximizing diversity within the bundle.
Computing Similarities. Once the target item is identified, the next step is to locate items that share similar features to enhance recommendation relevance and user satisfaction. Each item in the dataset can be represented by a feature vector f = [ 1 , 2 , … , ], where each represents a specific feature, such as brand or category. The similarity between the target item and another item is calculated using a similarity function, sim(f , f ). Items with the highest similarity scores are chosen, ensuring that the recommendations align closely with user preferences.
Finding similar items is crucial for creating efective product bundles, as it ensures relevance and increases the likelihood of high user ratings, enhancing engagement and satisfaction. To calculate item similarity, metadata is processed and vectorized using NLP techniques like TF-IDF, which converts text into numerical vectors, assigning greater importance to key terms. Cosine similarity is then used to measure the similarity between items by calculating the cosine of the angle between their feature vectors. Items with high similarity scores are considered closely related to the target item and are selected as potential recommendations, ensuring relevance and higher user satisfaction.
This section outlines three methods for forming product bundles. The first method focuses on item similarity, grouping highly similar items with user preferences without considering diversity. The second method introduces diversity by selecting a mix of related but varied items, ensuring a balance between relevance and diversity using pairwise dissimilarity and randomization. The third method aims to maximize intra-bundle diversity by choosing items that differ from both the target item and each other, providing a broader set of recommendations to enhance user experience. Similarity-based Bundling. The similarity-only bundling method creates product bundles based on items similar to those previously liked by the user, assuming similar items will be well-received. This method serves as the base model to compare with two diversity-aware models. The algorithm generates a list of items similar to the target item, identified from the user’s preferences, by comparing it with other items in the metadata. While this approach ensures relevance, the resulting bundles may lack variety, leading to repetitive suggestions. Despite this, it provides a useful baseline for comparing more diverse bundling strategies. Partition and Randomization Method. Formally, let represent the target item, and let = { 1, 2, … , } be the set of items similar to , determined based on the cosine similarity. The objective is to find items within that maximize dissimilarity to and place them in a list : = { ∈ ∣ maximize dissimilarity( , )} . This way, we ensure that the final bundle includes varied items that still reflect the user’s preferences, reducing redundancy.
After constructing the list , the items are shufled and divided into partitions. A random selection is made from each partition to add unpredictability, increasing novelty while maintaining relevance. This approach guarantees a fresh combination of items for each bundle, resulting in a dynamic and engaging recommendation process.
To measure diversity, TF-IDF vectors of item features are used, and Euclidean distance serves as the metric. This helps maintain a balance between similarity and diversity in the item list. For two points P = ( 1, 2, … , ) and Q = ( 1, 2, … , ) in a 2-dimensional space, the Euclidean distance is: ( P, Q) = √( 1 − 1)2 + ( 2 − 2)2. Euclidean distance helps identify items that are both relevant and diverse. This is an ideal measure in this work due to its simplicity and efectiveness in distinguishing diverse items, especially when using data like TF-IDF vectors. After generating the list, the items are shufled and divided into segments. One item is randomly selected from each segment: diversity and unpredictability in the final bundle. = { ∣ ∈ random(segment )}. This approach ensures
Algorithm 1 begins by retrieving and filtering texts of similar items, then computes TF-IDF vectors for these texts and the target item. It calculates Euclidean distances between the target item and similar items, shufles the list of similar items, and divides it into partitions. One item is randomly selected from each partition to form bundles, with the first bundle including the target item.
Bundle Function Method. The Bundle Function method aims to curate bundles by strategically selecting items that are distinct from each other while still aligning with user preferences. Items similar to the target item are identified from a pre-constructed list = { 1, 2, … , }, using precomputed similarities, like cosine similarity. The goal is to create a bundle from list , ensuring each successive item is as dissimilar as possible to previously selected items. This is achieved by calculating Euclidean distances between their feature vectors. Let f and f represent feature vectors of items and in . The Euclidean distance ( f , f ) is − )2. The algorithm segiven by: ( f , f ) = √
∑=1 ( lects items with the largest Euclidean distances to ensure variety: = { 1, 2, … , } ∣ maximize ( f , f ), ∀, ∈ , ≠ This selection ensures that the items within the bundle are . not just variations of the same product, but instead represent diverse choices that cater to user preferences.
Algorithm 2 analyzes target items and finds similar ones dle_idx ← [ _[0]] 3: Create a dictionary asin_idx mapping ASINs to their indices in
6: for each ∈ {1, 2, … , } do Start the bundle with the most similar item, bun
_[1 ∶ 10: while |bundle_idx| < and _ ≠ ∅ do
← [euclidean( last, [ idx]) ∀ idx ∈ _] dists Identify index of max distance: max_dist_idx ← Add _[
max_dist_idx] to bundle_idx and remove 17: Convert bundle_idx to ASINs using and append to arg max(dists) it from _ 16: end while
bundles 18: return bundles
The 2018 Amazon dataset provides rich user-item interactions and detailed metadata, making it valuable for recom14: for each bundle (10 total bundles) to select items do if it is the first bundle then
else end if 22: end for 23: return list of bundles
users, ratings, and timestamps, with 371,345, 5,722,988, and 602,777 ratings for the All_Beauty, Luxury_Beauty, and Appliances categories, respectively. The metadata file includes product information like title, features, description, price, brand, and category. The All_Beauty and Luxury_Beauty categories, with 32,992 and 12,308 products, were combined as the beauty dataset, while the Appliances dataset, with 30,459 products, was also analyzed.
Evaluating Diversity. In experiments, the Intra-List
Diversity (ILD) metric [
In Table 1, you can see the improvement in diversity by
the use of the methods Bundle Partition and Bundle Function
for beauty and appliances. The Bundle Partition method
consistently shows high ILD values ranging from 0.93 to
0.98, with an average of 0.96, indicating that it efectively
introduces diversity and prevents redundancy in the bundles.
Similarly, the Bundle Function method achieves high ILD
scores ranging from 0.87 to 0.99, with an identical average
of 0.96, suggesting that both methods are equally efective
in ensuring item diversity and enhancing user engagement.
Evaluating Relevance. The Average Rating (AVGr)
represents the mean of ratings for items within a bundle. In
recommendation systems, each item receives a rating, either
from user feedback or predictive algorithms. Based on the
idea in [
Variance of Ratings (VAR) measures the spread of ratings 2 indicate that increasing diversity does not compromise relevance or perceived quality.
Evaluating Novelty. The Global Long-Tail Novelty [
Bundles Appliances Dataset defined as: IUF = − log2 ( || | | ), where | | is the number of users interacted with , and | | the number of users in the system. For the average novelty of the recommended items, the Mean Inverse User Frequency (MIUF) is calculated by averaging the IUF values of all items in the recommenda∑∈ log2 ( || | | ), where is the set of
The MIUF score measures the novelty of recommended items by assessing how uncommon they are across the user base. To evaluate novelty, its distribution is analyzed, revealing that 90% of items have a MIUF below 13.33. All bundles generated by both methods meet or exceed this threshold, indicating their relative novelty. The Bundle Partition method achieves MIUF values between 13.00 and 18.54, demonstrating significant novelty, while the Bundle Function method shows even stronger novelty with MIUF values ranging from 15.50 to 18.54. Table 3 presents the novelty scores for each dataset.
Discussion. Bundle Partition and Bundle Function enhance the recommendations’ diversity and novelty while maintaining high relevance. ILD scores for both methods are high in All_Beauty and Luxury_Beauty (0.93-0.98) and slightly lower in Appliances (0.55-0.98), indicating varied recommendations. For relevance, AVGr and VAR demonstrate users’ preferences alignment, with AVGr scores up to 4.0. Novelty, measured by MIUF, is highest for the Bundle Function method, showing that the methods provide user-relevant bundles that outperform the Similarity-Based approach.
In this paper, we design, develop, and evaluate two bundling methods, alongside a baseline solution. Our goal is to improve product bundle recommendations by balancing relevance and diversity. We implemented a hybrid approach combining collaborative and content-based filtering, using NLP to analyze item features. Both methods successfully introduced diversity without sacrificing relevance, achieving promising results in maintaining high ratings and enhancing the overall diversity of recommendations.