Content-Based Item Image Bundle Recommendation

With the rapid development of the Internet industry and its e-commerce platforms,the massive amount of commodities has caused the problem of information overload.The recommendation system can quickly find the information that users are interested in from a large amount of data according to different user preferences,and solve the problem of information overload,so it has become one of the important topics of applied computer science.Traditional single-item recommendation systems usually only focus on users’ different preferences for items,while ignoring the compatibility relationship between items.More often than not,users not only care about their favorite items,but also hope to get recommendations on comptible items.For example,when a user likes a certain top,the recommendation system needs to consider the recommendation of matching pants or shoes.The recommendation for a single item does not consider the compatibility relationship between items,so it is easy to give multiple similar recommendations.In order to solve this problem,the bundle recommendation considering the compatibility relationship of multiple items has gradually become a new research hotspot.Item bundle recommendation has a wide range of application scenarios,such as recommending an outfit consisting of multiple pieces of clothing,a tourist route consisting of multiple attractions,a fitness plan consisting of multiple sports,etc.Compared with single-item recommendation,bundle recommendation provides users with more comprehensive results by recommending multiple compatible items at the same time.However,the complex item associations,diverse user preferences,and severe data sparsity brings greater challenges to bundle recommendation.The existing works have many limitations,such as low robustness,low computation efficiency and poor adaptability,which makes it difficult to meet the practical needs.Solving the bundle recommendation problem requires innovation at both the fundamental and application levels.At the principle level,bundle recommendation requires an additional explanation and exploration of the correlation between multiple items.At the application level,not only the direct recommendation of the budle needs to be considered,but also the item suggestion task related to the bundle,etc.Among the many item bundle recommendation applications,clothing matching recommendation is the most representative one,and it is also one of the most common tasks required by users in daily life.This work starts with the content analysis of item images,and focuses on the research of recommendation accuracy,computational efficiency and few-shot learning in bundle recommendation,which includes the following aspects:(1)A context-based item compatibility measurement algorithm is proposed to improve the recommendation accuracy.The existing recommendation approaches usually ignore the context of items when measuring the similarity,which restricts the performance.This work uses the theory of mutual information maximization to reasonably embed context conditions into the features of items,which improves the performance of compatibility measurement and the accuracy of bundle recommendation.(2)Abnormal items in the bundle are studied to improve the detection accuracy.Anomalous items are the opposite of similar items,which are items that do not match the whole item bundle.Detection of anomalous items can help users improve existing bundle at a lower cost.In this work,the self-attention mechanism is used to encode and decode the correlation between items,which realizes the detection of abnormal items and improves the accuracy of prediction.(3)A bundle representation algorithm is proposed to further improve the recommendation accuracy.Although existing bundle recommendation methods have proposed different bundle representation approaches based on different multi-item associativity assumptions,they do not consider the compact representation of bundles.To this end,this work constructs an intuitive and efficient compact representation for bundles that enables the complex dependency among items to be abstracted into a low-dimensional manifold,thereby simplifying the combinatorial recommendation problem.The combined representation method is applied to different types of recommendation tasks and achieves performance improvements under the respective tasks.(4)A hash-based bundle recommendation algorithm is proposed to improve the recommendation efficiency.Taking advantage of the efficiency of binary calculation,this work applies hash technology to the item bundle recommendation task,and obtains the binary codes of items and users respectively,which improves the computation efficiency.Furthermore,this work proposes a new Bernoulli distribution-based sampling hashing algorithm,which provides a probabilistic interpretation of learning to hash and a controllable performance-efficiency trade-off.(5)Parameter estimation for new uers is studied to solve the problem of few-shot learning in bundle recommendation.In the deployment of recommender systems,there are a large number of new users with limited data.In order to improve the recommendation performance for new users,this work proposes a multi-vector representation approach for users,which decouples each user’s diverse preferences into different vectors,so that the preferences of new users can be quickly reconstructed directly through existing users,which improves the recommendation performance for users with limited data.This work considers the clothing matching recommendation as the research object of bundle recommendation tasks,where a large-scale personalized clothing matching data set is collected,and a large number of experiments and theoretical analysis are carried out.The conducted study on clothing matching provides new algorithmic basis and insights for the subsequent research on bundle recommendation.