Learning Users’ Visual Preferences for Improving Recommendations

Wang, Shoujin; Yang, Chenlu; Qu, Tong; Yang, Kai; Wan, Wanggen

doi:10.1109/access.2022.3140215

Cited by 3 publications

(7 citation statements)

References 43 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The open source datasets are obtained through slight modifications of the existing Yelp (restaurants domain) and Amazon (movies&TV and clothing) datasets. Whereas the original datasets contain customer reviews with binary sentiment annotations, we instead transform the datasets to match customer reviews to item product descriptions [9,22], aligning the datasets to the modality of WhisperD. With this transformation, given a review and a set of possible targets, the task is to pair the review with the correct target description.…”

Section: Open-source Datasetsmentioning

confidence: 99%

Contrastive Learning for Interactive Recommendation in Fashion

Sevegnani¹,

Seshadri²,

Beniwal³

et al. 2022

Preprint

View full text Add to dashboard Cite

Recommender systems and search are both indispensable in facilitating personalization and ease of browsing in online fashion platforms. However, the two tools often operate independently, failing to combine the strengths of recommender systems to accurately capture user tastes with search systems' ability to process user queries. We propose a novel remedy to this problem by automatically recommending personalized fashion items based on a user-provided text request. Our proposed model, WhisperLite, uses contrastive learning to capture user intent from natural language text and improves the recommendation quality of fashion products. WhisperLite combines the strength of CLIP embeddings with additional neural network layers for personalization, and is trained using a composite loss function based on binary cross entropy and contrastive loss. The model demonstrates a significant improvement in offline recommendation retrieval metrics when tested on a real-world dataset collected from an online retail fashion store, as well as widely used open-source datasets in different e-commerce domains, such as restaurants, movies and TV shows, clothing and shoe reviews. We additionally conduct a user study that captures user judgements on the relevance of the model's recommended items, confirming the relevancy of WhisperLite's recommendations in an online setting.

show abstract

Section: Open-source Datasetsmentioning

confidence: 99%

Contrastive Learning for Interactive Recommendation in Fashion

Sevegnani¹,

Seshadri²,

Beniwal³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Nevertheless, they disregard the information in the sequential sets [18]. Similarly, sequential recommendation techniques [4] are used to predict the next basket by discovering the dynamic preferences of the users via sequential user-item interactions [4]. However, sequential models are incompatible with basket-based recommendations since the NBR approach deals with several concurrent items in a single set [1].…”

Section: Related Workmentioning

confidence: 99%

“…Another well-known solution is the Markov chain technique, which assumes that the previous few baskets largely determine the next basket. This technique, however, often reduces the performance of the corresponding recommendation while ignoring the highorder dependency from a long time ago [4]. However, results of most standard recommendation techniques are difficult to interpret for humans.…”

Section: Related Workmentioning

confidence: 99%

“…A basket is defined as all products purchased by the same consumer in a single transaction. 4 1 https://recsyswiki.com/wiki/Grocery_shopping_datasets 2 https://archive.ics.uci.edu/ml/datasets/Online+Retail…”

Section: B Datasetsmentioning

confidence: 99%

“…This recommendation task in e-commerce sites is formulated as the next-basket recommendation (NBR), which has received much attention recently [1]. Collaborative filtering (CF) [2], Session-based [3], Markov-chain [4], and Sequential-based [5] recommendation techniques are widely used to recommend a basket of items for the users to purchase next time they visit the shop based on their preferences. The four most common methods used by recommendation techniques to elicit user preferences are attribute weights, rating scales, implicit behavior, and critiques [6].…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

An Efficient Approach for Rational Next-Basket Recommendation

Fouad¹,

Hussein

Rady

et al. 2022

IEEE Access

View full text Add to dashboard Cite

E-commerce is one of the most valuable and popular application scenarios of recommendation technology. In this context, making rational decisions is required due to the necessity of objectivity and decision rationale tracking. However, a significant number of irrational recommendations may be generated as a result of using a narrow range of interestingness criteria, considering inconsistent user preferences, overlooking the temporal weights of items, and using outlier choices in making a prediction. In this paper, we propose an efficient approach called MONBR (Multi-Objective Next-Basket Recommendation) to improve the quality of recommendations. An approach promotes rational recommendations by measuring the temporal importance of items to the user using several interestingness criteria, namely utility, popularity, stability, frequency, occupancy, and novelty. Moreover, combining the multi-criteria weights of items obtained from the MONBR approach with the item-based collaborative filtering technique results in more accurate and reasonable recommendations. Experiment results on real-world datasets demonstrate that the proposed approach improves the quality of recommendations. The average quality of recommendations derived from the proposed approach is up to 60.95% more accurate than state-of-the-art algorithms with a recommended basket size ranging from 10 to 100. Furthermore, the results highlight the importance of adapting objectivity and rationality in the recommendation process to meet the needs of as many users as possible. The average user-level performance is up to 10.21% higher on dense datasets.

show abstract

A Hybrid Recommender System Combining Collaborative Filtering with Utility Mining

fouad¹,

Hussein

Rady

et al. 2022

IJICIS

View full text Add to dashboard Cite

Based on a variety of information sources, recommender systems can identify specific items for various user interests. Techniques for recommender systems are classified into two types: personalized and non-personalized. Personalized algorithms are based on individual user preferences or collaborative filtering data; as the system learns more about the user, the recommendations will become more satisfying. They do, however, suffer from data sparsity and cold start issues. On the other hand, non-personalized algorithms make recommendations based on the importance of the items in the database; they are very useful when the system has no information about a specific user. Their accuracy, however, is limited by the issue of personalization. In most cases, one of the recommendation categories can be used to make recommendations. Yet, it is a challenge to evaluate the importance of items to the user while simultaneously using personalized and non-personalized preferences functions and ranking a set of candidate items based on these functions. This paper addresses this issue and improves recommendation quality by introducing a new hybrid recommendation technique. The proposed hybrid recommendation technique combines the importance of items to the user obtained by the utility mining method with the similarity weights of items produced by the collaborative filtering technique to make the recommendation process more reasonable and accurate. This technique can provide appropriate recommendations whether or not users have previous purchasing histories. Finally, experimental results show that the proposed hybrid recommendation technique outperforms both implemented collaborative filtering and utility-based recommendation techniques.

show abstract

Learning Users’ Visual Preferences for Improving Recommendations

Cited by 3 publications

References 43 publications

Contrastive Learning for Interactive Recommendation in Fashion

Contrastive Learning for Interactive Recommendation in Fashion

An Efficient Approach for Rational Next-Basket Recommendation

A Hybrid Recommender System Combining Collaborative Filtering with Utility Mining

Contact Info

Product

Resources

About