PaperlayerView exact revision on arXiv

Source-linked AI summary

Try This Instead: Personalized and Interpretable Substitute Recommendation

Tong Chen, Hongzhi Yin, Guanhua Ye, Zi Huang, Yang Wang, Meng Wang

arXiv:2005.09344v1cs.IR

TL;DR

Existing substitute recommenders focus on pairwise item relationships, often overlooking personal preferences and producing hard-to-interpret results. A2CF uses review-derived attributes to model user preferences and item substitutions and generate attribute-level explanations. Experiments report superior recommendation and interpretation performance, while observed purchases do not reveal the query items users browsed.

  • Problem

    Existing substitute recommendation methods focus on item-item relationships rather than users’ preferences, and latent representations make their recommendations difficult to interpret.

    Source: §1 · p. 2

  • Method

    A2CF uses explicit attributes from reviews to model user-attribute and item-attribute interactions, combine personalization with substitution, and compare attributes for explanations.

    Source: §1 · p. 2, §3.4 · p. 4

  • Results

    Experiments on three million-scale datasets show A2CF has superior recommendation performance and interpretation quality.

    Source: §1 · p. 2, §8 · p. 9

  • Takeaways & Limitations

    Review-derived attributes bridge user preferences and item properties for personalized substitute recommendations with explanations.

    Source: §8 · p. 9

    Because these datasets record purchases rather than browsed alternatives, query items are unobserved and must be simulated with popularity-biased sampling.

    Source: §5.2 · p. 6

Abstract

from arXiv · show

As a fundamental yet significant process in personalized recommendation, candidate generation and suggestion effectively help users spot the most suitable items for them. Consequently, identifying substitutable items that are interchangeable opens up new opportunities to refine the quality of generated candidates. When a user is browsing a specific type of product (e.g., a laptop) to buy, the accurate recommendation of substitutes (e.g., better equipped laptops) can offer the user more suitable options to choose from, thus substantially increasing the chance of a successful purchase. However, existing methods merely treat this problem as mining pairwise item relationships without the consideration of users' personal preferences. Moreover, the substitutable relationships are implicitly identified through the learned latent representations of items, leading to uninterpretable recommendation results. In this paper, we propose attribute-aware collaborative filtering (A2CF) to perform substitute recommendation by addressing issues from both personalization and interpretability perspectives. Instead of directly modelling user-item interactions, we extract explicit and polarized item attributes from user reviews with sentiment analysis, whereafter the representations of attributes, users, and items are simultaneously learned. Then, by treating attributes as the bridge between users and items, we can thoroughly model the user-item preferences (i.e., personalization) and item-item relationships (i.e., substitution) for recommendation. In addition, A2CF is capable of generating intuitive interpretations by analyzing which attributes a user currently cares the most and comparing the recommended substitutes with her/his currently browsed items at an attribute level. The recommendation effectiveness and interpretation quality of A2CF are demonstrated via extensive experiments on three real datasets.

1 INTRODUCTION

The introduction frames substitute recommendation as an underexplored task whose existing approaches often neglect users’ attribute-level preferences and provide little interpretability. It introduces A2CF, which uses attributes to model personalization and substitution and to generate explanations.

1 INTRODUCTION

  • The paper identifies personalized and interpretable substitute recommendation as an underexplored problem and evaluates A2CF against state-of-the-art methods on three million-scale datasets.The authors report superior recommendation performance and high-quality interpretations.Source: §1 · pp. 1–2
  • Existing substitute recommenders often rank substitutes from query-item similarity alone, so users with different preferences can receive the same results.A user prioritizing camera quality and one preferring Apple may need different substitutes for the same iPhone query.Source: §1 · p. 2
  • Latent-factor and neural approaches make recommendations difficult to interpret, limiting users’ ability to understand why substitutes suit their needs..Source: §1 · p. 2
  • A2CF uses explicit user-attribute and item-attribute associations to model personalization and substitution, then explains recommendations through attribute comparisons.It infers a user’s current attribute demands and highlights attributes where recommended substitutes perform better.Source: §1 · p. 2

2 PRELIMINARIES

The preliminaries establish notation for users, items, attributes, and their interaction matrices, then describe how review-derived attribute-sentiment pairs produce user- and item-attribute matrices.

2.1 Notations

  • The notation uses u, v, and a for user, item, and attribute embeddings, and X and Y for user-attribute and item-attribute matrices.Vectors and matrices are bolded, sets use calligraphic uppercase letters, and vectors are columns unless specified.Source: §2.1 · p. 2

2.2 Extracting Attributes and Sentiments

  • The authors apply Sentires to extract attribute-sentiment pairs from user reviews, producing a sentiment lexicon for the entire review dataset whose entries are (user, item, attribute, sentiment).Source: §2.2 · p. 2
  • The user-attribute matrix X records how much each user cares about each attribute, while item-attribute matrix Y aggregates users’ feedback for each item attribute.Both matrices include entries for attributes without mentions or reviews, and Y also uses the mean sentiment of mentions.Source: §2.2 · p. 3

3 ATTRIBUTE-AWARE COLLABORATIVE FILTERING

A2CF uses explicit user-attribute and item-attribute associations to learn preferences and item characteristics, then combines personalization and substitution constraints in a ranking model. It estimates missing attribute values and uses attribute-aware representations to compare candidate substitutes with a query item.

3.1 User-Attribute Collaborative Filtering

  • User-attribute collaborative filtering learns user and attribute representations to estimate observed user-attribute preferences from the user-attribute matrix.A residual feed-forward network models nonlinear user-attribute interactions, and its predicted scores are constrained to the matrix’s [1, N] range.Source: §3.1 · p. 3

3.2 Item-Attribute Collaborative Filtering

  • Item-attribute collaborative filtering estimates item quality on attributes, sharing attribute embeddings with user-attribute learning to jointly optimize both objectives.The two collaborative-filtering components use separate network weights and biases.Source: §3.2 · p. 3

3.3 Estimating Missing Entries in X and Y

  • A2CF fills unobserved entries in the user-attribute and item-attribute matrices, producing estimated user preferences and item attribute-performance scores.The estimated matrices provide the attribute values used in subsequent recommendation modeling.Source: §3.3 · p. 4

3.4 Uniting Substitution and Personalization Constraints for Recommendation

  • A2CF combines substitution and personalization scores in a BPR-style ranking objective, training candidates to rank above negatives under both constraints.The score balances item-item substitution and user-item preference with a trade-off coefficient.Source: §3.4 · p. 4
  • The substitution score combines latent item similarity with an attribute-aware item representation, weighting attributes according to how similarly the query and candidate score on them.A scaling factor controls the strength of dominant attributes in the softmax weighting.Source: §3.4 · p. 4

3.5 Optimization Strategy

  • Training alternates between optimizing user-attribute and item-attribute losses and optimizing the recommendation loss; the authors use Adam and precompute estimated matrices before recommendation training.They tune hyperparameters by grid search and use five negative samples per positive training label.Source: §3.5 · p. 5

4 UTILIZING A2CF FOR RECOMMENDATION AND INTERPRETATION

A2CF ranks substitute candidates by both their fit to the browsed item and their interest to the user. It explains recommendations by identifying attributes the user values and comparing the substitute’s attribute performance with the query item.

4.1 Personalized Substitute Recommendation

  • A2CF ranks each candidate using both substitution fit to the query item and personalization for the user, recommending the top-K items.The combined score weights the substitution and personalization functions with γ and 1 − γ.Source: §4.1 · p. 5

4.2 Personalized Interpretation

  • A2CF interpretations identify attributes the user values when seeking an alternative and compare the recommended item’s performance with the query item on those attributes.The method selects a small number of top attributes, usually fewer than five, to keep template-based explanations concise.Source: §4.2 · p. 5

5 EXPERIMENTAL SETTINGS

The experiments use three Amazon review datasets, evaluate personalized substitute ranking and interpretation quality, and compare A2CF with substitute-recommendation baselines under a unified parameter setting.

5.1 Datasets

  • Cellphone, Automotive, and Office are the three Amazon product-category datasets, with substitution links defined by co-views or a view followed by purchase.Users with fewer than five interacted items and items visited by fewer than five users are filtered; attributes mentioned fewer than twice are discarded, and data are split 80/10/10 for training, validation, and test.Source: §5.1 · pp. 5–6

5.2 Evaluation Protocols

  • Recommendation effectiveness is measured with HR@K and NDCG@K after ranking each positive item against 1,000 random negatives at K = 5, 10, 20, and 50.The evaluation protocol follows established recommendation metrics and uses one positive test instance mixed with the sampled negatives.Source: §5.2 · p. 6
  • Query items are popularity-biased samples from substitutes the user has not interacted with, with sampling probability proportional to popularity raised to 0.75.One query is sampled per observed user–item interaction; the sampling rationale is that shoppers tend to start with popular products and broaden candidates as their demands clarify.Source: §5.2 · p. 6
  • Test ground-truth items are unseen in training for both the user and the query item, creating the paper's hardest test instances.The test setup excludes each ground-truth item from training pairs with either its user or its query item.Source: §5.2 · p. 6
  • Attribute trade-off coverage (ATC) combines MAP for review-mentioned attributes with NDCG for the ground-truth item, rewarding cases where both interpretation and recommendation are accurate.Interpretations are evaluated only for ground-truth items, since users have not visited or reviewed non-ground-truth items.Source: §5.2 · p. 6

5.3 Baseline Methods

  • The comparison includes Sceptre, PMSC, SPEM, and CLVA, whose approaches range from link prediction and product embeddings to variational autoencoders and personalized factorization.CLVA incorporates probabilistic matrix factorization for personalized substitute recommendation.Source: §5.3 · p. 6

5.4 Hyperparameter Settings

  • A2CF's overall recommendation performance is reported with d = 64, l = 1, γ = 0.7, and β = ϵ = d^0.5 = 8; baseline hyperparameters are selected by grid search.The effects of alternative A2CF hyperparameter settings are discussed in Section 6.4.Source: §5.4 · p. 6

6 EXPERIMENTAL RESULTS AND ANALYSIS

Across three datasets, A2CF outperforms recommendation baselines and leads Sceptre and CLVA on interpretation-quality scores. Ablations and sensitivity tests show that both constraints and attribute aggregation matter, while hyperparameter effects vary by dataset.

6.1 Recommendation Effectiveness (RQ1)

  • A2CF outperforms all baselines on three datasets (p-value < 0.01), improving over the second-best results by 28.1% on HR@5 and 29.2% on NDCG@5.These are average improvements across the datasets.Source: §6.1 · p. 7
  • Attribute-using recommenders outperform graph-only PMSC and SPEM; A2CF performs better on Office than similarly sparse Cellphone, but worse on more-sparse Automotive.The passage relates the dataset differences to attribute availability and interaction sparsity.Source: §6.1 · p. 7

6.2 Interpretation Quality (RQ2)

  • A2CF leads Sceptre and CLVA on ATC scores across all three datasets.The attribute ranking lists are truncated to 500 for evaluation.Source: Table 3 · p. 7, §6.2 · p. 7
  • Case studies show explanations align with review-mentioned attributes, including positive screensize and shield matches and printer colors and quality despite negative sentiment for colors.The printer review's negative colors sentiment refers to a previously purchased printer.Source: Figure 2 · p. 8, §6.2 · p. 8

6.3 Importance of Key Components (RQ3)

  • Removing either constraint lowers performance: substitution removal is most damaging on Automotive, while personalization removal hurts denser Cellphone and Office more.The study reports a severe decrease across datasets without the substitution constraint and a significant drop without personalization.Source: §6.3 · p. 8
  • Removing item attribute aggregation decreases performance, and removing user attribute aggregation also yields inferior results on all three datasets.The item aggregation supplies attribute information for substitution prediction; user aggregation incorporates preferred attributes into preference modeling.Source: §6.3 · p. 8

6.4 Impact of Hyperparameters (RQ4)

  • Greater depth helps across datasets until gains stop at l = 3–4 due to overfitting; the joint effect of the constraints is maximized at γ = 0.7.The reported γ balances substitution and personalization constraints.Source: §6.4 · pp. 8–9
  • β and ϵ effects vary by dataset; for Automotive, larger values spread probability across more attributes, adapting modeling to its higher sparsity.The passage attributes the dataset difference possibly to variation in the quantity of extracted attributes.Source: §6.4 · p. 9

7 RELATED WORK

Prior product-relationship work includes complement recommendation and several approaches to mining product relationships, but substitute recommendation remains challenged by limited personalization and explainability. Attribute-aware recommendation research provides a basis for explanations, yet existing methods do not account for users’ shifting attention across attributes.

7.1 Product Relationship Mining

  • Product relationship mining includes Sceptre’s topic-model approach, review-based RRN embeddings, and graph-based PMSC and SPEM methods.Sceptre compares product topic distributions using LDA; RRN combines textual reviews with manually crafted features.Source: §7.1 · p. 9
  • Existing substitute methods treat recommendation as item-based search, while CLVA adds personalization but retains latent representations that hinder explainability.The resulting lack of explainability can make it difficult to convince or attract potential customers.Source: §7.1 · p. 9

7.2 Attribute Modelling for Recommendation

  • Review-based attribute models use extracted attributes and sentiments, but existing methods offer generic explanations and miss users’ changing attribute attention while browsing.A2CF is presented as inferring desired attributes and using advantageous attributes of recommended items for personalized interpretations.Source: §7.2 · p. 9

8 CONCLUSION

The paper frames personalized and interpretable substitute recommendation as a new research problem and proposes A2CF. Experiments are reported to show superior recommendation and interpretation performance.

8 CONCLUSION

  • A2CF uses review-extracted attributes to bridge user preferences and item properties for personalized substitute recommendations and high explainability.The authors report superior performance on both recommendation and interpretation.Source: §8 · p. 9

Related papers

Loading 2005.09344v1…