PaperlayerView exact revision on arXiv

Source-linked AI summary

Addressing the Item Cold-start Problem by Attribute-driven Active Learning

Yu Zhu, Jinhao Lin, Shibi He, Beidou Wang, Ziyu Guan, Haifeng Liu, Deng Cai

arXiv:1805.09023v1cs.IRcs.LGstat.ML

TL;DR

The paper addresses item cold-start recommendation when no ratings exist for a new item and prior approaches use either content-based hybridization or active selection without content. It combines item attributes, user rating history, and active learning to select informative raters and predict unselected users’ ratings, with reported superiority over traditional methods on two real-world datasets.

  • Problem

    Item cold-start recommendation lacks prior ratings for new items, while earlier methods separately use content-based hybrid recommendation or active user selection without item attributes.

  • Method

    The method selects users using item attributes, rating history, willingness to rate, rating diversity, personalization, and a dynamic active-learning budget, then predicts ratings for unselected users.

  • Results

    Experimental results on two real-world datasets show that the proposed method outperforms traditional methods.

  • Takeaways & Limitations

    Combining content information with active user selection provides the paper’s supported approach for improving item cold-start recommendations.

  • Takeaways & Limitations

    The task differs from multi-armed bandits because users are selected in a batch without observing other users’ feedback, so many bandit solutions cannot be applied directly.

Abstract

from arXiv · show

In recommender systems, cold-start issues are situations where no previous events, e.g. ratings, are known for certain users or items. In this paper, we focus on the item cold-start problem. Both content information (e.g. item attributes) and initial user ratings are valuable for seizing users' preferences on a new item. However, previous methods for the item cold-start problem either 1) incorporate content information into collaborative filtering to perform hybrid recommendation, or 2) actively select users to rate the new item without considering content information and then do collaborative filtering. In this paper, we propose a novel recommendation scheme for the item cold-start problem by leverage both active learning and items' attribute information. Specifically, we design useful user selection criteria based on items' attributes and users' rating history, and combine the criteria in an optimization framework for selecting users. By exploiting the feedback ratings, users' previous ratings and items' attributes, we then generate accurate rating predictions for the other unselected users. Experimental results on two real-world datasets show the superiority of our proposed method over traditional methods.

1 INTRODUCTION

The paper addresses item cold-start recommendation by combining item attributes with active user selection, while accounting for willingness to rate and rating diversity. Its framework uses personalized, dynamic selection to balance user experience with prediction accuracy.

  • 1 INTRODUCTION: Attribute-only recommendation can fail when similar films receive different audience responses, as Taken averaged 8.0 while Taken 3 averaged 6.3.The example motivates collecting initial feedback instead of relying solely on shared genres, writers, and actors.
  • 1 INTRODUCTION: The framework combines item attributes and active learning to recommend new items before any user has rated them.It selects users using attribute information and rating history, then predicts ratings for unselected users from the feedback.
  • 1 INTRODUCTION: The method separately models whether users will rate the item and what ratings they will give, reducing selection and prediction bias.Selecting only users likely to give high ratings could produce uniformly inflated predictions for other users.
  • 1 INTRODUCTION: Personalized selection uses four criteria in an integer quadratic programming framework so different item attributes can lead to different selected users.This is intended to avoid repeatedly requesting ratings from the same users and to improve fairness and user experience.
  • 1 INTRODUCTION: Dynamic budgets allocate limited rating requests toward new items whose user preferences are difficult to predict.The paper distinguishes unpopular items, broadly favored items, and popular but controversial items that need more feedback.
  • 1 INTRODUCTION: The selection strategy balances exploitation of existing knowledge with exploration of unselected users’ preferences.It favors willing users during active learning while seeking feedback that improves later predictions.

2 RELATED WORK

Prior work addresses item cold-start through hybrid content-collaborative models or active learning without content information. The paper also distinguishes its batch-selection setting from sequential multi-armed bandit methods.

  • 2 RELATED WORK: Hybrid recommendation methods combine item content with collaborative filtering, while other methods estimate new-item factors from a small number of ratings without attributes.These approaches represent the main content-based and sparse-feedback alternatives discussed in the related work.
  • 2 RELATED WORK: Most recommender-system active-learning studies target user cold-start by selecting items for newly signed users using strategies such as popularity and coverage.The item cold-start setting instead selects users to rate a new item.
  • 2 RELATED WORK: Prior active-learning approaches for item cold-start focus on pure collaborative filtering and do not incorporate content information.This difference motivates the paper’s attribute-driven active-learning framework.
  • 2.3 The Exploitation-exploration Trade-off: Although the task shares exploitation-exploration trade-offs with multi-armed bandits, bandit methods generally assume sequential selections with immediate rewards.Here, a batch of users is selected simultaneously without observing other users’ feedback, limiting direct applicability of many bandit algorithms.

3 PRELIMINARIES AND MODEL

The paper formulates item cold-start recommendation as selecting users for feedback and then predicting ratings for the remaining users. Factorization Machines model both rating and willingness-to-rate tasks using user, item, and attribute features.

  • The task predicts ratings for a new item from the user-item matrix, item-attribute matrix, and the new item’s attribute vector.
  • The solution has two phases: active learning selects users to rate the new item, followed by prediction for the other users.
  • Factorization Machines are used for classification and regression tasks throughout both phases.
  • Rating regression uses user, item, and item-attribute features, with ratings as labels; feature interactions capture user–item or user–attribute preferences.

4 OUR METHOD

The method selects feedback users using willingness, rating diversity, rating objectivity, and representativeness, then allocates a dynamic budget across new items. It combines these criteria in an iterative optimization procedure and retrains a regression model with feedback.

  • Users are selected using four criteria: willingness to rate, diverse potential ratings, objective ratings, and representativeness.The criteria use classification, regression, rating-history similarity, and objective-rating measures.
  • The willingness criterion transforms observed ratings into a binary matrix and predicts whether each user will rate the new item.Factorization Machines use user and attribute features for this classification task.
  • Diverse potential ratings favor users with different attribute-based rating estimates, reducing redundant feedback and bias toward one rating region.The diversity values can be computed independently and in parallel.
  • Objective ratings favor users whose historical ratings stay close to item means, while penalizing users with few rated items.A smaller objective value indicates a more objective user.
  • Representativeness uses cosine similarity between users’ rating-history vectors to avoid selecting redundant users and cover unselected users.
  • After feedback collection, a Factorization Machine is retrained from previous ratings and new feedback to predict ratings for unselected users.
  • An iterative optimization process selects the user set, with binary entries indicating selected users and monotonic convergence under the stated matrix condition.
  • The active-learning budget is dynamically allocated across new items according to popularity and controversy rather than fixed equally per item.The budget score combines popularity with a weighted controversy score.

5 EXPERIMENTS

The experiments evaluate the proposed methods and baselines on Movielens-IMDB and Amazon using cold-start items and active-user selection. Compared methods include hybrid recommenders, factorization machines, and active-learning strategies.

  • 5.1 Dataset: 20% of items are randomly designated as new testing items, while ratings and attributes from the remaining 80% train the models.Movielens-IMDB combines MovieLens ratings with IMDb attributes; Amazon provides ratings and product attributes.
  • 5.1 Dataset: Users are split into an active-selection set and a prediction set, with selected users queried for ratings on each testing item.Half of all users are randomly assigned to each set.
  • 5.2 Compared Algorithms: The comparison includes hybrid methods, factorization machines, and active-learning variants using popularity, coverage, exploration, or the proposed criteria.The proposed methods are FMFC without a dynamic budget and FMFC-DB with one.
  • 5.2 Compared Algorithms: The user-selection objective balances diversity among selected users with representativeness of unselected users.The integer quadratic program can be relaxed to a quadratic problem and solved with existing solvers.

5.3 Evaluations

Evaluation measures both the experience of users receiving rating requests and prediction quality for unselected users. It also evaluates top-N recommendation quality using predicted ratings for new items.

  • 5.3 Evaluations: PFR measures the share of requested users who provide feedback, while AST measures average selection frequency per user.Higher PFR indicates better willingness to rate; higher AST indicates repeated selection and poorer user experience.
  • 5.3 Evaluations: RMSE and MAE evaluate rating predictions for unselected users using actual and predicted ratings on new items.Methods with active learning retrain using feedback ratings before prediction; non-active methods predict directly from training items.
  • 5.3 Evaluations: Top-N recommendation quality is evaluated by selecting 10 testing items with the largest predicted ratings and treating ratings above 3 as preferences.The evaluation counts preferred new items appearing in users’ recommendation lists.

5.4 Parameter Setting

Parameter setting calibrates the four selection criteria before tuning their relative weights. The experiments use fixed active-learning budgets and evaluate prediction and recommendation performance with multiple tables.

  • 5.4 Parameter Setting: The criteria are standardized, and matrix terms are divided by the user count so larger matrices do not dominate optimization.The scaling makes the tuned parameters comparable in magnitude.
  • 5.4 Parameter Setting: The active-learning budget is set to k = 25 users per testing item, with α fixed at 1 and the remaining free parameters tuned by grid search.RMSE on cross-validation over training data is used for tuning.
  • 5.4 Parameter Setting: For Amazon, the final tuned parameters are α = 1, β = 0.3, γ = 0.03, and σ = 0.1.Other experimental settings remain the same as for Movielens-IMDB.

5.5 Results and Analysis

FMFC and FMFC-DB outperform the baselines in prediction and top-N recommendation, while the dynamic budget further improves all reported metrics. Ablation results support the contribution of the criterion encouraging likely feedback.

  • 5.5.1 Algorithm Comparison: FMFC and FMFC-DB achieve the best RMSE, MAE, and top-N recommendation performance among the compared methods.The comparison covers both testing datasets and cites Tables 2–5.
  • 5.5.1 Algorithm Comparison: The dynamic-budget variant FMFC-DB further improves all reported metrics compared with FMFC.The improvement includes prediction, recommendation, and active-learning user-experience measures.
  • 5.5.1 Algorithm Comparison: FMFC and FMFC-DB jointly achieve high feedback rates and low prediction errors, supporting a trade-off between exploitation and exploration.The reported analysis connects high PFR with the active-learning phase and low RMSE and MAE with prediction.
  • 5.5.1 Algorithm Comparison: The experiments repeat holdout training and testing 100 times with independently selected testing items for significance testing.Paired t-tests compare the proposed methods with each other and with all baselines.
  • 5.5.2 Criterion Analysis: Removing the criterion for users likely to rate the new item dramatically decreases PFR for both proposed methods.This result supports the criterion’s effectiveness for obtaining feedback.

Criterion (2): Selected users’ potential ratings are diverse

Criterion (2) seeks selected users whose potential ratings are diverse, helping avoid predictions biased toward a fixed rating region.

  • Criterion (2): Selected users’ potential ratings are diverse: Removing Criterion (2) decreases the average diversity of selected users’ actual ratings in both FMFC and FMFC-DB.The result supports the criterion’s role in selecting users with diverse potential ratings.
  • Criterion (2): Selected users’ potential ratings are diverse: Diverse selected ratings help prevent the final prediction model from becoming biased toward a fixed region of ratings.
  • Criterion (2): Selected users’ potential ratings are diverse: The criterion is evaluated by comparing the average diverse value of selected users’ actual ratings with and without it.

Criterion (3): Selected users’ generated ratings are objective

Criterion (3) aims to make selected users’ average ratings resemble the overall user average, producing a more objective selection.

  • Criterion (3): Selected users’ generated ratings are objective: Criterion (3) improves objectivity by making the selected users’ average rating approximate the average rating of all users.
  • Criterion (3): Selected users’ generated ratings are objective: Figure 5 evaluates how removing each criterion affects FMFC’s RMSE across different numbers of selected users.
  • Criterion (3): Selected users’ generated ratings are objective: Without Criterion (3), selected users’ average ratings are farther from the overall user average for both FMFC and FMFC-DB.The comparison is based on |r̄_no c3 − r̄_all| versus |r̄_ours − r̄_all|.

Criterion (4): Selected users are representative

The experiments examine representativeness, criterion contributions, budget allocation, and the trade-off between prediction accuracy and active-learning performance.

  • Criterion (4): Selected users are representative: Without Criterion (4), the average similarity between selected and unselected users declines, indicating that the criterion improves representativeness.
  • Criterion (4): Selected users are representative: RMSE increases when any criterion is removed, with the largest increase occurring after removing Criterion (1).RMSE decreases as k increases because more feedback ratings provide more information about the new item.
  • Criterion (4): Selected users are representative: Dynamic budgeting improves both RMSE and PFR across total-budget settings, but its benefit narrows when the total budget becomes sufficient.
  • Criterion (4): Selected users are representative: As α varies, FMFC’s RMSE first decreases then increases while PFR keeps increasing, with the best RMSE occurring around α = 1.

6 CONCLUSION

The paper combines attribute information and active learning to address item cold-start recommendation. It selects users for feedback, retrains predictions with those ratings, and allocates active-learning resources dynamically.

  • 6 CONCLUSION: The proposed scheme leverages both active learning and item attributes for item cold-start recommendation.
  • 6 CONCLUSION: For a new item, the method selects a small portion of users using four criteria, adds their feedback ratings through retraining, and predicts ratings for unselected users.
  • 6 CONCLUSION: A dynamic active-learning budget distributes limited resources and contributes to better recommendation performance.The paper also identifies extending the regression-based method to ranking models for top-N recommendation as future work.
Loading 1805.09023v1…