Source-linked AI summary
MeLU: Meta-Learned User Preference Estimator for Cold-Start Recommendation
Hoyeop Lee, Jinbae Im, Seongwon Jang, Hyunsouk Cho, Sehee Chung
TL;DR
Sparse interactions and inadequate evidence candidates limit cold-start recommendation for new users. MeLU uses MAML-based personalization and selects distinguishing evidence candidates from users’ consumption histories. Across two benchmark datasets and a user study, the method outperformed comparison methods and produced more satisfying recommendations, although its update cycle requires verification before real-world use.
Problem
Sparse user-item interactions and inadequate evidence candidates limit preference estimation and recommendation quality for new users.
Method
MeLU applies MAML to estimate each new user’s preferences from a few consumed items and selects evidence candidates using gradient-based and popularity-based scores.
Results
MeLU outperformed two comparative methods in cold-start scenarios across the MovieLens and Bookcrossing datasets, while proposed-strategy evidence produced higher user satisfaction in an 84-user A/B test.
Takeaways & Limitations
Personalized updates based on individual consumption histories can support cold-start recommendations, and evidence selection can improve satisfaction despite lower candidate selection rates.
Takeaways & Limitations
The model update cycle must be verified before real-world applications because performance may not be guaranteed for an entirely new type of user.
Abstract
from arXiv · showhide
This paper proposes a recommender system to alleviate the cold-start problem that can estimate user preferences based on only a small number of items. To identify a user's preference in the cold state, existing recommender systems, such as Netflix, initially provide items to a user; we call those items evidence candidates. Recommendations are then made based on the items selected by the user. Previous recommendation studies have two limitations: (1) the users who consumed a few items have poor recommendations and (2) inadequate evidence candidates are used to identify user preferences. We propose a meta-learning-based recommender system called MeLU to overcome these two limitations. From meta-learning, which can rapidly adopt new task with a few examples, MeLU can estimate new user's preferences with a few consumed items. In addition, we provide an evidence candidate selection strategy that determines distinguishing items for customized preference estimation. We validate MeLU with two benchmark datasets, and the proposed model reduces at least 5.92% mean absolute error than two comparative models on the datasets. We also conduct a user study experiment to verify the evidence selection strategy.
1 INTRODUCTION
Existing recommender systems struggle with sparse interactions and new users, while privacy limits profile-based personalization. MeLU addresses these issues with meta-learning and distinguishing evidence candidates for preference estimation from few consumed items.
- Collaborative and hybrid systems struggle with user cold-start because new users lack sufficient user-item interaction data.
- Content-based systems use profiles and item contents but may recommend similar lists to users with similar profiles, regardless of their different ratings.
- New users who rate only a few items can receive poor initial recommendations, and privacy concerns make collecting personal information difficult.
- MeLU applies meta-learning to estimate a new user’s preferences from a small number of consumed items.
- The proposed strategy selects distinguishing evidence candidates for customized preference estimation and validates the system with two benchmark datasets and a user survey.
2 META-LEARNING
Optimization-based meta-learning adapts model parameters to sampled tasks through local updates and then improves shared parameters using query-set losses. MeLU treats each user’s preference estimation as a task and personalizes recommendations from that user’s consumption history.
- Meta-learning trains models to rapidly adapt to new tasks using only a few examples.
- Optimization-based meta-learning performs local task updates followed by a global update that fits the model across sampled tasks.Local updates use support-set training losses; the global update uses query-set test losses after adaptation.
- MeLU treats each recommendation task as estimating one user’s preferences from a few user-item interactions.
- MeLU provides personalized recommendations using each user’s unique item-consumption history rather than finding users with similar ratings.
- Items with large gradients can serve as reliable evidence candidates because ratings on them provide substantial information about a new user’s preferences.
3 METHOD
MeLU estimates user preferences from user and item contents, then personalizes the decision-making and output layers using each user’s consumption history. Its evidence-candidate strategy combines preference distinction with item awareness to select informative items.
- 3.1 User Preference Estimator: The preference estimator embeds categorical user and item contents, concatenates the resulting vectors, and processes them with a multilayer neural network.Continuous contents bypass embedding and connect directly to the concatenation layer.
- 3.1 User Preference Estimator: The decision-making layers handle potentially different user and item embedding dimensions, while the output layer estimates ratings, implicit feedback, or dwell times.The output activation depends on the preference type: linear for ratings and dwell times, or sigmoid for implicit feedback.
- 3.2 Meta-Learned User Preference Estimator: MeLU locally updates decision-making and output-layer parameters from each user’s consumed items, then globally updates all parameters using query sets.The local update personalizes the model to a user’s consumption pattern; the global update seeks parameters that perform well after personalization.
- 3.2 Meta-Learned User Preference Estimator: Each user’s consumed items form the support set, whose loss drives repeated local personalization before query-set evaluation and global training.The support set contains consumed items and actual preferences, while the query set represents new item-consumption history.
- 3.3 Evidence Candidate Selection: Evidence candidates are selected by multiplying normalized gradient-based distinction and popularity-based awareness scores, then retaining the top k items.Larger average Frobenius norms indicate stronger distinction among users’ preferences, while interaction frequency represents item awareness.
4 EXPERIMENTS
MeLU is evaluated across four user–item cold-start settings on MovieLens 1M and Bookcrossing, using MAE and nDCG. Results show strong cold-start performance, rapid adaptation with one local update, robustness to short histories, and personalized estimates beyond user profiles.
- Experimental Design: Experiments use MovieLens 1M and Bookcrossing, partitioning users and items into existing and new groups.The four partitions cover existing items for existing users, existing items for new users, new items for existing users, and new items for new users.
- Experimental Design: Performance is measured with mean absolute error (MAE) and normalized discounted cumulative gain (nDCG).The study reports nDCG at different cutoffs, including nDCG1 and nDCG3.
- Experimental Results: MeLU outperforms PPR and Wide & Deep in the three cold-start scenarios on both datasets.PPR performs best for existing items and existing users on MovieLens but performs poorly across the cold-start scenarios.
- Experimental Results: A single local update is sufficient for MeLU to adapt quickly to users, with only slight MAE differences after additional updates.The MAE drops dramatically after the first iteration across the datasets, and the authors connect this rapid adaptation to online recommendation.
- Experimental Results: MeLU maintains robust recommendation performance when the item-consumption history is short.The history experiment uses one local update; performance may be unstable for very long histories when fewer than five people are represented.
- Experimental Results: Local updates estimate different preferences for users with identical profiles by incorporating each user’s item-consumption history.Contour maps show differences both between groups with different profiles and within groups sharing the same profiles.
5 USER STUDY ON EVIDENCE CANDIDATES
The user study compared proposed and popularity-based evidence candidates using MovieLens surveys, finding that the proposed strategy produced better recommendation outcomes despite lower candidate selection. Qualitative responses similarly favored the proposed strategy for matching users’ preferences.
- 84 users participated, with 45 rating popularity-based candidates and the remainder rating candidates selected by the proposed strategy.
- Popularity-based candidates were chosen by approximately one-half of users, whereas proposed candidates were selected by only one-quarter.
- The proposed strategy achieved better performance on all indicators, including more selected recommendations, higher average ratings, and higher nDCG1.
- Although the proposed strategy used fewer evidence items for local updating, its evidence ratings were similar to those of the popularity-based strategy.
- Participants described popularity-based recommendations as familiar but poorly matched, while a proposed-strategy participant reported more preferred movies among the recommendations.
6 CONCLUSION
The conclusion presents MeLU as a MAML-based recommender that estimates personalized preferences from few consumed items and combines this with evidence-candidate selection. The model outperformed two methods on two benchmark datasets, while the user study found greater satisfaction with recommendations from the proposed evidence strategy.
- MeLU uses MAML to identify personalized preferences from only a small number of consumed items.
- The proposed method outperformed two methods on two benchmark datasets.
- Local updating with item-consumption history effectively identified user preferences and supported the proposed evidence-candidate selection strategy.
- In an 84-user A/B test, proposed candidates were selected less often than popularity-based candidates, but produced higher user satisfaction with recommendation results.
- Future work should verify update cycles for entirely new user types and explore alternative evidence-selection and meta-learning recommender variants.
A LISTS OF EVIDENCE CANDIDATES.
Table A.1 compares popularity-based and proposed strategy-based evidence candidates, ranking all movies by each method’s score. Only three movies overlap between the two lists.
- All movies in Table A.1 are sorted by the score assigned by each evidence-candidate selection method.
- The popularity-based and proposed strategy-based lists contain only three overlapping movies.