PaperlayerView exact revision on arXiv

Source-linked AI summary

How to Retrain Recommender System? A Sequential Meta-Learning Method

Yang Zhang, Fuli Feng, Chenxu Wang, Xiangnan He, Meng Wang, Yan Li, Yongdong Zhang

arXiv:2005.13258v1cs.IR

TL;DR

Recommender systems need periodic retraining, but full retraining is costly and this practical problem has received relatively little research attention. SML transfers prior model knowledge using new interactions and next-period performance; experiments examine its effectiveness, efficiency, and design rationale across two datasets.

  • Problem

    Periodic recommender retraining is practically important, yet has received relatively little scrutiny, while real-time model updates can be technically challenging.

    Source: §1 · pp. 1–2

  • Method

    SML transfers the previous model's knowledge using current-period interactions and trains across sequential tasks with next-period interactions as the test objective.

    Source: §1 · p. 2

  • Results

    Experiments on two datasets report results and analyses of SML's effectiveness, efficiency, and design rationale.

    Source: §1 · p. 2, §6 · p. 10

  • Takeaways & Limitations

    The paper formulates recommender retraining as learning to generalize from newly collected data to the next time period.

    Source: §6 · p. 10

    The current work does not consider online learning strategies within serving periods.

    Source: §6 · p. 10

Abstract

from arXiv · show

Practical recommender systems need be periodically retrained to refresh the model with new interaction data. To pursue high model fidelity, it is usually desirable to retrain the model on both historical and new data, since it can account for both long-term and short-term user preference. However, a full model retraining could be very time-consuming and memory-costly, especially when the scale of historical data is large. In this work, we study the model retraining mechanism for recommender systems, a topic of high practical values but has been relatively little explored in the research community. Our first belief is that retraining the model on historical data is unnecessary, since the model has been trained on it before. Nevertheless, normal training on new data only may easily cause overfitting and forgetting issues, since the new data is of a smaller scale and contains fewer information on long-term user preference. To address this dilemma, we propose a new training method, aiming to abandon the historical data during retraining through learning to transfer the past training experience. Specifically, we design a neural network-based transfer component, which transforms the old model to a new model that is tailored for future recommendations. To learn the transfer component well, we optimize the "future performance" -- i.e., the recommendation accuracy evaluated in the next time period. Our Sequential Meta-Learning(SML) method offers a general training paradigm that is applicable to any differentiable model. We demonstrate SML on matrix factorization and conduct experiments on two real-world datasets. Empirical results show that SML not only achieves significant speed-up, but also outperforms the full model retraining in recommendation accuracy, validating the effectiveness of our proposals. We release our codes at: https://github.com/zyang1580/SML.

1 INTRODUCTION

Periodic retraining refreshes recommendations as user interactions accumulate, but the common strategies trade efficiency against retaining long-term preference. The paper proposes learning to transfer prior model knowledge while optimizing for recommendations in the next period.

1 INTRODUCTION

  • Periodic retraining refreshes models with new interaction data because real-time updates are challenging, especially for expressive, computationally expensive models.Industry commonly retrains daily or weekly.Source: §1 · p. 1
  • Fine-tuning is efficient but can forget long-term preference, sampling may lose information, and full retraining uses the most resources while providing the highest fidelity.Sample-based retraining also requires careful selection of representative historical interactions.Source: §1 · pp. 1–2
  • Existing strategies lack an explicit objective to optimize recommendation performance on the next time period, whose interactions provide evidence of current-model generalization.The paper contrasts this objective with manually crafted heuristics for selecting training examples.Source: §1 · p. 2
  • SML transfers knowledge from the previous model to new-data training and learns that transfer using next-period interactions as the future-performance objective.It frames each period as a meta-learning task: current-period interactions train, and the next period tests.Source: §1 · p. 2
  • SML is evaluated on Adressa and Yelp, with experiments reporting its effectiveness and examining its efficiency and design rationale.The introduction describes extensive results but does not report specific benchmark values here.Source: §1 · p. 2

2 PROBLEM FORMULATION

The paper formulates each retraining as a task judged by performance on the following period. It seeks a next-period-effective model using the current period's data and the previous model, while avoiding full retraining's costs and lack of explicit future optimization.

2 PROBLEM FORMULATION

  • Retraining task τt is judged on Dt+1, the data collected in the next period, because the new model serves the near future.Periods may be defined by elapsed time or interaction volume.Source: §2 · p. 2
  • Full retraining uses all historical and new data but consumes substantial resources, takes longer over time, and does not explicitly optimize Dt+1 performance.Directly training on Dt+1 would leak information and worsen generalization.Source: §2 · p. 2
  • The proposed formulation uses newly collected data Dt and previous parameters Wt−1 to obtain a model that performs well on Dt+1.Figure 2 depicts the previous model and a model learned on Dt being combined by a transfer component.Source: §2 · p. 2, Figure 2 · p. 3
  • Tasks form a sequence, so offline training should mirror the serving order; this makes retraining a meta-learning problem focused on generalization to future tasks.Ordinary SGD from Wt−1 can forget long-term interest and lacks an objective for serving Dt+1.Source: §2 · p. 3

3 METHOD

The method transfers knowledge from the previous recommender model into a model for the next period, training the transfer component toward future recommendation performance. Its framework is designed to apply to differentiable recommenders with matching architectures.

3.1 Model Overview

  • The generic framework combines a previous model and a current-data model through a transfer function to produce the model used for next-period recommendations.The previous model is held constant, and the framework applies to differentiable models of the same architecture and parameter semantics.Source: §3.1 · p. 3

3.2 Transfer Design

  • The CNN transfer stack models relations between parameters at the same dimension and across different dimensions, unlike the described MLP alternative.Its stack combines previous and current parameter vectors with their element-wise product; the paper notes that this product can indicate dimensions that increased or decreased.Source: §3.2 · pp. 3–4

3.3 Sequential Training

  • SML trains a transfer input from current-period data and optimizes shared transfer parameters on next-period data, using first-order MAML to avoid expensive higher-order gradients.The transferred model is optimized on current data while the transfer parameters target future recommendations; the transfer input is optimized through the transfer output.Source: §3.3 · pp. 4–5

3.4 Instantiation on Matrix Factorization

  • For matrix factorization, SML uses separate transfer networks for user and item embeddings and shares each network across embeddings with common dimension semantics.The paper says this sharing reduces the number of transfer parameters and makes transfer more generalizable.Source: §3.4 · p. 5

4 EXPERIMENTS

The experiments are organized around comparing SML with existing retraining strategies and recommender models, then examining its components, CNN architecture, and sources of improvement.

4 EXPERIMENTS

  • The experiments ask how SML performs against existing retraining strategies and recommender models.Source: §4 · p. 6
  • They examine how SML's components affect its effectiveness.Source: §4 · p. 6
  • They investigate how CNN architecture choices affect the transfer network.Source: §4 · p. 6
  • They ask where SML's improvements come from.Source: §4 · p. 6

4.1 Experimental Settings

The evaluation uses Adressa and Yelp to represent different user-interest timescales, organizes their interactions into chronological periods, and compares SML with retraining and sequential recommendation baselines.

4.1.1 Datasets.

  • Adressa emphasizes short-term interest in time-sensitive news, whereas Yelp emphasizes longer-term interest in less time-sensitive business choices.The datasets contain 3,664,225 interactions for Adressa and 3,014,421 for Yelp.Source: §4.1.1 · p. 6
  • Adressa is split into 63 daily morning, afternoon, and evening periods, while Yelp is split into 40 roughly quarterly periods with equal interaction counts.The training/validation/testing period counts are 48/5/10 for Adressa and 30/3/7 for Yelp.Source: §4.1.1 · p. 6

4.1.2 Baselines.

  • The MF retraining baselines are full retraining, fine-tuning on new interactions only, and SPMF, which adds reservoir-sampled historical interactions.The SPMF reservoir size is tuned among 7,000, 15,000, 30,000, and 70,000.Source: §4.1.2 · p. 6
  • The comparisons also include GRU4Rec and Caser, and use Recall@K and NDCG@K at K=5, 10, and 20, evaluating each interaction against 999 sampled non-interacted items.Methods are tuned on validation data; the MF methods share the log loss, and Recall@20 is the main tuning reference.Source: §4.1.3 · p. 6, §4.1.2 · p. 6

4.2 Performance Comparison (RQ1)

SML outperforms Full-retrain and Caser on both datasets, with an 18% advantage over Full-retrain on Yelp. Its period-wise results and retraining-time comparisons also show strong performance across testing periods and faster, stable retraining.

4.2.1 Overall Comparison.

  • Full-retrain performs better than Fine-tune on Yelp but worse on Adressa, reflecting differences in users' long- and short-term interests.Full-retrain captures long-term interest but can underemphasize recent data.Source: §4.2.1 · pp. 6–7
  • SML achieves the best recommendation performance on Adressa and Yelp, outperforming Full-retrain and Caser, including by 18% over Full-retrain on Yelp.The authors attribute this to adapting long- and short-term interests by optimizing the transfer network on next-period data.Source: §4.2.1 · pp. 6–7, Table 1 · p. 7

4.2.2 Period-wise Performance.

  • SML achieves the best Recall@10 in most testing periods, while Adressa shows larger performance fluctuations than Yelp.The authors connect Adressa's fluctuations to the news domain's timeliness and quickly changing user interests.Source: Figure 3 · p. 7, §4.2.2 · p. 7
  • On Yelp, SML is about 18 times faster than Full-retrain with stable retraining time across periods; SML-S is faster than Fine-tune and more accurate.Table 2 reports retraining time in seconds at each Yelp testing period.Source: Table 2 · p. 7, §4.2.3 · p. 7

4.3 Ablation Studies (RQ2)

Ablations test the transfer network's structure and sequential training choices. SML's hybrid transfer design and next-period optimization outperform their ablated alternatives, while disabling transfer updates or learning the transfer input directly reduces performance.

4.3 Ablation Studies (RQ2)

  • SML outperforms SML-CNN and SML-FC, supporting a hybrid transfer network with both CNN and fully connected layers.The authors attribute the gain to jointly considering dimension-wise and cross-dimension relations between the previous and new models.Source: §4.3 · p. 8
  • Without next-period optimization, SML-N performs worse than SML by 18.81% and 34.53% on average.SML-N optimizes on current-period rather than next-period data.Source: §4.3 · p. 8
  • Disabling transfer-network updates during testing reduces SML performance by 7.87% and 9.43%.The authors suggest user-interest drift may explain the drop and identify online recommender updates as a future direction.Source: §4.3 · p. 8
  • SML-FP, which learns the transfer input directly from current-period recommendation loss, fails to match SML on either dataset.This supports learning the new model through the transfer network's output.Source: §4.3 · p. 8

4.4 Hyper-parameter Studies (RQ3)

The hyper-parameter study varies the transfer component's CNN filter count and layer count. The best settings differ across datasets, and additional layers can reduce performance.

4.4 Hyper-parameter Studies (RQ3)

  • The study varies the transfer component's filter count and CNN depth to assess their effects on recommendation performance.Filter-count tests fix one convolution layer; depth tests use 10 filters in the first layer and five in additional layers.Source: §4.4 · p. 8, §4.4.1 · p. 8, §4.4.2 · p. 8

4.4.1 Number of CNN Filters.

  • On Yelp, performance is stable across filter counts, while Adressa performs better with more than 10 filters than with 6 or 8.The study varies filters from 6 to 14 with one convolution layer.Source: §4.4.1 · p. 8, Figure 5 · p. 8
  • Yelp performs best with one convolution layer, whereas Adressa performs best with two or three; adding more layers degrades performance on both.The authors attribute the degradation to overfitting on Yelp and note that optimal depth depends on dataset features.Source: §4.4.2 · p. 8, Figure 5 · p. 8

4.5 In-depth Analyses (RQ4)

The analyses examine where SML’s gains over Full-retrain arise and how its transfer filters use previous and current model parameters. On Yelp, gains center on interactions involving new users or items, while old-user/old-item performance remains comparable.

4.5 In-depth Analyses (RQ4)

  • The in-depth analyses investigate the sources of SML’s improvements over Full-retrain and examine CNN filters to interpret their rationale.Source: §4.5 · p. 8

4.5.1 Performance of Different Interaction Types.

  • SML’s Yelp gains are largest for new-item interactions and also improve new-user/old-item interactions, while old-user/old-item performance is comparable to Full-retrain.The authors attribute the gains mainly to recommendations for new users and items, without degrading performance on old-user/old-item interactions.Source: §4.5.1 · pp. 8–9, Figure 6 · p. 9
  • Adressa filters generally give greater weight to the current model, while two Yelp filters use opposite-signed previous- and current-model values to capture interest evolution.The analysis concerns filters in the item transfer network; similar patterns are reported for the user transfer network.Source: §4.5.2 · p. 9

5 RELATED WORKS

Prior work studies sequential interactions and meta-learning, while streaming recommendation addresses real-time updates. This paper positions periodic retraining as a distinct sequential training problem and uses future-period performance to guide it.

5.1 Recommendation on Sequential Data

  • Sequential recommendation models interaction sequences to predict next interactions, using approaches including FPMC, higher-order modeling, and recurrent neural networks [18].Source: §5.1 · p. 9
  • Periodic retraining treats sequence as a need to adapt models to new data, an approach the authors describe as orthogonal to sequential interaction modeling.They present their training paradigm as technically viable for sequential recommendation models.Source: §5.1 · p. 9
  • Streaming methods trade off model freshness and accuracy; SML instead refreshes without historical data and optimizes for future performance.Prior approaches include local updates and sampling historical interactions; the latter may require domain-specific heuristic samplers.Source: §5.1 · p. 9

5.2 Meta-Learning

  • Meta-learning uses prior experience to adapt to new tasks, and SML applies that idea by optimizing retraining against future validation data [9].Earlier recommendation work applied meta-learning to cold-start tasks [8] [24] [28] and algorithm or regularization selection [31, 4].Source: §5.2 · p. 9
  • Lifelong-learning methods target the current task while protecting earlier-task performance, unlike this work’s focus on serving the next-period task [23] [27].Source: §5.2 · p. 9

6 CONCLUSION

The paper formulates recommender retraining around generalization to next-period data and proposes SML to transfer a previous model using new interactions. Its current scope excludes online learning within serving periods, and broader model coverage and personalized meta-learning remain future work.

6 CONCLUSION

  • SML combines a transfer network that converts the previous model using newly collected data with sequential training that uses next-period data to learn the transfer network.The authors report experiments on two datasets of different properties, with analyses of effectiveness, efficiency, and rationality.Source: §6 · p. 10
  • The current work does not consider online learning strategies within serving periods; extending the paradigm to more recommender models and personalizing meta-learning are future directions.The proposed extensions include graph neural networks, factorization machines, and sequential recommender models.Source: §6 · p. 10

Related papers

Loading 2005.13258v1…