Source-linked AI summary
Long-tail Session-based Recommendation
Siyi Liu, Yujia Zheng
TL;DR
Session-based recommendation methods largely prioritize accuracy without explicitly addressing long-tail items, whose recommendations can support diversity and serendipity. TailNet learns session preferences between short-head and long-tail items and softly adjusts recommendations; on two datasets, it reports the best long-tail performance among state-of-the-art deep learning methods and the best accuracy among traditional methods.
Problem
Session-based recommendation methods largely prioritize accuracy without explicitly addressing long-tail items, despite their relevance to diversity and serendipity.
Source: §1 · p. 1Method
TailNet uses session representations to learn preferences between short-head and long-tail items, generating rectification factors that softly adjust recommendations.
Source: §1 · pp. 1–2, §2.1 · p. 2Results
On both datasets, TailNet has the best long-tail performance among state-of-the-art deep learning methods and the best accuracy among traditional methods.
Source: §4.1 · p. 5Takeaways & Limitations
The comparison indicates that accuracy alone may not suffice to assess recommendation quality and that balancing accuracy with long-tail performance remains challenging.
Source: §4.1 · p. 5The authors cannot exclude that TailNet's competitive performance is largely due to its use of Pareto-principle prior knowledge.
Source: §4.2 · p. 5
Abstract
from arXiv · showhide
Session-based recommendation focuses on the prediction of user actions based on anonymous sessions and is a necessary method in the lack of user historical data. However, none of the existing session-based recommendation methods explicitly takes the long-tail recommendation into consideration, which plays an important role in improving the diversity of recommendation and producing the serendipity. As the distribution of items with long-tail is prevalent in session-based recommendation scenarios (e.g., e-commerce, music, and TV program recommendations), more attention should be put on the long-tail session-based recommendation. In this paper, we propose a novel network architecture, namely TailNet, to improve long-tail recommendation performance, while maintaining competitive accuracy performance compared with other methods. We start by classifying items into short-head (popular) and long-tail (niche) items based on click frequency. Then a novel is proposed and applied in TailNet to determine user preference between two types of items, so as to softly adjust and personalize recommendations. Extensive experiments on two real-world datasets verify the superiority of our method compared with state-of-the-art works.
1 INTRODUCTION
Long-tail recommendations are valued for diversity and serendipity, but session-based recommendation faces sparsity and lacks user profiles, making long-tail methods difficult to apply. TailNet adds a learned preference mechanism to adjust recommendations between short-head and long-tail items, with reported strong long-tail performance and competitive accuracy.
1 INTRODUCTION
- Long-tail recommendations can increase list diversity and serendipity, while also offering business benefits such as higher marginal profit and one-stop shopping convenience.The introduction connects these benefits to users and businesses, respectively.Source: §1 · p. 1
- Existing session-based methods largely prioritize accuracy, while data sparsity biases recommendation toward popular items and prior long-tail methods may sacrifice accuracy or require unavailable side information.The paper describes popular-item recommendation as a conservative but effective way to improve accuracy.Source: §1 · p. 1
- Session-based long-tail recommendation is difficult because sessions lack profiles and side information, traditional methods often ignore sequence order, and session data are sparse.The paper notes that traditional methods' accuracy can drop in session-based settings.Source: §1 · p. 1
- TailNet learns a preference between short-head and long-tail items from session representations and generates rectification factors that softly adjust the recommendation mode.The preference mechanism is jointly trained with TailNet end-to-end.Source: §1 · pp. 1–2
- Experiments on real-world datasets report state-of-the-art long-tail performance while maintaining relatively high accuracy.The introduction describes experiments on real-world datasets but does not give dataset names or metric values here.Source: §1 · p. 2
2 MODEL
TailNet encodes timestamp-ordered sessions and distinguishes short-head from long-tail items, then uses a learned preference mechanism to adjust item scores before producing recommendation probabilities.
2.1 Notations
- TailNet represents sessions as timestamp-ordered item sequences, encodes them with a GRU, and divides items into short-head and long-tail sets.Source: §2.1 · pp. 2–3
2.2 Preference Mechanism
- A preference mechanism processes session representations with attention and a sigmoid function to produce separate rectification factors for short-head and long-tail items.Source: §2.2 · p. 3
2.3 Session Pooling and Soft Adjustment
- Session pooling derives global and local interest representations, then multiplies each item’s score by its type-specific factor and applies softmax for recommendation.The model is trained jointly with the preference mechanism using back-propagation through time and summed cross-entropy loss.Source: §2.3 · p. 3
3 EXPERIMENTS
Experiments evaluate TailNet on YOO-CHOOSE and 30MUSIC against frequency-, sequence-, neighborhood-, attention-, matrix-factorization-, Markov-chain-, and graph-based baselines, using accuracy and long-tail diversity metrics.
3 EXPERIMENTS
- The evaluation uses YOO-CHOOSE and 30MUSIC, with the last day as test data and previously unseen test items filtered out.The YOO-CHOOSE experiments use the recent one-quarter fraction, and items are split into short-head and long-tail groups according to the Pareto Principle.Source: §3 · p. 3
- Baselines include POP, S-POP, GRU4REC, NARM, Item-KNN, RepeatNet, STAMP, FPMC, BPR-MF, and SR-GNN.The methods represent frequency-based, RNN-based, neighborhood-based, repeat-explore, attention-based, matrix-factorization or Markov-chain, and GNN-based approaches.Source: §3 · p. 3
- Table 1 compares TailNet with baseline methods on two datasets, while Figure 2 compares long-tail performance on those datasets.Boldface marks the best neural-network result and † marks the best result across all methods; selected neural-network Coverage, Tail_Coverage, and Tail values are those reported when most accurate.Source: Table 1 · p. 4, Figure 2 · p. 4, §3 · p. 4
- Recall@K and MRR@K measure accuracy, while Coverage@K, Tail-Coverage@K, and Tail@K evaluate long-tail recommendation.Coverage@K counts distinct recommended items, Tail-Coverage@K counts distinct long-tail items, and Tail@K measures long-tail items per recommendation list.Source: §3 · p. 4
4 RESULTS AND ANALYSIS
Across two datasets, TailNet combines stronger long-tail recommendation with competitive accuracy, and its preference mechanism also improves long-tail performance when applied to STAMP and SR-GNN. The reported comparison with a hard-adjustment variant attributes TailNet’s advantage primarily to soft adjustment rather than the Pareto-based item split.
4.1 Comparison with Baseline Methods
- Across both datasets, TailNet led deep-learning methods in long-tail performance, led traditional methods in accuracy, and remained comparable on MRR@20 and Recall@20.The authors report that baseline methods did not balance long-tail recommendation and accuracy as well as TailNet.Source: §4.1 · p. 5
4.2 The Effect of Soft Adjustment
- TailNet outperforms TailNet-proportion on all metrics, supporting soft adjustment rather than the Pareto-based hard quota as the main contributor.TailNet-proportion combines a session-proportion-based number of short-head and long-tail items in the top-k list, then sorts them by score.Source: §4.2 · p. 5
4.3 Application of Preference Mechanism in Other Models
- On YOOCHOOSE 1/4, adding the preference mechanism to STAMP and SR-GNN improves long-tail recommendation with only a small range of accuracy fluctuation.The comparison evaluates both models with and without the mechanism.Source: Table 2 · p. 5, §4.3 · p. 5