Source-linked AI summary
Deep Social Collaborative Filtering
Wenqi Fan, Yao Ma, Dawei Yin, Jianping Wang, Jiliang Tang, Qing Li
TL;DR
Deep social recommenders often rely on direct neighbors, treat neighbor information uniformly, and do not model neighbors’ opinions toward items. DSCF combines item-aware social sequences, neighbor opinions, and a Bi-LSTM with attention; experiments on two real-world datasets show its effectiveness.
Problem
Deep social recommenders often use only direct neighbors, treat neighbor information equally across recommendations, and omit neighbors’ opinions toward items.
Source: §1 · p. 2Method
DSCF uses random walks to generate item-aware social sequences, models neighbors’ opinions in user-item interactions, and applies a Bi-LSTM with attention to sequence features.
Source: §5 · p. 8Results
Comprehensive experiments on two real-world datasets show the effectiveness of DSCF.
Source: §5 · p. 8Takeaways & Limitations
Item-aware sequences and opinion information play a crucial role in modeling social information for recommendations.
Source: §5 · p. 8
Abstract
from arXiv · showhide
Recommender systems are crucial to alleviate the information overload problem in online worlds. Most of the modern recommender systems capture users' preference towards items via their interactions based on collaborative filtering techniques. In addition to the user-item interactions, social networks can also provide useful information to understand users' preference as suggested by the social theories such as homophily and influence. Recently, deep neural networks have been utilized for social recommendations, which facilitate both the user-item interactions and the social network information. However, most of these models cannot take full advantage of the social network information. They only use information from direct neighbors, but distant neighbors can also provide helpful information. Meanwhile, most of these models treat neighbors' information equally without considering the specific recommendations. However, for a specific recommendation case, the information relevant to the specific item would be helpful. Besides, most of these models do not explicitly capture the neighbor's opinions to items for social recommendations, while different opinions could affect the user differently. In this paper, to address the aforementioned challenges, we propose DSCF, a Deep Social Collaborative Filtering framework, which can exploit the social relations with various aspects for recommender systems. Comprehensive experiments on two-real world datasets show the effectiveness of the proposed framework.
1 INTRODUCTION
The introduction motivates social relations as a source of preference information, then identifies gaps in how deep social recommenders use that information. DSCF is proposed to address those gaps by using distant social information and neighbors’ opinions, with experiments on two real-world datasets.
1 INTRODUCTION
- Social relations may help describe user preferences through influence and homophily, and prior work reports recommendation gains from social information.The introduction also describes information diffusing through social networks.Source: §1 · p. 1
- Prior deep social recommendation approaches include GraphRec, which aggregates user-item and social interactions, and DASO, which uses adversarial learning.The introduction places these methods among work using deep neural networks for social recommendations.Source: §1 · pp. 1–2
- Most deep social recommenders use only direct neighbors, treat their information equally across recommendations, and omit neighbors’ opinions toward items.The paper argues that distant neighbors, item-relevant interactions, and positive versus negative opinions can each matter.Source: §1 · p. 2
- DSCF uses deep neural networks to extract information from distant social relations and introduces a way to model users’ opinions in interactions.The authors also report comprehensive experiments on two real-world datasets.Source: §1 · p. 2
2 THE PROPOSED FRAMEWORK
DSCF combines random-walk social sequences, opinion-aware interaction embeddings, sequence learning, and rating prediction to use distant neighbors and item-relevant information.
2.1 The Random walk layer: generating item-aware social sequences
- Random walks capture direct and distant neighbors, then pair each with its most similar interacted item to target item v using cosine similarity over NeuMF item embeddings.Multiple social sequences are generated to explore the target user's neighborhood.Source: §2.1 · p. 3
2.2 The embedding layer: modeling user-item interactions
- Each interaction's rating is embedded and fused with user and item embeddings by an MLP, representing neighbors' differing opinions.The rating embedding corresponds to the interaction's non-zero rating score.Source: §2.2 · p. 4
2.3 The sequence learning layer: learning representation for item-aware social sequences
- A Bi-LSTM reads each fused interaction sequence in both directions, and neighbor-level attention weights interaction embeddings when forming its sequence representation.The attention context vector acts as a learned query for informative neighbor-item interactions.Source: §2.3 · p. 4
- A second attention mechanism weights the item-aware sequence representations to form a unified representation for predicting the target rating.The sequence-level context vector queries which item-aware sequence is informative.Source: §2.3 · pp. 4–5
2.4 The output layer: rating prediction
- The output MLP combines the target user's embedding with the unified social representation, then uses the item embedding to predict the rating.The output layer takes the user embedding, item embedding, and unified item-aware social representation as inputs.Source: §2.4 · p. 5
2.5 Model Training
- Training optimizes rating-prediction error over observed interactions, jointly learns randomly initialized user, item, and rating embeddings, and uses Adam with dropout.Dropout is adopted to alleviate overfitting.Source: §2.5 · p. 5
3 EXPERIMENTS
The experiments assess the model through experimental settings, performance comparisons, and studies of component impact.
- The section moves from experimental settings to recommender comparisons, then studies the impact of DSCF components.Source: §3 · p. 5
3.1 Experimental Settings
Experiments use Ciao and Epinions, evaluate rating prediction with MAE and RMSE, and compare DSCF with traditional and neural recommendation baselines.
3.1.1 Datasets.
- Ciao and Epinions provide ratings and social connections; users rate items from 1 to 5, with dataset statistics reported in Table 1.Source: §3.1.1 · p. 5, Table 1 · p. 5
3.1.2 Evaluation Metrics.
- MAE and RMSE measure predictive accuracy, with smaller values indicating better accuracy; even small improvements may affect top-few recommendations.Source: §3.1.2 · p. 5
3.1.3 Baselines.
- Traditional baselines include PMF, SoRec, SoReg, and SocialMF, spanning interaction-only factorization and methods incorporating social information.Source: §3.1.3 · p. 5
- Additional baselines are TrustMF, NeuMF, DeepSoR, and GCMC+SN; NeuMF uses squared loss for rating prediction, and GCMC+SN uses node2vec user embeddings.Source: §3.1.3 · p. 6
- PMF and NeuMF are interaction-only models; other listed methods are social recommenders, and DSCF is compared with DeepSoR and GCMC+SN.Source: §3.1.3 · p. 6
- Training uses 80% or 60% of data, with validation and test splits, and tunes embedding size, batch size, and learning rate.The hidden-layer size matches the embedding size, ReLU is used, and training stops after five successive validation-RMSE increases.Source: §3.1.4 · p. 6
3.2 Performance Comparison
DSCF is reported to outperform every baseline in rating prediction; the section also identifies component and Bi-LSTM analyses on Ciao.
- Figure 3 presents a component analysis on Ciao, with DSCF-* denoting removal of component *.Source: Figure 3 · p. 6
- Figure 4 presents the effect of the Bi-LSTM model on Ciao.Source: Figure 4 · p. 6
- DSCF consistently outperforms all baseline methods in the reported rating-prediction comparison.Table 2 reports RMSE and MAE on Ciao and Epinions, but supplied passages give no metric values.Source: §3.2 · p. 6
3.3 Model Component Analysis
Component analyses examine how opinions, item relevance, and sequence modeling contribute to DSCF. The reported comparisons favor retaining opinions and item-aware interactions, using attention, and modeling sequence order.
Item-aware Social Sequences with Opinions
- Ignoring opinions in DSCF-Opinion significantly reduces performance, supporting the role of neighbors’ ratings in modeling user-item interactions.Source: §3.3 · p. 7
Item-aware Social Sequences
- DSCF-Item&Opinion performs worse than DSCF and DSCF-Opinion, supporting sequences that retain interactions with recommendation-relevant items.Source: §3.3 · p. 7
Attention Mechanisms
- DSCF-ATT performs worse than DSCF; the analysis attributes this to unequal contributions from interactions and item-aware sequences.DSCF-ATT removes the attention mechanisms.Source: §3.3 · p. 7
Bi-LSTM
- On Ciao, averaging and shuffling significantly reduce performance; the results favor Bi-LSTM for capturing heterogeneous influence by distance.Averaging assumes equal influence, while shuffling removes the sequence order.Source: §3.3 · p. 7
3.4 Parameter Analysis
The parameter analysis varies sequence length and sequence count on Ciao. Both can extend neighborhood information, but excessive length or count risks introducing noise.
- The analysis varies one parameter at a time while fixing the others and reports results only on Ciao.Source: §3.4 · p. 7
Effect of the length of sequences l
- Performance first tends to improve as sequence length increases beyond direct neighbors, then degrades when longer sequences introduce noise.Source: §3.4 · p. 8
Effect of the number of sequences H
- More sequences generally explore users’ neighborhoods more fully, but generating too many can introduce noise.Source: §3.4 · p. 8, Figure 6 · p. 8
4 RELATED WORK
Related work spans social recommendation and neural recommendation, including methods that integrate social signals in different ways. The paper situates DSCF among deep social recommenders and distinguishes its setting from related approaches.
- Social recommendation methods use social relations; SoRec shares user latent vectors across rating and social matrices, while SoDimRec models relation heterogeneity and weak dependencies.Source: §4 · p. 8
- Neural recommenders use deep networks for auxiliary features or user-item nonlinearity, as in NeuMF; NSCR adds social graph regularization for cross-domain ranking.Source: §4 · p. 8
- NSCR [40] uses an external social network for cross-domain recommendation, whereas DSCF learns from user-user interactions on the same e-commerce platform; ARSE targets dynamic preferences.Source: §4 · p. 8
- Deep social methods include DeepSoR, DASO, and GraphRec; GraphRec aggregates user-item interactions and direct social neighbors, while the paper says these methods do not fully exploit social networks.Source: §4 · p. 8
5 CONCLUSION AND FUTURE WORK
The conclusion presents DSCF as combining item-aware social sequences, neighbors’ opinions, and Bi-LSTM attention. Experiments on two real-world datasets report effectiveness, while item side information remains a future direction.
- DSCF uses random walks to generate item-aware sequences containing direct and distant neighbors, and models neighbors’ opinions with Bi-LSTM attention.Source: §5 · p. 8
- Experiments identify item-aware sequences and opinion information as important to modeling social information and report effectiveness on two real-world datasets.Source: §5 · p. 8
- Item similarity currently uses only user-item interactions; incorporating textual descriptions and image content is proposed as future work.Source: §5 · p. 8