Sequential Recommendation with Self-Attentive Multi-Adversarial Network

• MLE-based recommenders combine context factors with sequence representations or a black-box module, obscuring when and how each factor affects recommendations.Source: §1 · p. 1
• MFGAN uses a Transformer-based generator for next-item prediction and factor-specific discriminators to evaluate generated recommendations.Discriminator rewards guide generator learning through policy-gradient training.Source: §1 · p. 2
• The proposed multi-discriminator structure decouples factors and is designed to improve sequential recommendation performance.The paper also analyzes the architecture's effectiveness and stability.Source: §1 · p. 2
• Experiments on three real-world datasets show MFGAN benefits recommendation effectiveness and interpretability over state-of-the-art methods.Source: §1 · p. 2

• Each item is associated with m kinds of contextual information, corresponding to m factors.The paper gives artist, album, and popularity as examples in music recommendation.Source: §2 · p. 2
• Given a user's chronologically ordered item interactions and item context, the task is to predict the item they will interact with next.Source: §2 · p. 2

• The methodology section introduces the MFGAN framework and its generator and discriminators, then discusses their design and training process.Source: §3 · p. 2

3.1.1 Basic Components.

• The generator predicts next items from the historical sequence, while multiple discriminators judge generated sequences using their respective factors.The generator does not use item-side context; examples of discriminator factors include category, popularity, artist, and album.Source: §3.1.1 · p. 2, Figure 1 · p. 3

3.1.2 Overall Procedure.

• At each step, factor-specific discriminators evaluate the generated sequence and return feedback to guide the generator's next learning round.Discriminators are updated using generated sequences and actual user behavior sequences.Source: §3.1.2 · pp. 2–3
• Separate discriminators let the framework incorporate factors flexibly, keep the generator focused on sequence generation, and distinguish each factor's effect for interpretation.The paper also says this separation can make the model more resistant to useless or noisy context information.Source: §3.1.3 · p. 3

3.2 The Generator Component

• The generator stacks an embedding layer, self-attention blocks, and a prediction layer to generate target items.Its input is a sequence of historical interactions.Source: §3.2 · p. 3

3.2.1 Embedding Layer.

• Historical item embeddings are combined with learnable position encodings to form the generator’s input representations.The item embedding matrix maps items to dense vectors, which are summed with position encodings.Source: §3.2.1 · p. 3

3.2.2 Self-attention Block.

• Multi-head self-attention extracts sequence information from different representation subspaces, with future-position connections masked out.A point-wise feed-forward network adds non-linearity to the self-attention block.Source: §3.2.2 · pp. 3–4

3.2.3 Prediction Layer.

• The final generator layer applies softmax to calculate user preferences over the item set.The prediction uses the maintained item embedding matrix.Source: §3.2.3 · p. 4

3.3 Factor-specific Discriminator Components

• Each factor has an independent discriminator that classifies generated versus actual recommendation sequences as rational or not.The framework decouples context information from the generator by using factor-specific discriminators.Source: §3.3 · p. 4
• Each discriminator embeds discretized factor values with bin-specific vectors and adds positional encodings.The factor values are mapped to bins, and their embeddings are concatenated by input sequence.Source: §3.3 · p. 4

3.3.2 Architecture.

• A discriminator uses one self-attention block to avoid becoming too strong to provide suitable feedback to the generator.Its architecture is intentionally relatively weak compared with the generator.Source: §3.3.2 · p. 4
• Unlike the generator, a discriminator can attend to subsequent positions, modeling interactions between any two positions in the sequence.The discriminator removes the attention mask and uses overall sequential characteristics for its judgment.Source: §3.3.2 · p. 4
• An MLP turns each discriminator’s attention output into a rationality score, and the scores provide supervision signals for the generator.A rationality score reflects the discriminator’s estimate that a sequence comes from the actual data distribution.Source: §3.3.2 · p. 4

3.4 Multi-adversarial Training

• Discrete item sampling prevents direct gradient descent on the original GAN formulation, so recommendation is formalized as reinforcement learning.The state is the previously generated subsequence, the action is the next item, and the generator defines the policy.Source: §3.4.1 · p. 4

3.4.1 RL-based Formalization.

• The expected return combines discriminator rationality scores using λ-parameterized softmax weights.λ is a tuning parameter, and discriminator scores provide guidance from different factor perspectives.Source: §3.4.1 · p. 5
• The generator maximizes expected reward with the classic policy-gradient method.Its objective is to maximize the expected reward at each step.Source: §3.4.2 · p. 5

3.4.2 Learning Algorithm.

• Training pretrains the generator and discriminators, then updates the generator using discriminator rewards and policy gradients.For each discriminator update, actual sequences serve as positives and generated sequences as negatives.Source: §3.4.2 · p. 5

3.5 Discussion and Analysis

• A λ-parameterized softmax combines discriminator rewards as minimum, maximum, mean, or a soft combination, depending on λ.The minimum and maximum arise as λ approaches −∞ and +∞; λ = 0 gives the mean.Source: §3.5 · p. 5
• Minimum rewards can indicate where to decrease generator probability without identifying where to increase it, while maximum rewards can let a prominent factor dominate and leave others undertrained.The maximum strategy alleviates the stated training issue with the minimum strategy.Source: §3.5 · pp. 5–6
• Multiple discriminators make generator learning more stable; the authors adopt mean rewards because the last two reward cases show no significant difference.The generator's Q-function is zero only when all discriminators give zero reward.Source: §3.5 · p. 6
• Discriminator rewards can be treated as the importance of influencing factors at each time step, and the framework leaves room for more advanced generator or discriminator designs.The approach decouples factors from the prediction component.Source: §3.5 · p. 6

4.1 Dataset Construction

• The experiments use MovieLens-1M, a 10,000-user Yahoo! subset, and Steam, with user sequences sorted by time and filtered using dataset-specific k-core thresholds.The thresholds are k = 5, 10, and 5, respectively; the final item is test data, the preceding item validation data, and earlier interactions training data.Source: §4.1 · p. 6

4.2 Experimental Settings

• Evaluation uses MRR, NDCG@10, and HR@10, ranking each test positive against 100 sampled unobserved items and averaging scores over test users.MFGAN is implemented with two generator self-attention blocks and one discriminator block, with separate pretraining followed by alternating adversarial training.Source: §4.2.1 · p. 6, §4.2.2 · p. 6

4.3 Results and Analysis

• MFGAN consistently outperforms the competitive baselines in sequential recommendation, including when its generator uses no context directly at test time.The comparison is reported in Table 2; the authors attribute the improvement over the self-attentive generator to multi-adversarial training guided by factor information.Source: Table 2 · p. 7, §4.3 · p. 7

4.2.3 Comparison Methods.

• Among sequential baselines, SASRec performs best, while GRUF improves on GRU with context; MFGAN is compared with SASRec and other competitive methods.The authors omit other context-aware baselines because they found the Transformer architecture superior to other sequence neural networks for this task.Source: §4.2.3 · pp. 6–7, §4.3 · p. 7

4.4.1 Effect of Different Factors.

• Removing factors generally hurts MFGAN, except developer on Steam; knowledge graph, artist, and popularity are the most useful factors on MovieLens-1M, Yahoo!, and Steam, respectively.The factor-removal experiments report NDCG@10 while retaining the other factor information.Source: §4.4.1 · p. 8, Figure 2 · p. 8

4.4.2 Comparisons of Model Variants.

• The Uni-directional, SDSF, and SDAF variants all underperform complete MFGAN, supporting the value of bidirectional discriminators and multiple-discriminator adversarial training.The reported numerical improvement from bidirectionality appears small; the comparison is on MovieLens-1M.Source: §4.4.2 · p. 8, Table 3 · p. 8
• About 2,000 epochs: three- and four-discriminator losses stabilize, while the other configurations remain unstable after 4,000 epochs.The authors also observe larger loss with more discriminators and speculate that this may act like regularization against overfitting.Source: §4.4.2 · p. 8, Figure 3 · p. 8

4.5 Qualitative Analysis on the Recommendation Interpretability

• In the MovieLens case, popularity is prominent early and category at selected later steps; in the Yahoo! case, artist signals become dominant when a previously heard artist recurs.The examples illustrate factor-specific reward signals over time; the authors contrast this decoupling with methods that combine context into unified representations.Source: §4.5 · p. 9, Figure 4 · p. 9

Sequential Recommendation

• Prior sequential recommenders include Markov-chain methods, recurrent networks and variants, context-aware approaches, and self-attention methods such as SASRec.The cited examples include matrix factorization with first-order Markov chains, GRU recommendation [10], and self-attention recommendation [14].Source: §5 · p. 9

Generative Adversarial Network

• GAN work spans framework and training variants as well as recommendation applications, including information retrieval and adversarial approaches to recommendation diversity.The passage names IRGAN [28], Chae et al., and personalized determinantal point process.Source: §5 · p. 9
• The paper presents its work as the first application of adversarial training to sequential recommenders, using a multi-adversarial network to trace individual factor effects over time.The stated aim is improved recommendation interpretability through explicit characterization of factor effects.Source: §5 · p. 9

6 CONCLUSIONS

• MFGAN uses a user-sequence generator to propose next items and multiple factor-specific discriminators to evaluate generated subsequences.Experiments on three real-world datasets are reported to outperform several competitive baselines.Source: §6 · p. 10
• Multiple discriminators are reported as useful for stabilizing adversarial training and enhancing recommendation interpretability.The conclusion identifies separate discriminator design as the current simple setting and proposes principled sharing of discrimination information as future work.Source: §6 · p. 10