DisenReason: Behavior Disentanglement and Latent Reasoning for Shared-Account Sequential Recommendation

This paper proposes DisenReason, a novel two-stage sequential recommendation framework for shared accounts that overcomes the limitation of fixed user assumptions by disentangling collective account behaviors in the frequency domain to serve as a pivot for latent reasoning, thereby dynamically inferring the number of latent users and significantly improving recommendation accuracy.

The Gist

Imagine you and your family share a single Netflix or Amazon account. You all watch different things: maybe your dad watches sports, your sister watches reality TV, and you watch documentaries. To a computer, this looks like one person with very confusing, chaotic tastes.

Most recommendation systems today make a big mistake here: they assume one account equals one person. They try to guess what that single person wants next, but since the account is actually a mix of three or four different people, the guesses are often wrong.

Other researchers have tried to fix this by saying, "Okay, let's assume there are exactly 3 people on this account." But in real life, some families have 2 people, some have 5, and some accounts are shared by roommates who change over time. Guessing a fixed number is like trying to fit a square peg in a round hole.

This paper introduces a new system called DisenReason (short for "Disentangled Reasoning") that solves this problem without needing to know how many people are behind the account beforehand.

Here is how it works, using some simple analogies:

1. The Problem: The "Smoothie" vs. The "Fruit Salad"

Imagine the history of clicks on a shared account is like a smoothie. You've blended apples, bananas, and spinach together. If you take a sip (look at the last item clicked), you can't tell if the person who just drank it likes apples or spinach. It's just a messy mix.

Older systems tried to guess who was drinking the smoothie by looking at the last sip. But that only tells you what the last person tasted, not what the whole group likes.

2. Stage One: The "Magic Frequency Filter"

The authors realized that different people have different "rhythms."

• Low Frequency: One person might watch the same type of show every Tuesday night (a steady, slow rhythm).
• High Frequency: Another person might binge-watch random movies on a Saturday afternoon (a fast, bursty rhythm).

The system uses a mathematical tool called Fourier Transform (think of it as a sound equalizer on a stereo). Instead of looking at the smoothie as one messy blob, it separates the "sound" of the account into different frequencies.

• It filters out the "slow, steady" tastes.
• It filters out the "fast, bursty" tastes.

Once separated, it mixes these distinct "flavors" back together in a smart way to create a Unified Account Profile. This is no longer a messy smoothie; it's a clear picture of the entire household's collective behavior. This becomes the "starting point" for the next step.

3. Stage Two: The "Peeling an Onion" (Reasoning)

Now that the system has a clear picture of the account, it needs to figure out who is actually there. This is where the "Reasoning" part comes in.

Imagine the account profile is an onion. The system starts peeling it layer by layer:

1. Step 1: It looks at the onion and says, "I see a layer that looks like a sports fan." It peels that layer off and saves it as "User A."
2. Step 2: It looks at what's left. "Okay, now I see a layer that looks like a movie buff." It peels that off as "User B."
3. Step 3: It peels again. "Wait, this layer looks exactly like User A."

The Smart Stop: The system has a rule: "If the new layer looks too much like a layer I already found, stop peeling." This allows the system to count the users automatically. It doesn't guess "3 people"; it peels until it finds the last unique person and stops.

Why is this a big deal?

• No Guessing: It doesn't need you to tell it "There are 3 people." It figures it out on its own.
• Better Recommendations: Because it knows exactly who is on the account (e.g., "Oh, the sports fan just finished a game, let's recommend the next episode of the show they were watching"), the suggestions are much more accurate.
• Real World Ready: It works whether the account has 2 people or 10, and whether they are a family or roommates.

The Results

The researchers tested this on four different datasets (like real TV logs and Amazon shopping history). The new system, DisenReason, beat all the previous "state-of-the-art" methods. In some cases, it improved the accuracy of recommendations by over 12%.

In short: Instead of guessing how many people are sharing an account, this new AI listens to the "rhythm" of their activity, separates the different voices, and peels away the layers one by one to find the right people, resulting in much smarter recommendations for everyone in the household.

Technical Summary

Here is a detailed technical summary of the paper "DisenReason: Behavior Disentanglement and Latent Reasoning for Shared-Account Sequential Recommendation."

1. Problem Definition

Context: In digital services (streaming, e-commerce), multiple users often share a single account. This creates a Shared-Account Sequential Recommendation (SSR) challenge where the system must predict the next interaction based on a mixed sequence of behaviors from multiple latent users.

Limitations of Existing Methods:

• Fixed User Assumption: Most existing SSR methods assume a fixed, pre-determined number of latent users per account (e.g., exactly 2 or 3). This fails to adapt to real-world scenarios where the number of users varies dynamically (e.g., a family of 4 vs. a couple).
• Ineffective Reasoning Pivots: Recent "latent reasoning" techniques in sequential recommendation (SR) typically use the last item in a sequence as the starting point (pivot) for inference. In SSR, the last item only reflects the most recent user's behavior, ignoring the collective dynamics of the entire account. This makes it difficult to infer the number of hidden users or their distinct preferences.

Core Challenge: How to construct a unified representation of the shared account's collective behavior and adaptively infer the variable number of latent users without a pre-defined count.

---

2. Methodology: DisenReason Framework

The authors propose DisenReason, a novel two-stage reasoning framework that shifts the paradigm from "inferring preferences behind a user" to "inferring users behind an account."

Stage 1: Behavior Disentanglement for Pivot

The goal is to create a robust account-level reasoning pivot that represents the collective behavior of the shared account, rather than just the last interaction.

• Global Signal Propagation: Initial ID embeddings are enriched using LightGCN to capture global collaborative signals between accounts and items.
• Frequency-Domain Disentanglement:
  • The mixed interaction sequence is transformed into the frequency domain using the Fast Fourier Transform (FFT).
  • The frequency spectrum is partitioned into equal-width sub-bands. The authors hypothesize that different latent users exhibit distinct temporal rhythms (e.g., stable long-term preferences = low frequency; bursty short-term interactions = high frequency).
  • Each sub-band is transformed back to the time domain via Inverse FFT (IFFT) to isolate distinct behavioral patterns.
• Adaptive Fusion: A Mixture-of-Experts (MoE) mechanism learns adaptive weights to fuse these disentangled patterns. This creates a comprehensive, unified account representation ($\bar{H}_{seq}$).
• Pivot Extraction: The final hidden state of this fused sequence serves as the Reasoning Pivot ($\mathbf{p}$), representing the collective account state.

Stage 2: Progressive Residual Reasoning for Latent Users

The goal is to iteratively uncover the latent users behind the account using the pivot.

• Initialization: The reasoning process starts with the pivot $\mathbf{p}$ combined with a learnable reasoning position embedding ($\mathbf{r}_p$) to distinguish this stage from the disentanglement stage.
• Iterative Inference:
  • At each step $t$, a reasoning function $\phi(\cdot)$ (reusing the disentanglement module) extracts the most dominant undiscovered user representation $\mathbf{u}^{(t)}$.
  • Residual Update: The inferred user is subtracted from the current state: $\mathbf{r}^{(t)} = \mathbf{r}^{(t-1)} - \mathbf{u}^{(t)}$. This "residual" mechanism ensures subsequent steps focus only on unexplained signals, preventing redundant reasoning.
• Adaptive Termination: The loop continues until the semantic similarity between the newly inferred user and the previous one exceeds a threshold $\alpha$. This allows the model to autonomously determine the number of latent users ($T$) for each account.
• Final Aggregation: The final account representation is the average of all inferred user embeddings.

Training Objective

The model is optimized using:

1. Primary Loss: Cross-entropy loss for the next-item prediction.
2. Auxiliary Loss: Cross-entropy loss computed on intermediate user representations to encourage the reasoning process to learn meaningful user embeddings at every step.

---

3. Key Contributions

1. First Latent Reasoning in SSR: Introduces the concept of latent reasoning to the Shared-Account Sequential Recommendation task, addressing the challenge of identifying hidden users.
2. Two-Stage Framework: Proposes DisenReason, which uniquely combines frequency-domain behavior disentanglement (to create a collective pivot) with progressive residual reasoning (to infer variable user counts).
3. Frequency-Domain Insight: Demonstrates that mixed user behaviors can be effectively disentangled by analyzing their frequency components (stable vs. bursty patterns), providing a new perspective for SSR.
4. Adaptive User Counting: Eliminates the need for pre-defined user counts, allowing the model to dynamically infer the number of users based on semantic similarity thresholds.

---

4. Experimental Results

The model was evaluated on four benchmark datasets (two real-world smart TV datasets: HV-E, HV-V; and two synthetic Amazon datasets: HA-M, HA-B).

• Performance: DisenReason consistently outperformed all state-of-the-art baselines (including traditional SR models like SASRec and SSR models like LightGC2N, $\pi$-Net, and DA-GCN).
  • Improvements: Achieved relative improvements of up to 12.56% in MRR@5 and 6.06% in Recall@20 compared to the best baselines.
• Ablation Studies:
  • Removing Behavior Disentanglement caused the most significant performance drop, proving the necessity of a collective pivot over a single last-item pivot.
  • Removing LightGCN degraded performance, highlighting the importance of global collaborative signals.
  • Residual Operations provided semantic refinement, preventing redundant user inference.
• Robustness: The model showed superior stability across varying sequence lengths and training data proportions (performing well even with only 20% of training data).
• Inference Validation: Visualizations confirmed that DisenReason successfully infers different numbers of latent users for different accounts, aligning with the variability of real-world sharing scenarios.

---

5. Significance and Implications

• Theoretical Impact: Challenges the rigid assumption of fixed user counts in SSR. It bridges signal processing (Fourier Transform) with latent reasoning, offering a new mathematical approach to disentangle mixed user behaviors.
• Practical Application: Provides a scalable solution for streaming and e-commerce platforms where account sharing is ubiquitous. It enables more personalized recommendations without requiring explicit user identification or manual configuration of user counts.
• Future Directions: The authors suggest future work on adaptive frequency decomposition (to handle fine-grained boundaries) and decoupling the two stages to further improve interpretability.

In summary, DisenReason represents a significant advancement in handling the complexity of shared accounts by treating the problem as a signal disentanglement and iterative reasoning task, effectively solving the "who is behind the account" problem in a dynamic and data-driven manner.