Debiasing Sequential Recommendation with Time-aware Inverse Propensity Scoring

This paper proposes Time-aware Inverse Propensity Scoring (TIPS), a novel method that addresses selection and exposure biases in sequential recommendation by leveraging counterfactual reasoning to capture temporal dynamics and sequential dependencies, thereby significantly enhancing the performance of various existing recommenders.

The Gist

Imagine you are a chef trying to learn what your customers truly love to eat. You have a notebook where you write down every dish they order.

The Problem: The "Silent Menu" Bias
In the real world, you only write down what customers order. You don't write down the dishes they looked at but decided not to order.

• Scenario A: A customer sees a spicy curry on the menu, thinks "No thanks," and orders a salad. You only see the salad. You might wrongly think, "Oh, this customer hates spicy food!" (This is Selection Bias: assuming no order means no interest).
• Scenario B: You never put the spicy curry on the menu for this customer because you think they only like salads. They never get a chance to try it. You assume they don't like it because they never ordered it. (This is Exposure Bias: assuming they didn't see it, so they don't care).

Most current recommendation systems (like those on Netflix or Amazon) are like this chef. They only look at what you clicked or bought. They ignore everything you didn't click, assuming you didn't like it. This creates a feedback loop where the system only shows you more of the same, missing out on your true, hidden interests.

The Old Solution: The "Static Scale"
Scientists have tried to fix this using a method called Inverse Propensity Scoring (IPS). Think of IPS as a "fairness scale."

• If a rare, obscure item is clicked, the scale says, "Wow! This person really likes this rare thing! Give it double points!"
• If a super popular item is clicked, the scale says, "Well, everyone clicks this. Give it normal points."

The problem with the old IPS is that it's static. It treats every click as if it happened in a vacuum, ignoring when it happened or what the user was doing before. It's like a judge who gives a sentence based only on the crime, ignoring the defendant's history or the time of day. It doesn't understand that your taste in music changes after a breakup, or that you might buy a phone case only after buying a new phone.

The New Solution: HyperG (The "Time-Traveling Chef")
This paper introduces a new framework called HyperG (specifically using Time-aware Inverse Propensity Scoring or TIPS). Imagine HyperG as a chef who doesn't just look at your order history, but also uses a time machine to ask "What if?" questions.

Instead of just looking at the salad you ordered, HyperG runs three simulations for every single item you interacted with:

1. The "Similar Item" Test: "What if we had shown you a similar spicy curry instead of the salad? Would you have clicked that?"
2. The "Trending Item" Test: "What if we had shown you the most popular dish of the week? Would you have clicked that?"
3. The "Time Travel" Test: "What if we had shown you this exact salad 10 minutes later? Would your mood have changed?"

By creating these "What if" scenarios (called Counterfactuals), HyperG builds a much clearer picture of your true preferences. It learns to distinguish between:

• "I didn't click because I wasn't interested."
• "I didn't click because I never saw it."
• "I didn't click because I was in a hurry at that specific moment."

How It Works in Plain English

1. Dual Vision: The system keeps two separate mental lists. One list tracks what you clicked (your interest), and another tracks what was shown to you (the exposure). It doesn't mix them up, which prevents confusion.
2. Time Awareness: It understands that a click from yesterday is more important than a click from last year. It weighs recent actions heavier, just like you remember what you ate for dinner last night better than what you ate last month.
3. The Fairness Adjustment: When the system trains itself, it uses these "What if" simulations to adjust the score of every item. If an item was rarely shown but you clicked it, the system gives it a massive boost, realizing, "This user really loves this hidden gem!"

The Result
The paper shows that when you plug this "Time-Traveling Chef" (HyperG) into existing recommendation systems, they get much better at guessing what you actually want.

• They stop showing you the same old things.
• They start suggesting items you might have loved if you'd just seen them at the right time.
• They work better on huge datasets (like millions of users) because they can spot complex patterns in how your interests change over time.

In Summary
Current recommenders are like a blindfolded waiter who only serves you what you've ordered before. HyperG takes off the blindfold, looks at the whole menu, asks "What if we tried this?", and uses the passage of time to understand that your tastes evolve. It turns a biased, static guess into a dynamic, fair prediction of what you'll love next.

Technical Summary

Here is a detailed technical summary of the paper "Debiasing Sequential Recommendation with Time-aware Inverse Propensity Scoring" (referred to as HyperG or TIPS in the text).

1. Problem Statement

Sequential Recommendation (SR) systems predict a user's next interaction by modeling the temporal order of historical behaviors. However, existing SR models (both traditional sequential models like RNNs/Transformers and generative models) suffer from two critical biases due to the lack of item exposure logs:

1. Exposure Bias: Items that were never exposed to the user are implicitly treated as "not of interest." The model cannot distinguish between an item the user didn't see and an item the user saw but ignored.
2. Selection Bias: Items that were exposed but not clicked are treated as negative samples, ignoring the possibility that the user might have been interested if the context or timing had been different.

Limitations of Current Solutions:
Existing debiasing methods often rely on Inverse Propensity Scoring (IPS) to reweight interactions based on exposure probability. However, conventional IPS is static:

• It treats interactions independently, ignoring sequential dependencies (e.g., buying an iPhone 17 increases the likelihood of buying a case later).
• It ignores temporal dynamics (e.g., user interests shift over time, and item popularity changes).
• It fails to capture the causal chain where exposure ($E$) influences interaction ($C$) and user preference ($U$) evolves over time.

2. Methodology: Time-aware Inverse Propensity Scoring (TIPS / HyperG)

The authors propose HyperG, a model-agnostic plug-in framework that integrates Time-aware Inverse Propensity Scoring (TIPS) to correct biases without requiring explicit exposure logs. The framework is built on a Structural Causal Model (SCM) and consists of three main components:

A. Dual Encoding Strategy

To disentangle user preferences from exposure mechanisms, the model maintains two separate embedding tables for items:

• Interaction Embedding ($H^{(C)}$): Captures item semantics and collaborative filtering signals derived from explicit feedback (clicks/purchases).
• Exposure Embedding ($H^{(E)}$): Captures factors influencing whether an item is shown (e.g., popularity, promotions, display strategy).
• Time Embeddings: Normalized time intervals between interactions are encoded to capture temporal dynamics.

B. Counterfactual Sample Construction

Since exposure logs are missing, the model constructs counterfactual examples to estimate the exposure distribution $P(E=1|u)$. For every factual interaction $(u, v_i, t_i)$, three counterfactual positive exposure samples (where $E=1$ but $C=0$) are generated:

1. Similar Items ($v_{sim}$): Items with similar exposure embeddings (based on $H^{(E)}$) are assumed to be exposed at the same time.
2. Popular Items ($v_{pop}$): Highly popular items are assumed to be exposed concurrently with the factual item.
3. Same Item, Different Time ($t^*$): The same item is exposed at a slightly perturbed time step to model short-term re-ranking effects.

These counterfactuals serve as positive samples for the exposure estimation model but negative samples for the user preference model, helping the system learn what could have happened.

C. Exposure Influence Modeling (E $\to$ C and E $\to$ U)

• Exposure Estimation ($f_\phi$): A plug-in model uses Cross-Attention to aggregate the interaction sequence with the counterfactual item-time embeddings. It outputs a time-aware propensity score ($\pi_t$), representing the probability that an item was exposed at a specific time.
• Bias Correction: The propensity scores are used to reweight the training loss. The framework optimizes two objectives:
  1. Exposure Loss ($L_{EP}$): Trains the exposure estimation model to distinguish factual exposures from counterfactuals.
  2. Recommendation Loss ($L_{BPR-TIPS}$): A Bayesian Personalized Ranking loss where the weight of each interaction is adjusted by the inverse of the time-aware propensity score ($1/\pi_t$). This weight also includes a time-decay factor to prioritize recent interactions.

3. Key Contributions

1. Time-Aware IPS Framework: The first IPS method designed specifically for sequential recommendation that captures temporal dynamics and sequential dependencies, moving beyond static reweighting.
2. Counterfactual Exposure Estimation: A novel mechanism to estimate item exposure distributions in the absence of logs by constructing time-aware counterfactual samples (similar items, popular items, and time-shifted items).
3. Model Agnostic Plug-in: HyperG can be integrated into both traditional sequential models (e.g., Attention-based, GRU) and generative models (e.g., Diffusion, CVAE) to debias them.
4. Dual Embedding Architecture: Separating interaction and exposure embeddings prevents the amplification of bias during the propensity estimation process.

4. Experimental Results

The authors evaluated HyperG on four datasets: ML-1M, ML-10M, Music4All, and GoodReads, comparing it against state-of-the-art baselines (SASRec, TiSASRec, DiffuRec, CVAE, etc.).

• Performance Gains: HyperG consistently outperformed baselines across all datasets and backbone models.
  • On Music4All, it achieved up to 8.87% improvement in HR@10 and 8.72% in NDCG@10.
  • On ML-10M, it showed significant gains, particularly for generative models (e.g., +5.44% HR@10 for DiffuRec).
• Ablation Studies:
  • Removing the time component ($\text{HyperG}_{\neg \text{time}}$) caused performance drops, confirming the necessity of temporal modeling.
  • Removing IPS ($\text{HyperG}_{\neg \text{IPS}}$) led to significant degradation, proving the effectiveness of the reweighting mechanism.
  • Removing both exposure estimation and time ($\text{HyperG}_{\neg \text{EP\&time}}$) resulted in the largest performance drop, validating the holistic design.
• Hyperparameter Sensitivity: The study found that moderate time decay ($\mu \approx 0.5$) and a balanced exposure loss weight ($\gamma \approx 0.3$) yield optimal results.
• Propensity Discriminability: Analysis showed that HyperG produces more discriminative propensity scores between positive and negative items compared to traditional static IPS, leading to better bias mitigation.

5. Significance

This paper addresses a fundamental gap in sequential recommendation: the inability to distinguish between "not exposed" and "not interested" without explicit exposure logs.

• Theoretical Impact: It advances causal inference in SR by integrating counterfactual reasoning with temporal dynamics, offering a more rigorous approach to debiasing than static methods.
• Practical Impact: As a plug-in framework, it allows existing recommendation systems (which often lack exposure logs) to be easily upgraded to mitigate selection and exposure biases, leading to more accurate and fair recommendations.
• Scalability: The method demonstrates superior performance on large-scale, dynamic datasets (like Music4All), suggesting it is well-suited for real-world applications where user interests and item popularity evolve rapidly.