TL;DR
Standard sequential recommendation models like SASRec and BERT4Rec often fail in the music domain because they ignore a fundamental human habit: we love to listen to the same songs over and over again. By explicitly injecting a Personalized Popularity Score (PPS) into these Transformer-based architectures, researchers achieved performance boosts of up to 70%, finally allowing these complex models to outperform simple "play my favorites" baselines.
What Problem Does This Paper Address?
In domains like e-commerce, sequential recommendation makes perfect sense—if you just bought a smartphone, you probably want a case next, not another smartphone. However, music is different. Music consumption is characterized by high repetitive consumption. Users frequently return to their favorite tracks, creating a "personalized popularity" signal that standard models struggle to capture.
Current state-of-the-art (SOTA) models like SASRec and BERT4Rec are essentially adapted from Natural Language Processing (NLP). They treat items like words in a sentence, trying to predict the "next word." But while you rarely say the same word ten times in a row in a coherent sentence, you might very well listen to the same song ten times in a day. The authors identify a "disconnect" where complex models focus too much on discovering new items and fail to prioritize the tracks users actually want to hear again.
Methods
The authors propose a novel way to bake personalized popularity directly into the scoring mechanism of sequential models.
- Personalized Popularity Score (PPS): For every user, the system maintains a count of how many times they have interacted with each item in the catalog.
- Mathematical Integration: Instead of just adding a feature, the authors derived specific ways to inject these counts into the model's final output layer based on its activation function:
- For Softmax (e.g., BERT4Rec): They transform the popularity counts into a log-space score that aligns with the model's internal logits.
- For Sigmoid (e.g., SASRec/gSASRec): They use the additive inverse of the exponent to ensure the popularity signal correctly biases the probability.
- Smoothing: They introduce a smoothing parameter ($\epsilon$) to control the influence of the popularity signal, preventing it from completely overwhelming the model's ability to recommend new music.
Essentially, this forces the neural network to learn the delta—the difference between what a user likes based on habit and what they might like based on recent sequential context.
Datasets
The researchers evaluated their approach on two well-known music datasets, summarized in the table below:
- Yandex Music Event Dataset: A large-scale dataset from 2019 containing likes, dislikes, plays, and skips.
- Last.fm-1K: A classic dataset containing nearly a decade of listening history for 1,000 users.
To handle the massive scale of these catalogs, they employed a popularity-based item sampling method to select a representative subset of 30,000 items for training.
Key Insights and Ideas
- The "Simple" Baseline is a Giant: One of the most surprising findings was that a "Personalized Most Popular" (PMP) recommender—which simply recommends a user's most-played songs—outperforms standard SOTA Transformer models in music recommendation. This highlights how much these complex models were missing the mark on basic user behavior.
- Learning the Delta: By injecting the PPS, the Transformer models no longer have to "waste" capacity learning that a user likes their favorite songs. Instead, they can focus on the nuances of when to play those favorites versus when to introduce something new.
- Domain Specificity Matters: The paper serves as a reminder that "plug-and-play" SOTA models from NLP or e-commerce often require fundamental architectural tweaks to work in specialized domains like music.
Key Results
The results, shown in the table below, demonstrate a massive leap in performance across the board.
- Massive Gains: Adding PPS to BERT4Rec, SASRec, and gSASRec resulted in relative improvements in NDCG (Normalized Discounted Cumulative Gain) ranging from 25.2% to 69.8%.
- Beating the Baseline: While the simple PMP baseline was hard to beat at low ranking cutoffs (like NDCG@5), the PPS-enhanced models eventually surpassed it at higher cutoffs (NDCG@100). This suggests the hybrid approach successfully balances "playing the hits" with "discovering the new."
- Statistical Significance: The improvements were not just large but statistically significant ($p < 0.05$), proving the robustness of the popularity injection.
Limitations and Future Work
While effective, the authors acknowledge a few areas for further exploration:
- Filter Bubbles: There is a risk that over-prioritizing personalized popularity could trap users in a "filter bubble," limiting their exposure to new artists and genres.
- Beyond Score Integration: The current method integrates popularity at the final scoring stage. Future work could look at injecting this signal earlier in the self-attention layers.
- Cross-Domain Application: While this paper focuses on music, the authors suggest this approach could be valuable in other repetitive domains, such as grocery shopping or frequent "Point of Interest" (POI) visits.
Summary updated