Source-linked AI summary
Learning Disentangled Representations for Recommendation
Jianxin Ma, Chang Zhou, Peng Cui, Hongxia Yang, Wenwu Zhu
TL;DR
Disentangled representation learning from user behavior is challenging because intentions and granular preferences are entangled, while existing efforts largely neglect this setting. MacridVAE separates macro concepts and preferences and regularizes micro dimensions for independence, achieving substantial improvement over state-of-the-art baselines with interpretable and controllable representations.
Problem
User behavior contains entangled latent factors spanning high-level intentions and low-level preferences, but learning disentangled representations from such data remains challenging and largely neglected.
Method
MacridVAE uses prototypes to disentangle intentions and concept-specific preferences, then applies a strengthened VAE-derived regularizer to separate micro factors.
Results
MacridVAE achieves substantial improvement over state-of-the-art baselines, with improvements especially pronounced on small, sparse datasets.
Takeaways & Limitations
The learned representations are interpretable and controllable, supporting recommendation in which users can adjust targeted aspects of recommendation lists.
Takeaways & Limitations
Not all learned dimensions are human-understandable, and reliably identifying interpretable models requires external knowledge such as item attributes.
Abstract
from arXiv · showhide
User behavior data in recommender systems are driven by the complex interactions of many latent factors behind the users' decision making processes. The factors are highly entangled, and may range from high-level ones that govern user intentions, to low-level ones that characterize a user's preference when executing an intention. Learning representations that uncover and disentangle these latent factors can bring enhanced robustness, interpretability, and controllability. However, learning such disentangled representations from user behavior is challenging, and remains largely neglected by the existing literature. In this paper, we present the MACRo-mIcro Disentangled Variational Auto-Encoder (MacridVAE) for learning disentangled representations from user behavior. Our approach achieves macro disentanglement by inferring the high-level concepts associated with user intentions (e.g., to buy a shirt or a cellphone), while capturing the preference of a user regarding the different concepts separately. A micro-disentanglement regularizer, stemming from an information-theoretic interpretation of VAEs, then forces each dimension of the representations to independently reflect an isolated low-level factor (e.g., the size or the color of a shirt). Empirical results show that our approach can achieve substantial improvement over the state-of-the-art baselines. We further demonstrate that the learned representations are interpretable and controllable, which can potentially lead to a new paradigm for recommendation where users are given fine-grained control over targeted aspects of the recommendation lists.
1 Introduction
User-behavior representations must account for entangled latent factors spanning high-level intentions and low-level preferences. MacridVAE addresses this gap by disentangling both levels using prototypes and a micro-disentanglement regularizer.
- User behavior reflects entangled macro factors governing session intentions and micro factors describing granular preferences.
- Disentangled representations can improve robustness and interpretability, while controllability may support interactive recommendation experiences.
- Learning disentangled representations from discrete, sparse user behavior is challenging because macro and micro factors must remain hierarchically related.
- MacridVAE separates macro and micro factors by identifying intention-related concepts, learning preferences across concepts, and regularizing each dimension toward an independent micro factor.
2 Method
MacridVAE learns user representations that separate high-level concept preferences from low-level preference factors, using a VAE-based objective and prototype assignments. It also supports controllable recommendation by varying selected representation dimensions while preserving others.
- Macro and micro disentanglement: MacridVAE explicitly separates macro factors governing user intentions from micro factors describing granular preferences in user behavior.Macro components represent preferences for distinct high-level concepts, while micro dimensions are encouraged to capture isolated factors.
- Macro disentanglement: The model jointly infers user representations and one-hot item concept assignments, with each item assigned to exactly one high-level concept.The kth user-representation component captures preference regarding concept k.
- Generative model: The decoder predicts item interactions from the full user representation and concept assignments, while the kth component alone is sufficient for items from concept k.This design encourages each representation component to capture preference for only its corresponding concept when assignments are meaningful.
- Micro disentanglement: Micro disentanglement is encouraged by strengthening KL-derived regularization terms that promote dimension independence and suppress input noise.The KL decomposition separates mutual information from a divergence term that encourages independent representation dimensions under a factorized prior.
- Implementation: Prototype-based concept assignment uses cosine similarity between item representations and concept prototypes, helping prevent mode collapse.Inner-product assignments concentrate most items in one prototype, whereas cosine-based assignments distribute items across prototypes; mode collapse still requires careful prior design.
- User-controllable recommendation: Controllable recommendation varies one representation aspect over a suitable range, retrieves items across subranges, and preserves similarity in the remaining dimensions.The range keeps the nearest concept prototype fixed, while beam search approximately optimizes similarity outside the controlled dimension.
3 Empirical Results
Experiments evaluate MacridVAE on five collaborative-filtering datasets under controlled parameter budgets. The results show stronger recommendation performance, interpretable macro structure, and improved disentanglement, while cosine similarity avoids mode collapse.
- Experimental Setup: Experiments use Netflix Prize, three MovieLens datasets, and AliShop-7C, with ratings binarized and users retaining at least five watched movies.
- Experimental Setup: All methods are compared under around 2Md learnable parameters, with d = 100 for item representations.
- Recommendation Performance: MacridVAE significantly outperforms MultDAE and β-MultVAE, especially on small, sparse datasets.The authors associate this with macro disentanglement, information sharing within categories, and independent representation dimensions.
- Macro Disentanglement: Unsupervised item clusters align with ground-truth categories, while user-representation components lie near corresponding cluster centers and capture separate category preferences.
- Macro Disentanglement: Using inner product assigns most items to one prototype, whereas cosine similarity assigns significant numbers of items to all seven prototypes.The comparison supports cosine similarity as a means of preventing mode collapse.
- Micro Disentanglement: Higher representation independence generally accompanies higher recommendation performance, and K = 7 improves independence over K = 1.The approach achieves higher independence than β-MultVAE, suggesting macro disentanglement can help micro disentanglement.
- Interpretability: Representative dimensions have human-understandable semantics and may support fine-grained control, but not all dimensions are human-understandable.The authors recommend future semi-supervised or supervised efforts using external knowledge to identify interpretable models reliably.
4 Related Work
Prior work learns user representations and disentangled factors, but recommendation research largely neglects the entanglement of latent factors in user behavior.
- Deep user-behavior representation methods learn informative representations but generally neglect interactions among latent decision-making factors.
- Disentangled representation learning seeks factorized representations that uncover underlying explanatory factors in observed data.
- β-VAE promotes disentanglement by aggressively penalizing the VAE objective’s KL divergence term.
- Later approaches separate information bottleneck and total correlation terms to achieve greater disentanglement.
5 Conclusions
The paper addresses disentangled representation learning from user behavior by disentangling representations at both macro and micro levels. It identifies interpretability and controllability as directions for future applications.
- The proposed approach performs disentanglement from user behavior at both macro and micro levels.
- Future research should explore applications enabled by the interpretability and controllability of the learned representations.
Evidence lower bound (ELBO)
The supplied passage notes a qualification used in the preceding evidence lower bound derivation.
- The derivation uses an unstated qualification in its last line.
- The passage does not specify the equation or the qualification being applied.
- No further ELBO mechanism or result is given in the supplied text.
Information bottleneck (IB) and total correlation (TC)
The supplied passages state that pdata(xu | C) equals pdata(xu) and specify the joint distribution under which Iq(xu; zu) is evaluated.
- pdata(xu | C) = pdata(xu), so conditioning on C does not change the data distribution in this statement.
- Iq(xu; zu) is evaluated under qθ(zu, xu | C) = qθ(zu | xu, C)pdata(xu | C).
- Using pdata(xu | C) = pdata(xu), the same joint distribution is written as qθ(zu | xu, C)pdata(xu).
- The supplied table passage is captioned only as dataset statistics and provides no specific statistics to summarize.
A.2 Experimental Details
Experiments use preprocessed datasets with separate validation and test users, while tuning model and baseline hyperparameters automatically. The number of macro factors is selected without using ground-truth values, and an adaptive alternative was not significantly better than tuning K.
- Held-out users are split evenly between validation and testing after preprocessing with the β-MultVAE script.
- Hyperparameters are tuned automatically with Hyperopt’s TPE method using 200 trials per method and dataset.
- K is tuned rather than fixed to the ground-truth number of macro factors for fair comparison with baselines.
- The search varies σ0, β, K, learning rate, L2 regularization, dropout, network depth, and hidden-layer width.The specified ranges include K ∈ {1, 2, 3, . . . , 20} and β ∈ [0, 100].
- An adaptive strategy shrinks K when macro-factor distributions have negligible Jensen–Shannon divergence, but it was not significantly better than naïve hyperparameter tuning.The adaptive strategy also adds computational cost and a new hyperparameter.
A.3 Implementation Details
The training procedure clusters items into concepts, encodes user behavior into latent representations, decodes item probabilities, and updates parameters by maximizing the objective. Numerical safeguards and sampled softmax support stable and scalable training.
- The procedure adds 10^-8 where appropriate to prevent division by zero.
- Prototype clustering assigns each item to a concept using temperature-scaled cosine similarity and Gumbel-Softmax.At test time, the concept assignment is set to the mode.
- The encoder forms concatenated means and standard deviations across K components, samples z_u, and returns a KL-divergence term.At test time, z_u is set to its mean.
- The decoder converts latent user representations and item concepts into normalized item probabilities with softmax.Sampled softmax is used for very large item sets, computing probabilities only for clicked or sampled items.
- Each training iteration clusters items, encodes the user, decodes item probabilities, and maximizes the objective using gradient updates.