Exponential-Gaussian Mixture Network (EGMN)

• EGMN is a probabilistic framework that models watch time through a conditional mixture of one exponential and multiple Gaussian components.
• It uses a shared encoder with specialized branches to capture both quick-skip behavior and longer, multimodal viewing patterns.
• Empirical evaluations on industrial datasets show improved prediction accuracy with significant gains in MAE and XAUC metrics.

Exponential-Gaussian Mixture Network (EGMN) is a probabilistic watch-time prediction framework that models watch time as a conditional Exponential-Gaussian Mixture (EGM) distribution. In the formulation introduced for short-video recommendation, the model is motivated by two distributional regularities observed in industrial data: a coarse-grained skewness induced by a significant concentration of quick-skips, and a fine-grained diversity arising from heterogeneous user-video interaction patterns. EGMN addresses these by combining one exponential component, intended to capture the near-zero spike, with multiple Gaussian components, intended to represent bimodal, trimodal, or more generally multimodal watch-time behavior; a neural network then predicts the mixture parameters end-to-end from user, video, and context features (Zhao et al., 18 Aug 2025).

1. Definition and problem setting

EGMN was introduced for video watch time prediction in short-video recommendation, where direct regression with a single simple output distribution is argued to be inadequate. The motivating claim is that real watch-time data simultaneously exhibits two distinct statistical characteristics. First, there is coarse-grained skewness because many users quick-skip videos, creating a sharp concentration of probability mass near zero. Second, there is fine-grained diversity because different users, different videos, and different user-video pairs produce varied engagement patterns, including multimodal shapes (Zhao et al., 18 Aug 2025).

Within this formulation, the exponential component is assigned the role of modeling the dense near-zero region associated with quick-skips, while the Gaussian mixture component is assigned the role of modeling more structured, heterogeneous, and possibly multimodal viewing behavior. The design is explicitly contrasted with approaches based on label normalization, classification-bin transformation, and some debiasing models. Label normalization can make the distribution easier to learn but may lose absolute watch-time information; classification reformulations simplify learning but introduce discretization and reconstruction error; and some debiasing approaches may miss either quick-skips or fine-grained multimodality (Zhao et al., 18 Aug 2025).

A central architectural point is that EGMN is not primarily presented as a fixed backbone model. Rather, it is a distributional prediction framework in which a shared encoder produces a hidden representation and a parameter generator maps that representation to the parameters of the conditional EGM distribution. This makes EGMN functionally a probabilistic head or framework that can sit on top of standard recommendation backbones (Zhao et al., 18 Aug 2025).

2. Exponential-Gaussian Mixture distribution

The EGMN paper assumes that watch time $t \in \mathbb{R}^+$ follows a mixture of one exponential component and $K$ Gaussian components. The conditional density is written as

$$p(t \mid \mathbf{x}) = \omega_0(\mathbf{x}) f_{\text{exp}}(t \mid \lambda(\mathbf{x})) + \sum_{k=1}^{K} \omega_k(\mathbf{x}) f_{\text{gauss}}(t \mid \mu_k(\mathbf{x}), \sigma_k^2(\mathbf{x})).$$

The component densities are the standard exponential and Gaussian pdfs:

$$f_{\text{exp}}(t \mid \lambda) = \lambda e^{-\lambda t},$$

and

$$f_{\text{gauss}}(t \mid \mu, \sigma^2) = \frac{1}{\sqrt{2\pi\sigma^2}} e^{-\frac{(t-\mu)^2}{2\sigma^2}}.$$

The mixture weights satisfy

$\sum_{k=0}^{K} \omega_k = 1,$

with $\lambda > 0$, $\sigma_k^2 > 0$, and $\omega_k \ge 0$ (Zhao et al., 18 Aug 2025).

The paper interprets these terms asymmetrically. The exponential term $\omega_0 f_{\text{exp}}$ is intended to model the near-zero spike caused by quick-skips. The Gaussian sum $K$0 is intended to model longer or more structured watch behaviors, including multimodal patterns. This division of labor is the basis for the claim that EGMN performs coarse-to-fine distribution modeling: coarse scale via the exponential component, fine scale via the Gaussian mixture (Zhao et al., 18 Aug 2025).

The parameterization includes an identifiability-oriented constraint on the Gaussian means:

$K$1

This ensures that the Gaussian means are larger than the exponential component’s characteristic scale. In the paper’s interpretation, this helps reserve the exponential component for the short-time quick-skip region while allowing the Gaussian components to model longer or more complex viewing patterns (Zhao et al., 18 Aug 2025).

This distributional construction should be distinguished from two adjacent but non-identical uses of related terminology. In portfolio optimization, “EGM” refers to exponential utility with Gaussian mixture returns rather than a watch-time prediction network (Luxenberg et al., 2022). In Gaussian Mixture Convolution Networks, Gaussian mixtures are used as the representational substrate for continuous convolution, but the paper explicitly does not define that model as an Exponential-Gaussian Mixture Network (Celarek et al., 2022).

3. Network parameterization

EGMN predicts the parameters of the conditional EGM distribution from user-item-context features. For each user-video pair, the input feature vector $K$2 is built from user features, video features, and context features. The user features include demographics, historical behavior, and engagement statistics; the video features include content attributes, duration, category, and creator information; and the context features include time, weekday, and device type. These features are embedded and concatenated into $K$3 (Zhao et al., 18 Aug 2025).

The first stage is a shared backbone encoder,

$K$4

which produces a hidden representation shared by all mixture components. The paper emphasizes that this backbone is plug-and-play and architecture-agnostic, with examples including DCN, DIN, SENet, and Transformer. EGMN is therefore described more accurately as a parameterization framework for the EGM distribution than as a single fixed neural architecture (Zhao et al., 18 Aug 2025).

From the shared hidden vector $K$5, separate branches generate the mixture parameters. The exponential rate is predicted as

$K$6

For each Gaussian component $K$7, the mean and variance are predicted as

$K$8

and

$K$9

The mixture weights are generated by

$$p(t \mid \mathbf{x}) = \omega_0(\mathbf{x}) f_{\text{exp}}(t \mid \lambda(\mathbf{x})) + \sum_{k=1}^{K} \omega_k(\mathbf{x}) f_{\text{gauss}}(t \mid \mu_k(\mathbf{x}), \sigma_k^2(\mathbf{x})).$$0

Softplus is used for $$p(t \mid \mathbf{x}) = \omega_0(\mathbf{x}) f_{\text{exp}}(t \mid \lambda(\mathbf{x})) + \sum_{k=1}^{K} \omega_k(\mathbf{x}) f_{\text{gauss}}(t \mid \mu_k(\mathbf{x}), \sigma_k^2(\mathbf{x})).$$1, $$p(t \mid \mathbf{x}) = \omega_0(\mathbf{x}) f_{\text{exp}}(t \mid \lambda(\mathbf{x})) + \sum_{k=1}^{K} \omega_k(\mathbf{x}) f_{\text{gauss}}(t \mid \mu_k(\mathbf{x}), \sigma_k^2(\mathbf{x})).$$2, and $$p(t \mid \mathbf{x}) = \omega_0(\mathbf{x}) f_{\text{exp}}(t \mid \lambda(\mathbf{x})) + \sum_{k=1}^{K} \omega_k(\mathbf{x}) f_{\text{gauss}}(t \mid \mu_k(\mathbf{x}), \sigma_k^2(\mathbf{x})).$$3 to keep them positive where needed, and softmax is used for $$p(t \mid \mathbf{x}) = \omega_0(\mathbf{x}) f_{\text{exp}}(t \mid \lambda(\mathbf{x})) + \sum_{k=1}^{K} \omega_k(\mathbf{x}) f_{\text{gauss}}(t \mid \mu_k(\mathbf{x}), \sigma_k^2(\mathbf{x})).$$4 so that the weights sum to one (Zhao et al., 18 Aug 2025).

The paper’s architectural intuition is that the shared encoder learns a common representation of the user-video pair, after which specialized branches take on distinct distributional roles: one branch predicts how much quick-skip mass is present, and the others predict the shapes of more nuanced engagement modes. This suggests that EGMN is best understood as a conditional mixture-density model tailored to a two-regime watch-time distribution, rather than as a generic mixture-density network with interchangeable components (Zhao et al., 18 Aug 2025).

4. Training objective and estimation

EGMN is trained with a weighted combination of three losses. The main term is the negative log-likelihood,

$$p(t \mid \mathbf{x}) = \omega_0(\mathbf{x}) f_{\text{exp}}(t \mid \lambda(\mathbf{x})) + \sum_{k=1}^{K} \omega_k(\mathbf{x}) f_{\text{gauss}}(t \mid \mu_k(\mathbf{x}), \sigma_k^2(\mathbf{x})).$$5

which directly fits the conditional mixture distribution to observed watch times. The expanded form substitutes the exponential and Gaussian mixture terms into $$p(t \mid \mathbf{x}) = \omega_0(\mathbf{x}) f_{\text{exp}}(t \mid \lambda(\mathbf{x})) + \sum_{k=1}^{K} \omega_k(\mathbf{x}) f_{\text{gauss}}(t \mid \mu_k(\mathbf{x}), \sigma_k^2(\mathbf{x})).$$6 (Zhao et al., 18 Aug 2025).

To discourage collapse onto too few components, the model adds an entropy-based regularizer,

$$p(t \mid \mathbf{x}) = \omega_0(\mathbf{x}) f_{\text{exp}}(t \mid \lambda(\mathbf{x})) + \sum_{k=1}^{K} \omega_k(\mathbf{x}) f_{\text{gauss}}(t \mid \mu_k(\mathbf{x}), \sigma_k^2(\mathbf{x})).$$7

Because this term is minimized, the effect is to maximize the entropy of the mixture weights. The stated purpose is to encourage the use of multiple components when appropriate (Zhao et al., 18 Aug 2025).

A third term is a regression loss based on the expected value of the mixture. The point prediction is

$$p(t \mid \mathbf{x}) = \omega_0(\mathbf{x}) f_{\text{exp}}(t \mid \lambda(\mathbf{x})) + \sum_{k=1}^{K} \omega_k(\mathbf{x}) f_{\text{gauss}}(t \mid \mu_k(\mathbf{x}), \sigma_k^2(\mathbf{x})).$$8

and the regression loss is

$$p(t \mid \mathbf{x}) = \omega_0(\mathbf{x}) f_{\text{exp}}(t \mid \lambda(\mathbf{x})) + \sum_{k=1}^{K} \omega_k(\mathbf{x}) f_{\text{gauss}}(t \mid \mu_k(\mathbf{x}), \sigma_k^2(\mathbf{x})).$$9

The final objective is

$$f_{\text{exp}}(t \mid \lambda) = \lambda e^{-\lambda t},$$0

with the paper reporting typical values $$f_{\text{exp}}(t \mid \lambda) = \lambda e^{-\lambda t},$$1 and $$f_{\text{exp}}(t \mid \lambda) = \lambda e^{-\lambda t},$$2 (Zhao et al., 18 Aug 2025).

The ablation results reported for this loss construction assign distinct roles to each term. Removing maximum-likelihood training weakens distribution modeling most strongly; removing entropy regularization causes the model to collapse onto too few components; and removing the regression loss reduces absolute prediction accuracy. The same section reports that the number of Gaussian components matters materially: too few underfit the complexity, too many lead to optimization difficulty and overfitting, and about 8–12 components gives the best balance (Zhao et al., 18 Aug 2025).

5. Empirical evaluation

The reported evaluation covers four datasets: Indust, KuaiRec, WeChat, and CIKM. Indust is described as a real industrial short-video dataset containing 14 days of logs, 951,870 interactions, 25,947 users, and 7,155 videos. CIKM is described as a session-based dataset used to test scalability beyond short video. The baselines are VR, TPM, D2Q, CREAD, and D²CO, and the metrics are MAE for regression accuracy and XAUC for ranking consistency (Zhao et al., 18 Aug 2025).

Category	Items
Datasets	Indust, KuaiRec, WeChat, CIKM
Baselines	VR, TPM, D2Q, CREAD, D²CO
Metrics	MAE, XAUC

The paper reports that EGMN is best on all four datasets. The average offline improvements are about 14.11% in MAE and 7.76% in XAUC. Against the strongest baseline, the gains are reported as about 6.75% MAE and 5.09% XAUC on Indust, about 5.04% MAE and 4.67% XAUC on KuaiRec, about 1.72% MAE and 1.39% XAUC on WeChat, and about 2.42% MAE and 0.49% XAUC on CIKM (Zhao et al., 18 Aug 2025).

An online A/B test is also reported for Xiaohongshu’s industrial short-video feed. The deployment lasts 7 days, with about 10% of traffic assigned to EGMN against the baseline CREAD. The reported gains are Watch Time: +0.681%, Video Views: +0.189%, and Engage Actions: essentially unchanged / slightly negative but negligible. The paper also states that online prediction accuracy improves, with MAE decreasing, XAUC increasing, and KL divergence between predicted and real watch-time distributions dropping by nearly 20% (Zhao et al., 18 Aug 2025).

The ablation and visualization studies are directed specifically at the coarse-to-fine distributional claims. Removing the exponential component causes clear performance drops, with AUC degrading by more than 6% across quick-skip thresholds, alongside worse MAE and XAUC. Removing Gaussian components also worsens MAE and XAUC. The visual analyses further report that EGMN fits the overall distribution better than most baselines, reproduces bimodal duration-level watch-time distributions, and matches user-video-level shapes such as large quick-skip mass for “picky” users and multi-peak structures for some videos (Zhao et al., 18 Aug 2025).

6. Relation to adjacent Gaussian-mixture research and terminology

The term Exponential-Gaussian Mixture Network is used explicitly in the watch-time prediction paper, where “exponential” refers to an exponential distribution component and “Gaussian mixture” refers to the Gaussian components of the conditional target distribution (Zhao et al., 18 Aug 2025). This use should be distinguished from a different but related use of “EGM” in quantitative finance. In “Portfolio Construction with Gaussian Mixture Returns and Exponential Utility via Convex Optimization,” the combination is between a Gaussian mixture model for returns and an exponential utility objective,

$$f_{\text{exp}}(t \mid \lambda) = \lambda e^{-\lambda t},$$3

leading to an exact convex optimization problem. That paper does not introduce a neural network called EGMN, but it does provide a mathematically precise “GM + exponential” construction in another domain (Luxenberg et al., 2022).

A second neighboring line of work is Gaussian Mixture Convolution Networks (GMCNs). GMCNs are described as a continuous, function-based analogue of a CNN in which both signals and kernels are represented as multidimensional Gaussian mixtures and convolution is performed analytically in closed form. The paper explicitly states, however, that it does not introduce an “Exponential-Gaussian Mixture Network” terminology. Its relevance is structural rather than terminological: it is “EGMN-like in spirit” because it propagates Gaussian-mixture parameterizations through a layered architecture, but it is not an EGMN in the explicit sense used in the watch-time paper (Celarek et al., 2022).

These distinctions matter because the phrase “Exponential-Gaussian Mixture” can denote different mathematical compositions in different literatures. In the watch-time model, the exponential and Gaussian terms are components of a conditional predictive density over positive durations. In the portfolio paper, the exponential term belongs to the utility function while the Gaussian mixture describes the return law. In GMCNs, Gaussian mixtures define the representational substrate for analytic convolution, and no exponential-family or exponential-component distribution of the EGMN type is introduced (Zhao et al., 18 Aug 2025).

A plausible implication is that EGMN occupies a specific position within a broader family of Gaussian-mixture-based methods: it is not merely a generic mixture-density network, and it is not interchangeable with Gaussian-mixture functional networks or exponential-utility Gaussian-mixture optimization. Its defining feature is the explicit decomposition of watch-time uncertainty into a near-zero exponential regime and a multi-modal Gaussian regime, coupled to a neural parameter generator trained by likelihood, entropy regularization, and expectation-based regression (Zhao et al., 18 Aug 2025).

7. Significance, limitations, and interpretive issues

The principal significance claimed for EGMN is its ability to fit watch-time distributions across multiple granularities while preserving absolute watch-time values. The paper argues that this is preferable to normalization- or discretization-based strategies when the goal is to model the full watch-time distribution directly. The combination of an exponential quick-skip component and Gaussian multimodal components is presented as the mechanism by which the model captures overall skewness, duration-level bimodality, user-level differences, video-level multimodality, and joint user-video interaction effects (Zhao et al., 18 Aug 2025).

The main empirical support for this claim comes from three sources: best offline results on all four reported datasets, positive online A/B test outcomes in an industrial short-video feed, and ablation evidence showing that both the exponential term and the Gaussian components are functionally necessary. The exponential component is validated by the degradation in quick-skip recognition and predictive metrics when it is removed, while the Gaussian mixture is validated by worse MAE and XAUC when it is excluded (Zhao et al., 18 Aug 2025).

At the same time, the reported design implies several constraints. The modeling assumption is that watch time is well described by one exponential component plus $$f_{\text{exp}}(t \mid \lambda) = \lambda e^{-\lambda t},$$4 Gaussian components. The paper’s own component-count study indicates sensitivity to the choice of $$f_{\text{exp}}(t \mid \lambda) = \lambda e^{-\lambda t},$$5: too few Gaussian components underfit, while too many create optimization difficulty and overfitting, with about 8–12 components giving the best balance. This suggests that the model’s effectiveness depends materially on matching component complexity to the data regime (Zhao et al., 18 Aug 2025).

Another interpretive issue concerns terminology. Because “EGM” already has an established meaning in the portfolio paper and because Gaussian-mixture neural architectures exist in other forms such as GMCNs, the term “EGMN” should be used with domain specificity. In current arXiv usage represented here, EGMN refers explicitly to the short-video watch-time prediction framework built around a conditional Exponential-Gaussian Mixture distribution, not to Gaussian-mixture convolutional networks and not to exponential-utility portfolio construction under Gaussian-mixture returns (Zhao et al., 18 Aug 2025).