arXiv subjects

Shivam Verma

Publications and source records attributed to Shivam Verma.

9 recordsLinked to original sources

arXiv2026

Cold-Starting Podcast Ads and Promotions with Multi-Task Learning on Spotify

We present a unified multi-objective model for targeting both advertisements and promotions within the Spotify podcast ecosystem. Our approach addresses key challenges in personalization and cold-start initialization, particularly for new advertising objectives. By leveraging transfer learning from large-scale ad and content interactions within a multi-task learning (MTL) framework, a single joint model can be fine-tuned or directly applied to new or low-data targeting tasks, including in-app promotions. This multi-objective design jointly optimizes podcast outcomes such as streams, clicks, and follows for both ads and promotions using a shared representation over user, content, context, and creative features, effectively supporting diverse business goals while improving user experience. Online A/B tests show up to a 22% reduction in effective Cost-Per-Stream (eCPS), particularly for less-streamed podcasts, and an 18-24% increase in podcast stream rates. Offline experiments and ablations highlight the contribution of ancillary objectives and feature groups to cold-start performance. Our experience shows that a unified modeling strategy improves maintainability, cold-start performance, and coverage, while breaking down historically siloed targeting pipelines. We discuss practical trade-offs of such joint models in a real-world advertising system.

cs.IR↗

arXiv2025

Implications of portal vector-like lepton on associated Higgs production at a multi-TeV muon collider

We have explore a portal vector-like lepton (pVLL) extension of the Standard Model (SM) and study its implications for Higgs and vector-boson associated production ($hV$, with $~V = Z$-boson or dark photon) at a future muon collider facility. We show that while the $~\mu^+ \mu^- \to hZ~$ production rate remains close to its SM prediction in a wide range of parameter space, the rate for $~\mu^+ \mu^- \to h\gamma_d~$ can be substantially enhanced owing to the non-decoupling nature of the interaction involving the heavy lepton, the muon and the dark photon. We demonstrate that the $h\gamma_d$ production rate can exceed the corresponding $hZ$ rate by a factor of $1$-$100$ within the perturbative unitarity limit, making it a promising channel for probing Higgs interactions and potential new physics effects. We also examine the role of the pVLL state in the context of dark matter (DM) phenomenology and identify regions of parameter space consistent with the observed relic abundance by extending the simplified setup with a viable DM candidate. The $h\gamma_d$ production can also be used to constrain the dark photon mass ($m_{\gamma_d}$) and/or the dark gauge coupling ($g_d$) consistent with various constraints including the current muon $g-2$ measurements within the pVLL framework. We perform a detailed collider analysis of the $h\gamma_d$ process in the $b\bar{b}~+$ missing energy final state. A $2\sigma$ exclusion limit for $m_{\gamma_d}$ up to $80$ GeV is obtained assuming $~g_d=0.05$, $~\sin\theta_L=4\times10^{-5}$, $~\sin\theta_s=0.05$, for a heavy lepton mass $~\sim 3$ TeV at a $3$ TeV muon collider with an integrated luminosity of $1$ ab$^{-1}$.

hep-ph↗

arXiv2025

Thermal Tracks: A Gaussian process-based framework for universal melting curve analysis enabling unconstrained hit identification in thermal proteome profiling experiments

Thermal Tracks is a Python-based statistical framework for analyzing protein thermal stability data that overcomes key limitations of existing thermal proteome profiling (TPP) work-flows. Unlike standard approaches that assume sigmoidal melting curves and are constrained by empirical null distributions (limiting significant hits to approximately 5 % of data), Thermal Tracks uses Gaussian Process (GP) models with squared-exponential kernels to flexibly model any melting curve shape while generating unbiased null distributions through kernel priors. This framework is particularly valuable for analyzing proteome-wide perturbations that significantly alter protein thermal stability, such as pathway inhibitions, genetic modifications, or environmental stresses, where conventional TPP methods may miss biologically relevant changes due to their statistical constraints. Furthermore, Thermal Tracks excels at analyzing proteins with un-conventional melting profiles, including phase-separating proteins and membrane proteins, which often exhibit complex, non-sigmoidal thermal stability behaviors. Thermal Tracks is freely available from GitHub and is implemented in Python, providing an accessible and flexible tool for proteome-wide thermal profiling studies.

cs.LG↗

arXiv2025

An Audio-centric Multi-task Learning Framework for Streaming Ads Targeting on Spotify

Spotify, a large-scale multimedia platform, attracts over 675 million monthly active users who collectively consume millions of hours of music, podcasts, audiobooks, and video content. This diverse content consumption pattern introduces unique challenges for computational advertising, which must effectively integrate a variety of ad modalities, including audio, video, and display, within a single user experience. Traditional ad recommendation models, primarily designed for foregrounded experiences, often struggle to reconcile the platform's inherent audio-centrality with the demands of optimizing ad performance across multiple formats and modalities. To overcome these challenges, we introduce Cross-modal Adaptive Mixture-of-Experts (CAMoE), a novel framework for optimizing click-through rate (CTR) prediction in both audio-centric and multi-modal settings. CAMoE enhances traditional mixture-of-experts models by incorporating modality-aware task grouping, adaptive loss masking, and deep-cross networks (DCN) to capture complex feature interactions within a multi-modal ad ecosystem. Through extensive ablation studies, we demonstrate that this approach achieves near Pareto-optimal performance across audio, video, and display ad formats, significantly improving AUC-PR compared to conventional single-task and content-based multi-task learning baselines. When deployed at scale on Spotify's ad serving platform, CAMoE delivered substantial gains, yielding a 14.5% increase in CTR for audio ads, a 1.3% increase for video ads, and a 4.8% reduction in expected cost-per-click (eCPC) for audio slots.

cs.IR↗

arXiv2024

Machine learning tagged boosted dark photon: A signature of fermionic portal matter at the LHC

We use a Hybrid Deep Neural Network (HDNN) to identify a boosted dark photon jet as a signature of a heavy vector-like fermionic portal matter (PM) connecting the visible and the dark sectors. In this work, the fermionic PM, which mixes only with the Standard Model (SM) third-generation up-type quark, predominantly decays into a top quark and a dark photon pair. The dark photon then promptly decays to a pair of standard model fermions via the gauge kinetic mixing. We have analyzed two different final states, namely, (i) exactly one tagged dark photon and exactly one tagged top quark jet, and (ii) at least two tagged dark photons and at least one tagged top quark jet at the 13 and 14 TeV LHC center of mass energies. Both these final states receive significant contributions from the pair and single production processes of the top partner. The rich event topology of the signal processes, i.e., the presence of a boosted dark photon and top quark jet pair, along with the fact that the invariant mass of the system corresponds to the mass of the top partner, help us to significantly suppress potential SM backgrounds. We have shown that one can set a $2\sigma$ exclusion limit of $\sim 2.3$ TeV on the top partner mass with $\sin\theta_L=0.1$ and assuming $100\%$ branching ratio of the top partner in the final state with exactly one tagged dark photon and exactly one tagged top quark jet at the 14 TeV LHC center of mass energy assuming 300 fb$^{-1}$ of integrated luminosity.

hep-ph↗

arXiv2023

Exploring maverick top partner decays at the LHC

In this work, we have considered an extension of the standard model (SM) with a $SU(2)_L$ singlet vectorlike quark (VLQ) with electric charge $Q=+2/3$. The model also contains an additional local $U(1)_d$ symmetry group and the corresponding gauge boson is the dark photon. The VLQ is charged while all the SM particles are neutral under the new $U(1)_d$ gauge group. Even though in this model the VLQ possesses many properties qualitatively similar to that of the traditional top partner ($T_p$), there are some compelling differences as well. In particular, its branching ratio to the traditional modes ($T_p \to bW, tZ, th$) are suppressed which in turn helps to evade many of the existing bound, mainly coming from the LHC experiments. In an earlier work, such a VLQ is referred to as ``maverick top partner". It has been shown that the top partner in this model predominantly decays to a top quark and a dark photon/dark higgs pair ($T_p \to tγ_d ,~th_d$) over a large region of the parameter space. The dark photon can be made invisible and consequently, it gives rise to the missing transverse energy ($\not\!\!{E} _{T}$) signature at the LHC detector. We have mainly focused on the LHC signatures and future prospects of such top partners. In particular, we have studied the $t\bar{t}+\not\!\!{E}_{T}$ and $t+\not\!\!{E}_{T}$ signatures in the context of the LHC via pair and single productions of the top partner, respectively at 13 and 14 TeV LHC center of mass energies assuming that the dark photon either decays into an invisible mode or it is invisible at the length scale of the detector. We have shown that one can exclude $\sinθ_L \sim 0.025$ (0.05) for $m_{_{T_p}} \leq $ 2.0 (2.6) TeV at $\sqrt{s}=14$ TeV with an integrated luminosity of 3 ab$^{-1}$ using the single top partner production channel.

hep-ph↗

arXiv2022

Novel STT/SHE MTJ Compact Model Compatible with NGSPICE

Ensuring high performance, while meeting the power budget is a challenging task as the world is moving towards next-generation computing. Researchers and designers are in search of new solutions for efficient computation. Spintronics devices have been viewed as a promising way to deal with the escalating difficulties of CMOS downscaling, explicitly, the Magnetic Tunnel Junction (MTJ) devices have been the focal point of investigation. They possess some essential features from the aforementioned perspective such as nonvolatility, low power, and scalability. In light of the significance of MTJ devices in next-generation computing, this paper presents a physics-based STT/SHE MTJ model for hybrid MTJ/CMOS circuit simulation, that accurately emulates the device physics and stochastic thermal noise behavior of the MTJ. It is vital to have an MTJ compact model which is compatible with the open-source NGSPICE simulation framework since previously developed models are reliant on commercial EDA tools. In addition, for developing hybrid circuits with random process fluctuations, a simulator-independent Monte-Carlo simulation capability has been incorporated Finally, the STT/SHE-MTJ model is demonstrated using PCSA read/write operation and the implementation of neuron MTJ.

physics.app-ph↗

arXiv2021

Lessons Learned Addressing Dataset Bias in Model-Based Candidate Generation at Twitter

Traditionally, heuristic methods are used to generate candidates for large scale recommender systems. Model-based candidate generation promises multiple potential advantages, primarily that we can explicitly optimize the same objective as the downstream ranking model. However, large scale model-based candidate generation approaches suffer from dataset bias problems caused by the infeasibility of obtaining representative data on very irrelevant candidates. Popular techniques to correct dataset bias, such as inverse propensity scoring, do not work well in the context of candidate generation. We first explore the dynamics of the dataset bias problem and then demonstrate how to use random sampling techniques to mitigate it. Finally, in a novel application of fine-tuning, we show performance gains when applying our candidate generation system to Twitter's home timeline.

cs.IR↗

arXiv2020

SIMPler realisation of Scalar Dark Matter

With growing agony of not finding a dark matter (DM) particle in direct search experiments so far (for example in XENON1T), frameworks where the freeze-out of DM is driven by number changing processes within the dark sector itself and do not contribute to direct search, like Strongly Interacting Massive Particle (SIMP) are gaining more attention. In this analysis, we ideate a simple scalar DM framework stabilised by $Z_3$ symmetry to serve with a SIMP-like DM ($χ$) with additional light scalar mediation ($ϕ$) to enhance DM self interaction. We identify that a large parameter space for such DM is available from correct relic density and self interaction constraints coming from Bullet or Abell cluster data. We derive an approximate analytic solution for freeze-out of the SIMP like DM in Boltzmann Equation describing $3 \to 2$ number changing process within the dark sector. We also provide a comparative analysis of the SIMP like solution with the Weakly Interacting Massive Particle (WIMP) realisation of the same model framework here.

hep-ph↗