SearcharxivSearch

arXiv subjects

Francesca Catino

Publications and source records attributed to Francesca Catino.

10 recordsLinked to original sources

VCG: A Multimodal Retrieval Framework for E-Commerce Video Feeds under Extreme Cold-Start Conditions

The digital commerce landscape is shifting from static, search-driven catalogs to dynamic, immersive video feeds. This transition introduces an ``extreme cold-start'' problem: unlike traditional items, new short-form videos lack the dense interaction history required for collaborative filtering. Furthermore, immersive feeds introduce strong position and duration biases that distort standard engagement signals. In this paper, we demonstrate the Video Candidate Generation (VCG) system, a scalable multimodal retrieval engine designed to solve these challenges in a large-scale e-commerce environment. By leveraging a domain-adapted vision-language model (based on CLIP), we map users and videos into a shared semantic space, enabling zero-shot retrieval based on visual content rather than behavioral history. We detail the system's architecture and present a rigorous evaluation comparing generative (LLM) vs. discriminative (CLIP) embeddings. Our results show that while generative models excel at attribute prediction, they suffer from embedding space collapse in retrieval tasks. Online A/B testing demonstrates that VCG effectively mitigates engagement biases, yielding a 50\% uplift in deep video completion. To showcase the system's capabilities, we present an interactive demonstration featuring three bi-directional retrieval scenarios: Product-to-Video, Video-to-Product, and Zero-Shot Semantic Search.

cs.IR

Short-Form Video Recommendations with Multimodal Embeddings: Addressing Cold-Start and Bias Challenges

In recent years, social media users have spent significant amounts of time on short-form video platforms. As a result, established platforms in other domains, such as e-commerce, have begun introducing short-form video content to engage users and increase their time spent on the platform. The success of these experiences is due not only to the content itself but also to a unique UI innovation: instead of offering users a list of choices to click, platforms actively recommend content for users to watch one at a time. This creates new challenges for recommender systems, especially when launching a new video experience. Beyond the limited interaction data, immersive feed experiences introduce stronger position bias due to the UI and duration bias when optimizing for watch-time, as models tend to favor shorter videos. These issues, together with the feedback loop inherent in recommender systems, make it difficult to build effective solutions. In this paper, we highlight the challenges faced when introducing a new short-form video experience and present our experience showing that, even with sufficient video interaction data, it can be more beneficial to leverage a video retrieval system using a fine-tuned multimodal vision-language model to overcome these challenges. This approach demonstrated greater effectiveness compared to conventional supervised learning methods in online experiments conducted on our e-commerce platform.

cs.LG

Scalar masses in general N=2 gauged supergravity theories

We readdress the question of whether any universal upper bound exists on the square mass m^2 of the lightest scalar around a supersymmetry breaking vacuum in generic N=2 gauged supergravity theories for a given gravitino mass m_3/2 and cosmological constant V. We review the known bounds which apply to theories with restricted matter content from a new perspective. We then extend these results to theories with both hyper and vector multiplets and a gauging involving only one generator, for which we show that such a bound exists for both V>0 and V<0. We finally argue that there is no bound for the same theories with a gauging involving two or more generators. These results imply that in N=2 supergravity theories metastable de Sitter vacua with V< >m^2_3/2 can also arise when a single isometry is gauged.

hep-th

On the moduli space of spontaneously broken N = 8 supergravity

We analyze the moduli space of spontaneously broken N=8 supergravity theories in 4 dimensions with classical Minkowski vacua. We find that all the known classical vacua, as well as the several new ones we construct here, can be connected by sending some of the moduli to their boundary values. We also show that Cremmer-Scherk-Schwarz models can be viewed as special limits of more general CSO* gaugings, which allow for non-Abelian residual symmetries on the vacuum. Finally, we find that all the classical Minkowski vacua with fully broken supersymmetry found so far are unstable with respect to 1-loop corrections, which drive the effective potential to negative values.

hep-th

Simple metastable de Sitter vacua in N=2 gauged supergravity

We construct a simple class of N=2 gauged supergravity theories that admit metastable de Sitter vacua, generalizing the recent work done in the context of rigid supersymmetry. The setup involves one hypermultiplet and one vector multiplet spanning suitably curved quaternionic-Kahler and special-Kahler geometries, with an Abelian gauging based on a single triholomorphic isometry, but neither Fayet-Iliopoulos terms nor non-Abelian gauge symmetries. We construct the most general model of this type and show that in such a situation the possibility of achieving metastable supersymmetry breaking vacua crucially depends on the value of the cosmological constant V relative to the gravitino mass squared m_{3/2}^2 in Planck units. In particular, focusing on de Sitter vacua with positive V, we show that metastability is only possible when V >= 2.17 m_{3/2}^2. We also derive an upper bound on the lightest scalar mass in this kind of model relative to the gravitino mass m_{3/2} as a function of the cosmological constant V, and discuss its physical implications.

hep-th

Metastable de Sitter vacua in N=2 to N=1 truncated supergravity

We study the possibility of achieving metastable de Sitter vacua in general N=2 to N=1 truncated supergravities without vector multiplets, and compare with the situations arising in N=2 theories with only hypermultiplets and N=1 theories with only chiral multiplets. In N=2 theories based on a quaternionic manifold and a graviphoton gauging, de Sitter vacua are necessarily unstable, as a result of the peculiar properties of the geometry. In N=1 theories based on a Kahler manifold and a superpotential, de Sitter vacua can instead be metastable provided the geometry satisfies some constraint and the superpotential can be freely adjusted. In N=2 to N=1 truncations, the crucial requirement is then that the tachyon of the mother theory be projected out from the daughter theory, so that the original unstable vacuum is projected to a metastable vacuum. We study the circumstances under which this may happen and derive general constraints for metastability on the geometry and the gauging. We then study in full detail the simplest case of quaternionic manifolds of dimension four with at least one isometry, for which there exists a general parametrization, and study two types of truncations defining Kahler submanifolds of dimension two. As an application, we finally discuss the case of the universal hypermultiplet of N=2 superstrings and its truncations to the dilaton chiral multiplet of N=1 superstrings. We argue that de Sitter vacua in such theories are necessarily unstable in weakly coupled situations, while they can in principle be metastable in strongly coupled regimes.

hep-th

On Fayet-Iliopoulos terms and de Sitter vacua in supergravity: some easy pieces

We clarify a number of issues on Fayet-Iliopoulos (FI) terms in supergravity, keeping the formalism at a minumum and making use of explicit examples. We explain why, if the U(1) vector is massive everywhere in field space, FI terms are not genuine and can always be redefined away or introduced when they are not present. We formulate a simple anomaly-free model with a genuine FI term, a classically stable de Sitter (dS) vacuum and no global symmetries. We explore the relation between N=2 and N=1 FI terms by discussing N=1 truncations of N=2 models with classically stable dS vacua.

hep-th

On the large spin limit of twist operators in N=4 SYM

The long range Bethe Ansatz solution of the mixing problem in N=4 SYM allows to compute in a very efficient way multiloop anomalous dimensions of various composite operators. In the case of sl(2) twist operators it is important to obtain closed expressions for the anomalous dimensions in terms of the Lorentz spin. Conjectures are available altough analytical proofs are missing beyond one-loop. In this paper, we will present a method to expand at large spin the solution of the long range Baxter equation in twist 2 and 3. We will also propose sum rules for special singlet states at higher twist.

hep-th

Sum rules for higher twist sl(2)operators in N=4 SYM

The spectrum of anomalous dimensions of twist sl(2) operators in N=4 SYM has an intriguing feature in low twist 2 or 3. The anomalous dimension of the lowest state, dual a folded string on AdS_5 X S^5, can be computed by Bethe Ansatz at 3, 4 loops respectively as a simple closed function of the Lorentz spin. This feature is apparently lost at higher twist. We propose sum rules for the excited anomalous dimensions where closed expressions can still be provided, even at higher twist. We present several explicit three loop examples. Many structural regularities can be observed leading to closed expressions which depend parametrically both on the spin and the twist. They allow to compute the subleading term in the logarithmic large spin expansion of the sum rules as a compact simple function of the twist, in analogy with the recent results by Freyhult, Rej and Staudacher in arXiv:0712.2743 [hep-th].

hep-th

Large spin expansion of the long-range Baxter equation in the sl(2) sector of N=4 SYM

Recently, several multi-loop conjectures have been proposed for the spin dependence of anomalous dimensions of twist-2 and 3 operators in the sl(2) sector of N=4 SYM. Currently, these conjectures are not proven, although several consistency checks have been performed on their large spin expansion. In this paper, we show how these expansions can be efficiently computed without resorting to any conjecture. To this aim we present in full details a method to expand at large spin the solution of the long-range Baxter equation. We treat the twist-2 and 3 cases at two loops and the twist-3 case at three loops. Several subtleties arise whose resolution leads to a simple algorithm computing the expansion.

hep-th