arXiv subjects

Himel Dev

Publications and source records attributed to Himel Dev.

5 recordsLinked to original sources

arXiv2026

From Feasibility to Desirability: Plan, Learn, Adapt (PLA) Framework for Personalized On-Device Itinerary Generation

Generating personalized trip itineraries is a complex planning task and involves a tension between hard combinatorial feasibility and soft latent desirability. Classical optimization enforces constraints but fails to capture subjective traveler preferences. While learning-based approaches model preferences, they cannot guarantee feasibility. Mobile deployment imposes additional resource constraints on both. To address this, we propose Plan, Learn, Adapt (PLA), a three-stage framework for personalized on-device itinerary generation. The Plan stage builds a heterogeneous ensemble of lightweight planners that produces structurally diverse feasible candidates. From pairwise itinerary comparisons, Learn fits a compact Bradley-Terry reward model that captures emergent schedule properties such as pacing, geographic coherence, and day balance, which per-POI signals miss. Finally, Adapt applies feasibility-preserving local refinement within a device-aware compute budget; every intermediate state is feasible by construction. On 2,519 pairwise human comparisons across more than 100 U.S. cities, the reward-guided ensemble achieves a 67.8% win rate, 11.2 percentage points above the best single planner, with 100% feasibility. Three frontier LLMs, GPT-5, Claude Opus 4.5, and Gemini 3 Pro, achieve 0% feasibility under the same constraints. The reward model generalizes across held-out cities, with a 67.6% mean leave-one-city-out accuracy. In production deployment within FlyEnJoy, PLA increased itinerary completion rates by 91%, with 109.9 ms average on-device latency.

cs.LG↗

arXiv2026

Hard Rules, Soft Preferences: Bridging Reasoning, Learning, and Optimization for Personalized Packing Checklist Generation

Packing for air travel is recurring and error-prone: the checklist must be personal and context-aware, yet feasible under safety rules, item dependencies, and luggage limits. Existing packing assistants are template-driven and generic, or recommendation-driven but unconstrained, leaving users to manually patch regulatory and capacity violations. We propose a reasoning-guided learning framework with three stages: (1) a symbolic engine that generates a regulation-aware seed checklist with explicit dependency structure, (2) a two-stage preference learner that estimates inclusion and priority utilities from user add and remove actions while mitigating survivorship bias, and (3) a CP-SAT optimizer that selects a compact, compliant subset. The architecture instantiates a general pattern for constrained personalization, applicable wherever hard feasibility coexists with sparse preference signals. On 604 labeled trip scenarios, comprising 29K inclusion labels and 343K pairwise comparisons, the symbolic engine attains 99.7% recall and 0.96 rubric validity, compared with 0.78 to 0.81 for frontier LLMs. Gradient-boosted trees and LambdaMART reach an AUC-ROC of 0.943 and an NDCG@5 of 0.923. CP-SAT attains 100% constraint satisfaction, compared with 28% for greedy selection and 10% for random selection. Deployment in FlyEnJoy, a production iOS travel app, doubled checklist completions and reduced editing and completion time.

cs.LG↗

arXiv2020

Profiling US Restaurants from Billions of Payment Card Transactions

A payment card (such as debit or credit) is one of the most convenient payment methods for purchasing goods and services. Hundreds of millions of card transactions take place across the globe every day, generating a massive volume of transaction data. The data render a holistic view of cardholder-merchant interactions, containing insights that can benefit various applications, such as payment fraud detection and merchant recommendation. However, utilizing these insights often requires additional information about merchants missing from the data owner's (i.e., payment company's) perspective. For example, payment companies do not know the exact type of product a merchant serves. Collecting merchant attributes from external sources for commercial purposes can be expensive. Motivated by this limitation, we aim to infer latent merchant attributes from transaction data. As proof of concept, we concentrate on restaurants and infer the cuisine types of restaurants from transactions. To this end, we present a framework for inferring the cuisine types of restaurants from transaction data. Our proposed framework consists of three steps. In the first step, we generate cuisine labels for a limited number of restaurants via weak supervision. In the second step, we extract a wide variety of statistical features and neural embeddings from the restaurant transactions. In the third step, we use deep neural networks (DNNs) to infer the remaining restaurants' cuisine types. The proposed framework achieved a 76.2% accuracy in classifying the US restaurants. To the best of our knowledge, this is the first framework to infer the cuisine types of restaurants by analyzing transaction data as the only source.

cs.LG↗

arXiv2019

Quantifying Voter Biases in Online Platforms: An Instrumental Variable Approach

In content-based online platforms, use of aggregate user feedback (say, the sum of votes) is commonplace as the "gold standard" for measuring content quality. Use of vote aggregates, however, is at odds with the existing empirical literature, which suggests that voters are susceptible to different biases -- reputation (e.g., of the poster), social influence (e.g., votes thus far), and position (e.g., answer position). Our goal is to quantify, in an observational setting, the degree of these biases in online platforms. Specifically, what are the causal effects of different impression signals -- such as the reputation of the contributing user, aggregate vote thus far, and position of content -- on a participant's vote on content? We adopt an instrumental variable (IV) framework to answer this question. We identify a set of candidate instruments, carefully analyze their validity, and then use the valid instruments to reveal the effects of the impression signals on votes. Our empirical study using log data from Stack Exchange websites shows that the bias estimates from our IV approach differ from the bias estimates from the ordinary least squares (OLS) method. In particular, OLS underestimates reputation bias (1.6--2.2x for gold badges) and position bias (up to 1.9x for the initial position) and overestimates social influence bias (1.8--2.3x for initial votes). The implications of our work include: redesigning user interface to avoid voter biases; making changes to platforms' policy to mitigate voter biases; detecting other forms of biases in online platforms.

cs.SI↗

arXiv2017

The Size Conundrum: Why Online Knowledge Markets Can Fail at Scale

In this paper, we interpret the community question answering websites on the StackExchange platform as knowledge markets, and analyze how and why these markets can fail at scale. A knowledge market framing allows site operators to reason about market failures, and to design policies to prevent them. Our goal is to provide insights on large-scale knowledge market failures through an interpretable model. We explore a set of interpretable economic production models on a large empirical dataset to analyze the dynamics of content generation in knowledge markets. Amongst these, the Cobb-Douglas model best explains empirical data and provides an intuitive explanation for content generation through concepts of elasticity and diminishing returns. Content generation depends on user participation and also on how specific types of content (e.g. answers) depends on other types (e.g. questions). We show that these factors of content generation have constant elasticity---a percentage increase in any of the inputs leads to a constant percentage increase in the output. Furthermore, markets exhibit diminishing returns---the marginal output decreases as the input is incrementally increased. Knowledge markets also vary on their returns to scale---the increase in output resulting from a proportionate increase in all inputs. Importantly, many knowledge markets exhibit diseconomies of scale---measures of market health (e.g., the percentage of questions with an accepted answer) decrease as a function of number of participants. The implications of our work are two-fold: site operators ought to design incentives as a function of system size (number of participants); the market lens should shed insight into complex dependencies amongst different content types and participant actions in general social networks.

cs.SI↗