Searching for images using text has limitations because language has difficulties in expressing certain abstract intentions, e.g. artistic styles are difficult to describe for non-experts. As for the image search image model, images can convey abstract intentions, but cannot express the specific purpose, so many of the current graph search works only have a single function, such as content search and style search. Our work aims to combine the strengths of both, merging the ability of text to express specific ideas with the ability of images to convey abstract concepts, thus achieving a better capture of the user's intentions. To this end, we propose CCSR, a multimodal conditional content-style joint retrieval model. Our model is the first to apply contrastive learning to conditional retrieval and introduces a novel Mixture-of-Expert models (MOE) system to enable collaboration between multiple expert systems. We adopt a novel prompt learning strategy that allows the model to adaptively select specific prompts, thereby enhancing its focus on the current task. In addition, to evaluate the joint content-style retrieval capability of our model, we present a new dataset, StyleCoco, containing rich content categories and style categories. The experimental results indicate that CCSR has achieved state-of-the-art performance in conditional style retrieval, content retrieval, and style-content retrieval. The dataset and code will be publicly available on https://mccsr.github.io/.

Xiaoda Yang, Xize Cheng, Jiaqi Duan, Hongshun Qiu, Minjie Hong, Minghui Fang, Shengpeng Ji, Jialong Zuo, Zhiqing Hong, Zhimeng Zhang, Tao Jin. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Abstract The next Point of Interest (POI) recommendation is the core technology of smart city. Current state-of-the-art models attempt to improve the accuracy of the next POI recommendation by incorporating temporal and spatial intervals or by partitioning the POI coordinates into grids. However, they all overlook a detail that in real life, people always want to know where to go at an exact time point or after a specific time interval instead of aimlessly asking where to go next. Moreover, due to individual preferences, different users may visit different places at the same timestamp. Therefore, utilizing timestamp queries can enhance the personalized recommendation capability of the model and mitigate overfitting risks. These implies that using timestamp can achieve more precise recommendations. To the best of our knowledge, we are the first to use the next timestamp for next POI recommendation. In particular, we propose a Time-Stamp Cross Attention Network (TSCAN). TSCAN is a two-layer cross-attention network. The first layer, Time Stamp Cross Attention Block (TSCAB), uses cross-attention between the next timestamp and historical timestamps, and multiplies the attention scores on corresponding POI to predict the next POI that is most related to the history. The other layer, Cross Time Interval Aware Block (CTIAB), applies the time intervals between the next timestamp and historical timestamps to the POI obtained by TSCAB and historical POIs, allowing temporally adjacent POIs to have a greater similarity. Our model not only has a significant improvement in accuracy but also achieves the goal of personalized recommendation, effectively alleviating overfitting. We evaluate the proposed model with three real-world LBSN datasets, and show that TSCAN outperforms the state-of-the-art next POI recommendation models by 5~9%. TSCAN can not only recommend the next POI, but also recommend the possible POI to visit at any specific timestamp in the future.

Large geometry (e.g., orientation) variances are the key challenges in the scene text detection. In this work, we first conduct experiments to investigate the capacity of networks for learning geometry variances on detecting scene texts, and find that networks can handle only limited text geometry variances. Then, we put forward a novel Geometry Normalization Module (GNM) with multiple branches, each of which is composed of one Scale Normalization Unit and one Orientation Normalization Unit, to normalize each text instance to one desired canonical geometry range through at least one branch. The GNM is general and readily plugged into existing convolutional neural network based text detectors to construct end-to-end Geometry Normalization Networks (GNNets). Moreover, we propose a geometry-aware training scheme to effectively train the GNNets by sampling and augmenting text instances from a uniform geometry variance distribution. Finally, experiments on popular benchmarks of ICDAR 2015 and ICDAR 2017 MLT validate that our method outperforms all the state-of-the-art approaches remarkably by obtaining one-forward test F-scores of 88.52 and 74.54 respectively.

LPMLNextends the language of Answer Set Programming (ASP) by assigning a weight degree to each rule so that its stable models do not have to satisfy all LPMLNrules, which is rooted in the manner of Markov Logic Networks (MLN) to handle the uncertainties and inconsistencies in knowledge representation and reasoning. Due to its expressibility, LPMLNhas been employed in several real world applications. However, an LPMLNprogram is much harder to solve than its unweighted counterpart (an ASP program), and only some preliminary solvers have been implemented so far, which is preventing further studies in both theoretical and practical sides. There are three main contributions in this paper. Firstly, we present an LPMLNsolver: LPMLNModels, which is able to run concurrently. Secondly, we present parallel methods in LPMLNModels. For splitting set method, we present an algorithm to generate a proper splitting set, which is an essential part of the method. For augmented subset method, we present a heuristic method to improve its performance. Finally, we present hybrid methods in LPMLNModels to better utilize the parallel methods. The experimental results show that our algorithms and improvements in this paper works and hybrid methods have better performance in general.

Spectrum sharing is envisaged to allow multiple cognitive radio (CR) nodes to jointly use the vacant spectrum resource for opportunistic transmission. Most previous work on the spectrum sharing did not take into account the spectrum heterogeneity, such as different transmission ranges, error rates, etc. In this paper, a spectrum sharing algorithm is presented that accommodates spectrum heterogeneity for centralized CR networks, meanwhile the individual users are free to be mobile. Each node is uniquely assigned a channel while considering its channel occupancy time, the location of the node, and the fairness of channel access opportunity. Each of these metrics contributes in part to a larger optimization problem and the final objective is to maximize the total utility of system. Simulations results reveal significant improvement in reducing spectrum handoffs.