These are the most intriguing AI research articles from this year. It blends AI with data science advancements. It is organized chronologically and includes a link to a longer article.

Multi-Modal Summary Generation using Multi-objective Optimization

The rapid advancement of communication technologies in recent years has fueled interest in multi-modal summarization techniques. Most prior efforts on multi-modal summarization have been on text and images. The researchers present a novel extractive multi-objective optimization-based methodology to generate a multi-modal summary combining text, graphics, and videos in this work. 

Multi-objective optimization optimizes intra-modality salience, cross-modal redundancy, and cross-modal similarity for effective multi-modal output. The suggested model was assessed individually for several modalities and shown to outperform state-of-the-art techniques.

Deep Exogenous and Endogenous Influence Combination for Social Chatter Intensity Prediction

Modelling social media user interaction dynamics has exciting applications in user-persona detection and political discourse mining. Most existing techniques rely substantially on understanding the underlying user network. However, many discussions occur on platforms that either need a trustworthy social network or display only a portion of inter-user links (Reddit, Stackoverflow). Many methodologies necessitate observing a debate for an extended time before making useful predictions. Observations incur expenses in real-time streaming scenarios. Finally, most models do not account for intricate interactions between exogenous events (such as externally released news stories) and in-network impacts (such as follow-up debates on Reddit) in determining engagement levels.

To overcome the three limitations mentioned above, the researchers offer ChatterNet, the first framework to model and predict user interaction without considering the underlying user network. The aim is to observe streams of timestamped news stories and talks for a brief period leading up to a time horizon, then predict chatter: the volume of discussions through a set period following the horizon. ChatterNet uses a novel time-evolving recurrent network architecture to analyze text from news and talks, capturing both temporal aspects within the news. The researchers present the results of comprehensive tests conducted with a two-month-long Reddit conversation corpus and a contemporaneous corpus of online news stories from the Common Crawl. ChatterNet outperforms current state-of-the-art engagement prediction models by a wide margin. Controlling observation and prediction windows in detail over 43 different subreddits yields additional exciting insights.

Game Action Modeling for Fine-Grained Analyses of Player Behavior in Multi-player Card Games

The researchers provide a deep learning approach for game action modelling that allows for fine-grained player behaviour analysis. The researchers create CNN-based supervised models that effectively learn crucial gameplay decisions from expert players and then utilize these models to evaluate player qualities in the system, such as retention, engagement, deposit buckets, etc. They demonstrate that the model accurately learns the game's strategies using a carefully developed input format that efficiently captures the game state and history as a multi-dimensional image and a custom architecture. 

It is further improved with self-play simulation look-ahead to estimate the game state better, and this knowledge is employed in a new loss function. They then demonstrate that analyzing players using these models as a guide significantly impacts understanding the player's potential in terms of engagement and income. Researchers utilize the approach to analyze the different circumstances where players make mistakes and upskill them.

Image source: Unsplash