Plantilla LaTeX Póster de investigación académica

\documentclass[a1paper,landscape,25pt]{tikzposter}
\usepackage{amsmath,amssymb}
\usepackage{booktabs}
\usepackage{enumitem}
\usepackage{pgfplots}
\usepackage{xcolor}

\pgfplotsset{compat=1.18}

\definecolorstyle{FormatExStyle}{
  \definecolor{colorOne}{HTML}{0F2744}
  \definecolor{colorTwo}{HTML}{2563EB}
  \definecolor{colorThree}{HTML}{EFF6FF}
}{
  \colorlet{backgroundcolor}{colorThree}
  \colorlet{framecolor}{colorTwo}
  \colorlet{blocktitlebgcolor}{colorOne}
  \colorlet{blocktitlefgcolor}{white}
  \colorlet{blockbodybgcolor}{white}
  \colorlet{blockbodyfgcolor}{black}
  \colorlet{innerblocktitlebgcolor}{colorTwo}
  \colorlet{innerblocktitlefgcolor}{white}
  \colorlet{innerblockbodybgcolor}{colorThree}
  \colorlet{titlebgcolor}{colorOne}
  \colorlet{titlefgcolor}{white}
}

\usetheme{Default}
\usecolorstyle{FormatExStyle}

\title{\parbox{0.9\linewidth}{\centering <<poster_title>>}}
\author{<<authors>>}
\institute{<<institute>>}

\makeatletter
\renewcommand\TP@maketitle{%
  \begin{minipage}{\linewidth}
    \centering
    \color{titlefgcolor}
    {\bfseries\Huge\@title\par}
    \vspace{1em}
    {\Large\@author\par}
    \vspace{0.5em}
    {\large\@institute\par}
    \vspace{0.3em}
    {\normalsize <<conference>> \quad|\quad \texttt{<<email>>}}
  \end{minipage}
}
\makeatother

\begin{document}
\maketitle

\begin{columns}
  \column{0.33}

  \block{Introduction}{
    Anomaly detection in distributed systems is critical for maintaining service reliability. Traditional threshold-based approaches fail to capture complex inter-service dependencies.

    \vspace{6pt}
    \textbf{Our approach:} Model the system as a dynamic graph where nodes represent services and edges represent communication patterns. We apply a Graph Neural Network (GNN) to detect anomalous subgraph patterns in real-time.

    \vspace{6pt}
    \textbf{Key contributions:}
    \begin{itemize}[leftmargin=*, nosep]
      \item Temporal graph attention mechanism
      \item Sub-second detection latency
      \item 94.7\% F1-score on production data
    \end{itemize}
  }

  \block{Problem Formulation}{
    Given a system graph $G_t = (V, E_t, X_t)$ at time $t$:
    \begin{align*}
      V &= \{v_1, \ldots, v_n\} \quad \text{(services)} \\
      E_t &\subseteq V \times V \quad \text{(active connections)} \\
      X_t &\in \mathbb{R}^{n \times d} \quad \text{(feature matrix)}
    \end{align*}
    We seek a function $f_\theta$ such that:
    \[ f_\theta(G_{t-w:t}) \to \{0, 1\}^n \]
    classifying each node as normal (0) or anomalous (1) using a sliding window of $w$ graph snapshots.
  }

  \column{0.34}

  \block{Method}{
    Our architecture consists of three components:

    \vspace{6pt}
    \textbf{1. Temporal Graph Encoder.} We use a stack of $L=3$ graph attention layers with temporal self-attention across the window:
    \[ h_i^{(l)} = \sigma\!\left(\sum_{j \in \mathcal{N}(i)} \alpha_{ij}^{(l)} W^{(l)} h_j^{(l-1)}\right) \]

    \textbf{2. Anomaly Scorer.} A two-layer MLP produces per-node anomaly scores $s_i \in [0,1]$.

    \textbf{3. Graph-Level Aggregation.} We aggregate node scores via attention pooling to produce a system-level anomaly indicator for alert routing.

    \vspace{6pt}
    \textbf{Training:} We use a combination of binary cross-entropy on labeled incidents and a contrastive loss on normal operation windows.
  }

  \block{Experimental Setup}{
    \begin{center}
    \begin{tabular}{lr}
      \toprule
      \textbf{Parameter} & \textbf{Value} \\
      \midrule
      Training samples  & 48,000 graphs \\
      Window size $w$   & 10 snapshots \\
      Feature dim $d$   & 32 \\
      GNN layers $L$    & 3 \\
      Learning rate     & $3 \times 10^{-4}$ \\
      Batch size        & 64 \\
      \bottomrule
    \end{tabular}
    \end{center}
  }

  \column{0.33}

  \block{Results}{
    \begin{center}
    \begin{tikzpicture}
      \begin{axis}[
        width=0.8\linewidth,
        height=8cm,
        ybar,
        bar width=18pt,
        xlabel={Method},
        ylabel={F1-Score (\%)},
        symbolic x coords={Threshold,LSTM,GCN,GAT,Ours},
        xtick=data,
        ymin=60, ymax=100,
        nodes near coords,
        every node near coord/.append style={font=\small},
      ]
        \addplot[fill=blue!40, draw=blue!70] coordinates {
          (Threshold,68.2) (LSTM,78.5) (GCN,85.1) (GAT,89.3) (Ours,94.7)
        };
      \end{axis}
    \end{tikzpicture}
    \end{center}

    \vspace{6pt}
    \begin{center}
    \begin{tabular}{lrrrr}
      \toprule
      & \textbf{Precision} & \textbf{Recall} & \textbf{F1} & \textbf{Latency} \\
      \midrule
      Threshold & 71.3\% & 65.4\% & 68.2\% & 12ms \\
      LSTM      & 82.1\% & 75.2\% & 78.5\% & 85ms \\
      Ours      & \textbf{95.1\%} & \textbf{94.3\%} & \textbf{94.7\%} & 42ms \\
      \bottomrule
    \end{tabular}
    \end{center}
  }

  \block{Conclusion}{
    Our temporal graph attention approach significantly outperforms baselines on real production data while maintaining sub-second detection latency. The method generalizes across different microservice topologies without retraining.

    \vspace{6pt}
    \textbf{Future work:} Extend to multi-cloud deployments and integrate causal reasoning for root cause analysis.

    \vspace{6pt}
    {\small\textit{Funded by <<funding>>}}
  }

\end{columns}
\end{document}