Modeling user purchase behavior is a critical challenge in display advertising systems, necessary for real-time bidding. The difficulty arises from the sparsity of positive user events and the stochasticity of user actions, leading to severe class imbalance and irregular event timing. Predictive systems usually rely on hand-crafted "counter" features, overlooking the fine-grained temporal evolution of user intent. Meanwhile, current sequential models extract direct sequential signal, missing useful event-counting statistics. We enhance deep sequential models with self-supervised pretraining strategies for display advertising. Especially, we introduce Abacus, a novel approach of predicting the empirical frequency distribution of user events. We further propose a hybrid objective unifying Abacus with sequential learning objectives, combining stability of aggregated statistics with the sequence modeling sensitivity. Experiments on two real-world datasets show that Abacus pretraining outperforms existing methods accelerating downstream task convergence, while hybrid approach yields up to +6.1% AUC compared to the baselines.

SAR time series are powerful assets for forest monitoring. In recent years, they were involved in various classical forest applications such as forest mapping [9] . These applications largely benefited from the advances of Deep Learning, particularly Unsupervised Learning, using Convolutional Autoencoders in applications such as wildfire detection [4] . Not only did purely temporal approaches offer high prediction performance compared to spatiotemporal variants, but the unsupervised autoencoder rivaled its supervised counterparts. The monitoring of forests also involves the detection of under-canopy targets, which could disturb protected environments. The literature mostly relies on classical SAR approaches PolSAR change detection [8] . A recent shift towards the usage of SAR time series displayed promising performance [10] . Thus, to fully exploit the potential of multi-temporal SAR imagery, this paper proposes the usage of unsupervised Deep Learning, particularly the Convolutional Autoencoder, to detect under forest cover objects.

The C-Band radiometric signature of boreal forests is highly seasonal, with apparent correlations to temperature changes. Within these seasonal components, we assume that information related to tree height can be extracted. We apply a one-dimensional Convolutional Neural Network to assess this assumption, intending to retrieve tree height measured by Airborne Laser Scanning from C-Band Sentinel-1 time series. A study site in the Parc National des Grands Jardins, in Québec, Canada, was selected for this analysis. Prediction-wise, we reach an R2 score of 0.45 and an RMSE of 1.84m, following a 4-fold cross-validation, which exhibits a non-negligible influence of the tree height parameter on boreal forest radiometric response in C-Band Synthetic Aperture Radar, despite the presumed fast saturation of this wavelength, when observing forested environments. In addition to performance metrics, we use a gradient-based explainability tool to diagnose the most contributing periods of the input time series to predict tree height to better correlate the seasonal conditions of this parameter’s influence on the forests’ radiometry.

This paper introduces a tool for explaining the latent space generated by applying convolutional autoencoders to satellite image time series, entitled Grad-SLAM. We rely on backpropagated gradient interpretation combined with network activation localization. We use the proposed formula for multiple layers of the encoder, then scale and merge the results to generate a single date contribution metric for the generation of the latent space. We illustrate the potential of this method with the study of the unsupervised classification of agricultural Sentinel-1 time series. We show that critical characterizing dates for unsupervised retrieval of a given class are conditioned by the crop type’s radiometric signature and class count. We also present how Grad-SLAM can be used to enhance the understanding of unsupervised classification confusion.

We present a fully unsupervised learning pipeline, which involves both a projection method and a clustering algorithm dedicated to the pixel-wise classification of multitemporal SAR images. We design a Convolutional Autoencoder as the method to project our time series onto a lower dimensional latent space, where semantically similar temporal signals are placed close together. The additional use of convolutional layers as feature extraction steps allows us to exploit the sequential nature of time series, exhibiting higher representation performance than fully connected layers. The extracted clusters can encapture different semantic levels to either separate classes or extract outlying temporal signals. The application of this method to crop-types mapping enables the extraction of major crop-types within a scene, without supervision. In a labeled context, this method also allows for the extraction of outlying profiles which can lead to the discovery of mislabeled time series.

Temporal Convolutional AutoEncoders are used as feature extractors to project time series onto a latent space where similarity detection can be easily performed. This model can generate accurate descriptors of the temporal profile of the input time-series. We apply this algorithm to PolSAR S1 uncoherent SAR time series where the model learns highly discriminative data representations. This reduction method is compared to others such as PCA or Temporal Averaging and is shown to outperform them when leveraging the learnt representation using K-Means clustering.

We introduce a multi-branch Deep Learning architecture that allows for the extraction of multi-scale features. Exploiting the data multi-modality structure through the combined use of various feature extractors provides high performance on data fusion tasks. Furthermore, the representation of the multi-temporality of the data using sensor-specific 3D convolutions with custom kernel size extracts temporal features at an early computation stage. Our methodology allows reaching performance up to 0.8876 F1 Score on the development phase dataset and around 0.8798 on the test phase dataset. Finally, we demonstrate the contribution of each sensor to the prediction task with the design of data-focused experiments.