(
methodological focus,
technical focus)
Master Thesis at ISMLL
Past Master theses
Character- Level Text Classification via Deep Networks
The ultimate objective of NLP is to accomplish human-like language processing. Text classification is the process of labeling different texts with some labels according to its content. Text classification lies in a number of disciplines like document classification, web search, ranking, and information retrieval. Very Deep Networks has shown to outperform the baselines only in Computer Vision domain. We demonstrate that we can apply deep learning to text understanding from character level inputs using Deep Convolutional Neural Network. The sentences are as a sequence of tokens (characters) and process them with a Deep Convolutional Neural Network. A data augmentation method is used for increased accuracy. We experimented data augmentation by using an English thesaurus, which use synonyms to replace words or phrases in the text. Another experiment of my research is to study the effect of POS tags on the performance of the generated text classifier. Part of speech tags are features that can help enhance classifiers by allowing us to better understand the purpose of word choice and the meaning of sentences as a whole. We also give attention to a challenging part in NLP, dealing with different language text using Binary Encoding. We treat the text (in UTF-8) as a sequence of bytes and encode at byte-level. The advantage of byte-level processing is that they can handle different language at byte-level. For the experimental evaluation, we make use of the public dataset with both English and Chinese character including sentiment analysis and text categorization.
Deep Collaborative Filtering
Everyone is using the internet, then everyone has experienced a website which suggests him a song, a video or a product to buy. Most of the websites nowadays are heavily relying on the fact that the better they recommend to the user, the better the user experience on that website. In this thesis, I am proposing a study, research and develop techniques based on neural networks to tackle the key problem in recommendation (collaborative filtering) on the basis of implicit feedback. Recent research adopted deep learning in dealing with the recommendation problem, but they are still applying matrix factorization which implicitly applying the inner product on the latent features of users and items. Here I propose to take over the inner product with a multi-layer perceptron neural network that can learn an arbitrary function from data, which can express and generalize matrix factorization. I also propose adding new features and changing of the already-exist architecture for producing better results
Prediction of Driver’s Destination and its partial trajectory Using Deep Learning Vehicle destination prediction is an essential task in various intelligent vehicles for a number of reasons which includes optimal route finding, optimizing vehicle energy savings strategy and driving costs. As navigation system becomes part of our daily life we progressively benefit from various position-based services for example upcoming point of interest, traffic hazards, social networking based on location, recommending advertisements, and to automatically set destination in navigation systems. This Thesis aims at predicting the future destination of the driver based on its partial trajectory using deep learning architecture. Various neural network approaches have been explored such as Sub-Trajectory Synthesis (SubSyn), Hidden Markov model, Multi-layer perceptron, Spatial- Temporal recurrent neural network, Bi-directional recurrent neural network, and Long Short- term memory. A common approach for trajectory prediction is to input the historical trajectories and to derive the probability of the future destination location. For the experimental evaluation, we make use of the public historical spatial trajectories acquired from observing 182 users in a period of four years. However, lots of challenges are involved in using the real-world GPS data, such as long raw trajectories, which limits the performance of various models and data sparsity problem. In this thesis, you Investigate various hybrid models to select the best model which gives the most accurate predictions for a user’s trajectory. Keywords: Destination prediction, Sub-Trajectory Synthesis, Hidden Markov model, Multi-layer perceptron, Spatial-Temporal recurrent neural network, Bi-directional recurrent neural network, Long Short-term memory, Spatial trajectory, Data sparsity, Hybrid models.
- 1. Endo Y, Nishida K, Toda H, Sawada H: Predicting Destinations from Partial Trajectories Using Recurrent Neural Network. NTT Service Evolution Laboratories. PAKDD 2017
- 2. A. de Brebisson, E. Simon, A. Auvolat, P. Vincent, and Y. Bengio. Artificial neural networks applied to taxi destination prediction. ECML/PKDD 2015
- 3. Q. Liu, S. Wu, L. Wang, and T. Tan: Predicting the Next Location: A Recurrent Model with Spatial and Temporal Contexts, 194–200, AAAI 2016. 4. Fan Wu and Kun Fu and Yang Wang and Zhibin Xiao and Xingyu Fu: A Spatial-Temporal- Semantic Neural Network Algorithm for Location Prediction on Moving Objects. Algorithms 2017.
Exploration in RL using Curiosity and Distortion Reinforcement Learning has a strong foundation with continuous rewards for the exploration of the environment with On-policy and Off-policy settings. An agent learns by receiving different rewards from the environment that comes in multiple formats either dense, sparse or no reward and makes exploration difficult. Curiosity-Driven Learning (ICM) has proved to be one of the key components of learning in sparse environments as it uses error for its reward using non-linear functions with two forward and inverse modules. Curiosity was published with on-policy agent A3C (asynchronous advantages actor-critic) and PPO (Proximal Policy Gradient). Here we propose two experiments, first to use interest with off-policy agent ACER (Sample Efficient Actor-Critic with Experience Replay) as its the counterpart of A3C with experience replay. Previously random features have also been shown as essential for exploration, so for our second experiment, we propose to have to have distortion on our ICM model to avoid getting stuck in some local minima and also to have a better generalization of the environment. For evaluation, we use returned extrinsic reward as a result but without the use of gradients on dense extrinsic rewards for training.
- 1. Pathak, Deepak, et al. "Curiosity-driven exploration by self-supervised prediction." International Conference on Machine Learning (ICML). Vol. 2017. 2017.
- 2. Burda, Yuri, et al. "Large-scale study of curiosity-driven learning." arXiv preprint arXiv:1808.04355 (2018).
- 3. Mnih, Volodymyr, et al. "Asynchronous methods for deep reinforcement learning." International conference on machine learning. 2016.
- 4. Wang, Ziyu, et al. "Sample efficient actor-critic with experience replay." arXiv preprint arXiv:1611.01224 (2016).
- 5. Mnih, Volodymyr, et al. "Playing atari with deep reinforcement learning." arXiv preprint arXiv:1312.5602 (2013).
Boosting Multi-step forecasts via Multi-task learning
Time series forecasting helps organizations with capacity planning and goal setting. We look into a multivariate setting where a target and auxiliary series are to be forecasted for multi- step ahead. This results in an interesting multi-task learning problem formulation where the learning tasks come from future horizons of target and auxiliary series both. Multi-task approaches rely on enumerating multiple network architectures to balance the amount of shared and non-shared layers between tasks. Also, as multi-horizon strategies minimize forecast errors over multiple horizons rather than just over the horizon of interest, loss functions would be at different scales based on the time steps being in near or distant future respectively. This thesis aims to show performance is sensitive with respect to the choice of sharing layers and weighting between the auxiliary tasks and the multi-step ahead tasks both. The proposed method balances these configurations optimally, resulting in superior performance, compared with several baselines. An extensive ablation study is carried out over real-world datasets to establish the effectiveness of proposed approach.
Dynamic Recommender Systems A Hybrid Approach
Recommender systems are used in many fields of society. Current methodologies rely mostly on static information to predict ratings given to items by users. These static recommender systems assume no change in user behavior over time and all timesteps are assumed to have the same importance. Nevertheless, in real life scenarios, temporal context is an important matter. A user's taste can change, and an item's perception can differ from season to season. A model is needed that can predict the future behavior of users and items, without the need of manually updating the system. This can be achieved by combining a simple static matrix factorization model with LSTMs. While the static model learns stationary effects, the LSTMs can capture long and short-term dynamics and thus a change in behavior. By passing additional temporal information, rating predictions can be made for any point in time. Experiments on three datasets show that such hybrid recommender system outperforms multiple baselines when predicting future ratings.
Generating Time Series’ Latent Factors with Variational Autoencoders This thesis studies the behavior of time series by investigating its latent components using a deep learning architecture coupled with a Bayesian inference, known as Variational Autoencoder. Applied to the Euro/Dollar daily exchange rate series, this methodology generates two latent components that could be used as proxy for a noise and a cycle, as it describes the time series’ movements and encompasses main statistical properties of this type of data.
Spatio-Temporal Forecasting: Parking Prediction Finding a parking space is is always a strenuous task for drivers in urban sce- narios, often due to the lack of actual parking availability information for drivers. Therefore, it becomes necessary to provide a car parking availability prediction ser- vice which could inform car drivers about empty parking locations before reaching their final destination. The goal of parking prediction is to predict the future occu- pancy status of a parking location over a relatively short period of time. However, traditional parking prediction methods mostly rely on time series forecasting tech- niques, which fail to model the complex non-linear spatotemporal correlations. In this work, we formulate parking prediction as a spatiotemporal sequence forecast- ing problem. Further, we investigate the use of deep learning approaches to model this problem. The proposed network architectures utilize the decoder-encoder framework based with different types of layers. In addition, two loss techniques were introduced to enable the proposed models to reach better performance in less training time. Experimental results on the parking dataset demonstrate that the spatiotemporal approaches provide better results than traditional time series state-of-the-art methods that totally ignore the spatial correlations. Specifically, one model, named ConvLSTM, outperforms all remaining models thanks to its ConvLSTM cells that are able to preserve the spatiotemporal correlations of the data.
Bounding Box Regression: Replacing the Proxy Loss in Bounding Box Localization with Intersection Over Union
Object detection is an important part of Computer Vision domain that has seen remarkable progress in the recent years with the introduction of Convo- lutional Neural Networks (CNN). Object detection is a Multi-task learning problem where both the position of the objects in the images as well as their class needs to be correctly identified. The outcome of an object detec- tion task are the bounding boxes around the objects and their correct labels. The idea here is to maximize the overlap between the ground truth bounding boxes and the predictions i-e the overlap or more directly, the Intersection over Union (IoU). In the current scope of work seen in this domain, IoU is approximated by using the Huber loss as a proxy loss but this indirect method is not the best way since the optimization and evaluation is done on different metrics. In this thesis we have formulated two novel losses namely, the Smooth IoU and the Hard Switch loss, which directly optimize the IoUs for the bounding boxes. These losses have been evaluated on the Oxford IIIT Pets, Udacity self driving car and PASCAL VOC datasets and have shown performance gains over the standard Huber Loss.
DECISION SUPPORT FOR FINANCIAL DISCLOSURES USING DATA SCIENCE (STOCK FORECASTING USING DEEP LEARNING)
Traditional financial trading by a large industry has grown up around the implication proposition that some analysts can do the job well and predict stocks better than others. Also, not to forget there are a lot of big companies spending tremendous amount of money on hiring experts to predict stock prices using statistical models. This work explores the efficiency of Deep Learning in Time Series Forecasting, precisely to say Stock Market Forecasting. More precisely recent advancements and improvements in Recurrent Neural Networks called Gated Recurrent Units (GRU) have been explored in Stock Market Forecasting and the results seem promising when compared with standard baselines. The entire work of thesis explores three different models in predicting the stock movement.
Mini-Batch Optimization in a Distributed Setting
The gradient based methods are used in large scale optimization problems. Especially, Mini-Batch optimization is well suited for learning in distributed settings, because it reduces the communication cost. However, the larger the Mini-Batch size is, the more it diminishes the convergence rate of the model parameters. In an ideal scenario the Mini-Batch size should be set as large as possible without influencing negatively the convergence rate. In this work, we would like to experiment with different algorithms and techniques that address the problems associated with Mini-Batch optimization in distributed settings. The experiments reach from different Mini-Batch sizes and a varying number of features over the error rate and the time to search for the final solution. Another aspect to be considered is the influence of memory size. It is expected that an optimal algorithm will be proposed that will overcome the shortcomings in a vanilla Mini-Batch algorithm.
Wissenstransfer in Empfehler Systemen mit teilweiser Überschneidung durch Ähnlichkeitsbedingung
Transfer learning (TL) bietet eine Möglichkeit, die Vorhersage von Empfehler Systemen zu verbessern, indem ein oder mehrere Systeme als Hilfssysteme das Zielsystem mit zusätzlichen Informationen unterfüttern. Dabei können unterschiedliche Anforderungen an die Hilfssysteme gestellt werden. Während die Forderung, dass sich sowohl die Kunden als auch die Produkte in allen Systemen gleichen, in der Realität nur bedingt anzutreffen ist, führt die Annahme, dass weder Kunden noch Produkte übereinstimmen dazu, dass man sich auf eine gewisse Verwandtschaft der Kunden und Produkte zwischen den Systemen verlassen muss. Weitaus realistischer scheint es, dass sich Kunden oder Produkte zum Teil überschneiden. Wie man diese teilweise Überschneidung von Produkten in unterschiedlichen Empfehler Systemen nutzen kann, versucht diese Arbeit zu erörtern. Dafür wird eine nicht negative Matrixfaktorisierung angewendet, die als zusätzliche Bedingung die Ähnlichkeit von Produkten in den Systemen enthält. Dabei ist die Grundidee, dass die Ähnlichkeit von zwei Produkten, die in mehreren Systemen vorkommen, innnerhalb eines Systems in etwa genau so groß sein sollte wie die Ähnlichkeit dieser zwei Produkte in einem anderen System. Die Experimente zeigen, dass die nicht negative Matrixfaktorisierung durch diesen Ansatz verbessert werden kann.
Non-cyclic Feature Extraction for Accelerometer-Based Gait Recognition on Smartphones
The arise of smartphones with their capability to store various types of sensitive data, have had a big impact on information science and particularly challenges the field of information security where access- control is a crucial task to protect data against compromisation.Utilizing the smartphone's sensors, methods for user-to-computer authentication have been developed that overcome the inconvenience of methods that are requiring the subject to memorize a given PIN, password or gesture and provide it to the system during the authentication process. In contrast to methods that are authenticating subjects by biometric properties that needs to get captured explicitly (like the fingerprint or the pattern of the iris), accelerometer-based methods as discussed in this thesis are based on the subject's gait, which is a natural and implicit action. A method for accelerometer-based gait-recognition has been developed by Nickel et al., having the distinctness of extracting frequency cepstral coefficients, which are known from the field of audio recognition, from a subject's gait signal. Amongst other features, these coefficients are passed to a SVM.The major steps of the cepstral coefficient extraction will get examined in detail. The goal of the thesis is an implementation of the aforementioned method in JAVA that is evaluated on the ZJU-GaitAcc dataset.
Visualization of machine data
In today's industry 4.0 machine generate information. By machine data collection, it is possible to store this information. A Manufacturing Execution System (MES) enables prepare this machine data and visualize. This work gives examples of the prior art and documented its own prototype, which visualizes the machine data. For visualization, there are different points of view in the machine data. An example of the machine data to visualize the overall equipment effectiveness (OEE) analysis, which is used to compare similar machines or different time periods. Means of comparing, improvement can be recognized by a machine.
An Experimental Study of Image Feature Representations for Embedded Traffic Sign Recognition
The department CC-DA at Robert Bosch GmbH is currently developing the third generation of camera platforms for video-based driver assistance systems. Among other subsystems in development for that platform, the module RSR (Road Sign Recognition) is in charge of identifying and displaying sign related information to the driver. Continually increasing customer requirements with respect to sign coverage, recognition rates and runtime performance necessitate the analysis of alternative approaches. The proposed master thesis aims for the analysis of different feature representations of traffic signs and their influence on the classification performance. Observing the limitations of the project's hardware resources constitutes a crucial factor in the selection of algorithms. One traditional family of representations is based on projecting data into a lower dimensional representation, such that Principal Component Analysis (PCA), Discrete Cosine Transformation (DCT), or label-supervised decompositions such as Linear Discriminant Analysis (LDA).Another direction to be assessed is the bag-of-local-features approach where images are characterized as histograms of local patterns. Pioneering alternatives are scale-invariant-feature-transform representation (SIFT) and gradient histograms (HOG). Recent state of-the-art representations focus on variations of spatial-pyramids of local bag-of-features and sparse codings of local feature histograms. The thesis aims at evaluating of the proposed methods for RSR regarding classification and runtime performance.
Personalized Recommender Systems in the Context of Built-to-Stock Vehicles
Auf den meisten europäischen Märkten bieten Händler Lagerfahrzeuge an, die vor dem Eingang einer konkreten Kundenbestellung beim Hersteller bestellt werden müssen. Ursächlich dafür ist, dass Fahrzeuge mit gewissen Ausstattungsmerkmalen als Vorführwagen und auch Exemplare zum direkten Verkauf bereitgehalten werden sollen. Dies stellt die Händler vor eine komplexe Entscheidungsproblematik, da sie Fahrzeugkonfigurationen für zukünftige Kundenaufträge voraussehen müssen. Um die Bewältigung dieser schwierigen Aufgabe zu unterstützen, wird in dieser Arbeit ein Empfehlungssystem entwickelt. Es soll dem Händler bei der Zusammenstellung der grundlegenden Komponenten, so genannter Heavy Items , repräsentativer Lagerfahrzeuge mit hohem Marktpotential helfen. Ein solches Unterstützungssystem ist im aktuellen Prozess der händlerseitigen Lagerfahrzeug-Bestellungen nicht vorhanden. Vielmehr verlässt sich der Händler derzeit beim Bestellvorgang überwiegend auf seine Erfahrung und seine Intuition.
Das Empfehlungssystem soll in der Lage sein, für jeden Händler eine individuelle, nach Empfehlungsrelevanz geordnete, Vorschlagsliste zurückzuliefern. Eine Besonderheit in diesem speziellen Fall stellen die Quelldaten dar. Denn sie bestehen aus Fahrzeugen, die sich wiederum aus verschiedenen Ausstattungsmerkmalen zusammensetzen. Neben dem neuartigen Anwendungskontext wird im Rahmen dieser Arbeit, auch aufgrund der Factorization Machine und des Bayesian Personalized Ranking , ein neues Verfahren, welches speziell auf das Empfehlungsproblem und die Quelldaten abgestimmt ist, entwickelt.
Design and Implementation of a Web Extraction Tool (Wrapper), using Machine Learning ApproachesWrappers are used to extract relevant information from Webpages and convert them into a structured format. There are several steps in achieving this. For basic wrapper, the structure of the webpage has to be examined to identify tags and retrieve information related to this tag. However, coding wrappers by hand is time consuming and tedious. Therefore, more sophisticated algorithms have been researched in this area such as applying induction techniques for automatically learning wrapper. However, most of these algorithms can only handle extracting data from a single webpage. Kovalev et al. have proposed to extend these algorithms to extract and transform a set of hyperlinked hidden web query results to XML.
The tasks of this Master work are to investigate:
Evaluation of compression methods for the IBM BI AcceleratorTBA
Past Master theses at University of Freiburg