Detailed Program
| Standard time zone: | UTC/GMT +5 hours | ||
| 09.00 - 09.50 | Registration is open | ||
| Coffee and tea available - Dining Hall | |||
| 09.50 - 10.00 | Opening remarks | ||
| Aleksey Buzmakov, Kai Heinrich, Dmitry Ignatov, Rustam Tagiew | |||
| 10.00 - 11.00 | KEYNOTE TALK | ||
| Professor Ayelet Gneezy, Rady School of Management, UC San Diego, USA | |||
| "Field Experimentation in Marketing Research" | |||
| This talk reviews the results published in Journal of Marketing Research 54(1) in 2016 under the same title. "Despite increasing efforts to encourage the adoption of field experiments in marketing research (e.g., Campbell 1969; Cialdini 1980; Li et al. 2015), the majority of scholars continue to rely primarily on laboratory studies (Cialdini 2009). For example, of the 50 articles published in Journal of Marketing Research in 2013, only three (6%) were based on field experiments. The goal of this article is to motivate a methodological shift in marketing research and increase the proportion of empirical findings obtained using field experiments. The author begins by making a case for field experiments and offers a description of their defining features. She then demonstrates the unique value that field experiments can offer and concludes with a discussion of key considerations that researchers should be mindful of when designing, planning, and running field experiments." | |||
| 11.00 - 12.30 | Session 1 | Regular Talks | |
| 11.00 - 11.30 | Anastasiia Saltykova and Agata Lozinskaia | Fundamental Factors Affecting the IMOEX Russia Index: Retrospective Analysis | |
| This paper is an empirical study of the changing nature of the dependence of fundamental factors on the stock market index, which is the trend identified earlier in the Russian stock market. We empirically test the impact of daily values of fundamental factors on the MOEX Russia Index from 2003 to 2018. The analysis of the ARIMA-GARCH (1,1) model with a rolling window reveals that the change in the power and direction of the influence of the fundamental factors on the Russian stock market persists. The Quandt-Andrews breakpoint test and Bai-Perron test identify the number and likely location of structural breaks. We find multiple breaks probably associated with the dramatic falls of the stock market index. The results of the regression models over the different regimes, defined by the structural breaks, can vary markedly over time. This research is of value in macroeconomic forecasting and the investment strategy development. | |||
| 11.30 - 12.00 | Dmitry Akimov and Ilya Makarov | Deep Reinforcement Learning with VizDoom First-Person Shooter | |
| In this work, we study deep reinforcement algorithms for partially observable Markov decision processes (POMDP) combined with Deep Q-Networks. To our knowledge, we are the first to apply standard Markov decision process architectures to POMDP scenarios. We propose an extension of DQN with Dueling Networks and several other model-free policies to training agent using deep reinforcement learning in VizDoom environment, which is replication of Doom first-person shooter. We develop several agents for the following scenarios in VizDoom first-person shooter (FPS): Basic, Defend The Center, Health Gathering. We compare our agent with Recurrent DQN with Prioritized Experience Replay and Snaphot Ensembling agent and get approximately triple increase in per episode reward. | |||
| 12.00 - 12.30 | Daria Teterina and Maria Temirkaeva | The Comparison of Methods for Individual Treatment Effect Detection | |
| Today, treatment effect estimation at the individual level is vital problem in many areas of science and business. At the same time, the issue of the most effective methods, that is, those that give the smallest predictive error (for instance, RMSE), remains open. In this paper we contribute to the answer on the above question by comparing the effectiveness of machine learning methods for estimation of individual treatment effects. The comparison is performed on the Criteo Uplift Modeling Dataset. According to the research results, the combination of the Random Forest method and the modified outcome approach proved to be the best one with respect to total treatment effect on the top 30% observations of the test dataset. | |||
| 12.30 - 13.00 | Coffee break - Dining Hall | ||
| 13.00 - 14.30 | Session 2 | Regular Talks | |
| 13.00 - 13.30 | Ilya Makarov and Pavel Zolnikov | Effective Algorithms for Constructing Multiplex Networks Embedding | |
| Network embedding has become a very promising technique in analysis of complex networks. It projects nodes of a network into a lower-dimensional vector space but still retaining the structure of the network. There are many methods of network embedding developed for traditional single layer networks. On the other hand multilayer networks can provide more information about relationships between nodes. In this paper we compare methods of embedding multilayer networks with methods of embedding multilayer networks and demonstrate their effectiveness. For this purpose we use several classic datasets usually used in network embedding experiments. We also introduce new method of network embedding which uses random walks generated in specific way. We test it against several baseline methods on our new dataset we obtained to measure its effectiveness. | |||
| 13.30 - 14.00 | Anna Muratova, Robiul Islam and Ekaterina Mitrofanova | Sequence Mining for Searching Interpretable Demographic Patterns | |
| Network embedding has become a very promising technique in analysis of complex networks. It projects nodes of a network into a lower-dimensional vector space but still retaining the structure of the network. Nowadays there is a large amount of demographic data which should be analyzed and interpreted. From accumulated demographic data, more useful information can be extracted by applying modern methods of data mining. Two kinds of experiments are considered in this work: 1) generation of additional secondary features from events and evaluation of its influence on accuracy; 2) exploration of features influence on classification result using SHAP (SHapley Additive exPlanations). An algorithm for creating secondary features is proposed and applied to the dataset. The classifications were made by two methods: SVM and neural networks, and the results were evaluated. Also features impact on the classification results was evaluated using SHAP. It was demonstrated how to tune model for improving accuracy based on features impact. | |||
| 14.40 - 14.30 | Vasiliy Osipov, Nataly Zhukova and Dmitriy Miloserdov | Neural Network Associative Forecasting of Demand for Goods | |
| This article discusses the applicability of recurrent neural networks with controlled elements to the problem of forecasting market demand for goods on the four month horizon. Two variants of forecasting are considered. In the first variant, time series are used to train the neural network, including the real demand values, as well as pre-order values for 1, 2 and 3 months ahead. The forecast horizon is 4 months. In the second variant, there is an iterative forecasting method. It predicts the demand for the next month at each step, and the training set is supplemented by the values predicted for the previous months. It is shown that the proposed methods can give a sufficiently high result. At the same time, the second approach demonstrates greater potential. | |||
| 14.30 - 15.30 | Lunch (self serve) | ||
| 15.30 - 16.10 | INDUSTRY TALK | ||
| Vasily A. Leksin, PhD, Unit Leader in Recommender Systems, AVITO.ru, Moscow | |||
| Recommender Systems for online classified advertisements | |||
| I’ll tell about our journey from simple linear models to neural networks in the development of Recommender Systems in Avito and the current recommendations architecture. I will share successes and failures in building a highly loaded Recommendation Systems. We will discuss the gap between the latest scientific approaches and methods applicable in production systems and why we claim that the “keep it simple” principle works well in production systems. | |||
| 16.10 - 17.50 | Session 3 | Short Talks and Project Proposals | |
| 16.10 - 16.30 | Ilya Makarov and Alsu Zaynutdinova | Deception Detection in Online Media | |
| Russian Federation and European Union are fighting against fake news together with other countries. The consequences can be observed today in different countries in various topics. The disinformation affected British referendum of existing EU, the US election and Catalonia's referendum are broadly studied. A need for automated fact-checking increases, European Commission's Action Plan 8 is an evidence. In this work we develop a model for detecting disinformation in Russian language in online media. We use reliable and unreliable sources to compare named entities and verbs extracted using DeepPavlov library. Our method shows four time greater recall compared to chosen baseline. | |||
| 16.30 - 16.50 | Timur Kadyrov and Dmitry Ignatov | Attribution of Customer’s Actions Based on Machine Learning Approach | |
| Multichannel attribution model based on gradient boosting over trees is proposed, which was compared with the state of the art models: bagged logistic regression, markov chains approach, shapley value. Experiments on digital advertising datasets showed that the proposed model is better than the solutions considered by ROC AUC metric. In addition, the problem of probability prediction of conversion by the consumer using the ensemble of the analyzed algorithms was solved, the meta-features obtained were enriched with consumers and offline activities of the advertising campaign data. | |||
| 16:50 - 17.10 | NameNikita Benkovich, Roman Dedenok and Dmitry Golubev | DeepQuarantine | |
| In this paper, we introduce DeepQuarantine (DQ), a cloud technology to detect and quarantine potential spam messages. Spam attacks are becoming more diverse and can potentially be harmful to email users. Despite the high quality and performance of spam filtering systems, detection of a spam campaign can take some time. Unfortunately, in this case some unwanted messages get delivered to users. To solve this problem, we created DQ, which detects potential spam and keeps it in a special Quarantine folder for a while. The time gained allows us to double-check the messages to improve the reliability of the anti-spam solution. Due to high precision of the technology, most of the quarantined mail is spam, which allows clients to use email without delay. Our solution is based on applying Convolutional Neural Networks on MIME headers to extract deep features from large-scale historical data. We evaluated the proposed method on real-world data and showed that DQ enhances the quality of spam detection. | |||
| 17.10 - 17.30 | Friedrich Michael | General Game Playing B-to-B Price Negotiations | |
| This papers discusses the scientific and practical perspectives of using general game playing in business-to-business price negotiations. The status quo of digital price negotiations software, which emerged from intuitive business needs and referred to as electronic auctions in industry, is summarized for the scientific world. Description of such aspects as auctioneers interventions, asymmetry among players and time depended elements reveals the nature of nowadays electronic auctions be rather termed as price games. This paper strongly suggests general game playing as the crucial technology for automation of human ruler setters in those games. Game theory, experimental economics and AI human player simulation are also discussed as satellite topics. State-of-art game descriptions languages are presented, as well as their formal game theoretic foundations. | |||
| 17.30 - 17.50 | Sobana Sumi P and Radhakrishnan Delhibabu | Glioblastoma Multiforme Classification On High Resolution Histology Image Using Deep Spatial Fusion Network | |
| Glioblastoma Multiforme is a kind of brain tumor. Glioblastoma Multiforme (GBM) tumor is one of the cancerous tumors with scarce symptoms and classifying this tumor is tedious. Brain tumor can be normal (benign) or cancerous (malignant) based on its characteristics. Brain tumor is classified into Low Grade Gliomas (LGG) as grade 1, grade 2 and High Grade Gliomas (HGG) as grade3, grade 4. Grade 1 and grade 2 is Astrocytomas, grade 3 and grade 4 is Oligodendrogliomas and Glioblastoma Multiforme. Histology image diagnosis is a suitable way to classify brain tumor. Biopsy is a process of taking tissue from tumor and analysing it through microscope is called histology image. Sometimes this histology image analysis is not agreed due to its variations in morphological features. To overcome this problem, a systematic way of feature extraction and classification is needed. This can be satisfied by deep learning technique of Convolutional Neural Network (CNN). Training directly CNN with high resolution histology image is computationally high. Also uncertain selective features given over the histology image, causes challenges in patch based CNN classification. Here proposed a method for classifying high resolution histology image. CNN architecture of InceptionResNetV2 is adapted first and extracted features without loss. Then generated deep spatial fusion network to make use of spatial features between patches and to predict true features from uncertain selective features. 10 fold cross validation is performed on histology image. This achieve 96 percent accuracy on 4 class classification (benign, malignant, Glioblastoma, Oligodendrogliomas). 99 percent accuracy and 99.5 percent AUC on 2 way classification (necrosis and non necrosis) is obtained by proposed work. | |||
| 17:50 - 18:00 | Closing remarks | ||