Dissertation
Building interpretable machine learning models for sequential data
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Summer 2023
DOI: 10.25820/etd.006992
Abstract
Machine learning is progressing at an astounding rate.The past decade has seen an explosion in the amount of machine learning research, including deep learning techniques. These modern techniques have proven highly effective in a variety of domains, such as healthcare, finance, genomics, image processing, text analytics, etc. Despite the successes, machine learning has its own limitations and drawbacks. The most significant one is the lack of transparency behind their behaviors, which leaves users with little understanding of how particular decisions are made by these models. There has recently been an increasing amount of work in explanatory artificial intelligence (XAI). This research area tackles the important problem that complex machines and algorithms often cannot provide insights into their behavior and thought processes. XAI allows parts of the internal system to be more transparent to users, providing explanations of its decisions in some level of detail. These explanations are important to identify potential biases/problems of the predictive models, and to ensure the fairness of the algorithms.
Although there is a large quantity of research and empirical success of XAI research, most of the existing work has focused on tabular data and image processing techniques. Interpreting machine learning models for sequential data has received much less attention comparatively. Given recent machine learning advancements in highly sequential domains such as natural language processing, e-commerce, healthcare, etc., the need for sequential explanations is, more than ever, critical.
In this thesis, we propose several methods for explaining and understanding sequential machine learning models. This thesis consists of three parts. In the first part, we discuss AdaAX - an effective method to learn deterministic finite automaton (DFA) as a post-hoc explanation of recurrent neural networks (RNN). A DFA, which defines a set of states and transition rules between them, can be used to explain the internal behavior of any sequence model. Comprehensive evaluations show that AdaAX can generate DFA that is easy to interpret and outperforms other baseline models on both synthetic and real-world datasets. In the second part, we introduce ProtoryNet, a model with inherent interpretability based on the concept of prototype trajectory. ProtoryNet will transform any text input into a prototypical text sequence, which is a sequence of trained prototypical sentences, before generating the prediction. Extensive experimental results on a variety of text datasets demonstrate the superior performance of ProtoryNet. Finally, in the third part, we introduce Personalized Path Recourse (PPR), a novel method that generates alternative action paths for a given path from an agent, with the goal of achieving better outcomes while satisfying both similarity and personalization requirements. The proposed method is applicable to both reinforcement learning and supervised learning settings for correcting or improving sequences of actions of sequences of data to achieve a pre-determined goal. The method is evaluated in various settings and demonstrates promising results.
Details
- Title: Subtitle
- Building interpretable machine learning models for sequential data
- Creators
- Dat Hong
- Contributors
- Tong Wang (Advisor)Alberto Maria Segre (Committee Member)Padmini Srinivasan (Committee Member)Bijaya Adhikari (Committee Member)Qihang Lin (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Computer Science
- Date degree season
- Summer 2023
- Publisher
- University of Iowa
- DOI
- 10.25820/etd.006992
- Number of pages
- xiii. 102 pages
- Copyright
- Copyright 2023 Dat Hong
- Language
- English
- Date submitted
- 06/12/2023
- Description illustrations
- illustrations, tables, graphs
- Description bibliographic
- Includes bibliographical references (pages 93-102).
- Public Abstract (ETD)
- Machine learning is a branch of artificial intelligence where we teach computers how to learn. For instance, we can train a program to recognize if an image shows a dog or a cat. The trained program is called a model, and it can predict the label of a new image. While machine learning models can be highly accurate, their decision-making process is often unclear to humans. As a result, there is a growing demand for interpretable machine learning due to increased requirements for transparency in algorithms and data by governments and businesses. This field focuses on developing techniques to explain how machine learning models work and has gained attention from both researchers and industry.
Interpretable machine learning encompasses two main categories of methods: those that explain pre-trained models and those that create new models that are easier for humans to understand. In this thesis, we propose various methods that belong to both categories.
The first method explains a recurrent neural network, a specialized network used for analyzing time-series data. Our method generates a diagram to explain how the network makes predictions after training it to determine the sentiment (positive or negative) of a text. For example, the diagram may reveal that if a text begins with words like “wonderful” or “great food,” it will be classified as having a positive sentiment.
In the second method, we propose a new, self-explanatory network architecture. This network is built based on the concept of prototypes, which are representative examples in the training data that can stand for multiple other examples. Before generating a prediction, the network maps any input text sequence to a sequence of prototypes called a prototype trajectory. For instance, sentences like “The food is fantastic,” “Amazing food” and “All are delicious” can all be represented by a single prototype sentence like “Food is great.” By examining the prototypes, users can understand how the model reaches its predictions.
Finally, our third method explains predictions at the instance level using personalized recourse or counterfactual explanations. Given an input sequence with a specific label, our method generates a similar sequence, but with the label reversed. Additionally, the new sequence can be personalized based on the desired style. For example, if the original sequence is “The food is terrible” with a negative sentiment, our method can generate a new sentence that is similar but with a positive sentiment, such as “The food is fantastic,” and add a specific flair, like “The street food in the city is amazing.” In another application, if a player makes a series of incorrect decisions resulting in a loss in a game, our method can suggest a similar series of steps that would help the player win and adjust their playing style.
- Academic Unit
- Computer Science
- Record Identifier
- 9984454644102771
Metrics
5 File views/ downloads
49 Record Views