The Role of Human Knowledge in Explainable AI

Machine learning and AI are facing a new challenge: making models more explainable.

This means to develop new methodologies to describe the behaviour of widely adopted black-box models, i.e., high-performing models whose internal logic is challenging to describe, justify, and understand from a human perspective.

The final goal of an explainability method is to faithfully describe the behaviour of a (black-box) model to users who can get a better understanding of its logic, thus increasing the trust and acceptance of the system.

Unfortunately, state-of-the-art explainability approaches may not be enough to guarantee the full understandability of explanations from a human perspective. For this reason, human-in-the-loop methods have been widely employed to enhance and/or evaluate explanations of machine learning models. These approaches focus on collecting human knowledge that AI systems can then employ or involving humans to achieve their objectives (e.g., evaluating or improving the system).

Based on these assumptions and requirements, we published a review article that aims to present a literature overview on collecting and employing human knowledge to improve and evaluate the understandability of machine learning models through human-in-the-loop approaches. The paper features a discussion on the challenges, state-of-the-art, and future trends in explainability.

The paper starts from the definition of the notion of “explanation” as an “interface between humans and a decision-maker that is, at the same time, both an accurate proxy of the decision-maker and comprehensible to humans”. Such a description highlights two fundamental features an explanation should have. It must be accurate, i.e., it must faithfully represent the model’s behaviour, and comprehensible, i.e., any human should be able to understand the meaning it conveys.

The Role of Human Knowledge in Explainable AI

The figure above summarizes the four main ways to use human knowledge in explainability, namely: knowledge collection for explainability (red), explainability evaluation (green), understanding human’s perspective in explainability (blue), and improving model explainability (yellow). In the schema, the icons represent human actors.

You may cite the paper as:

Tocchetti, Andrea; Brambilla, Marco. The Role of Human Knowledge in Explainable AI. Data 2022, 7, 93. https://doi.org/10.3390/ data7070093

The VaccinEU dataset of COVID-19 Vaccine Conversations on Twitter in French, German, and Italian

Despite the increasing limitations for unvaccinated people, in many European countries, there is still a non-negligible fraction of individuals who refuse to get vaccinated against SARS-CoV-2, undermining governmental efforts to eradicate the virus.

Within the PERISCOPE project, we studied the role of online social media in influencing individuals’ opinions about getting vaccinated by designing a large-scale collection of Twitter messages in three different languages — French, German, and Italian — and providing public access to the data collected. This work was implemented in collaboration with Observatory on Social Media, Indiana University, Bloomington, USA.

Focusing on the European context, we devised an open dataset called VaccinEU, that aims to help researchers to better understand the impact of online (mis)information about vaccines and design more accurate communication strategies to maximize vaccination coverage.

The dataset is openly accessible in a Dataverse repository and a GitHub repository.

Furthermore, a description has been published in a paper at ICWSM 2022 (open access), which can be cited as:

Di Giovanni, M., Pierri, F., Torres-Lugo, C., & Brambilla, M. (2022). VaccinEU: COVID-19 Vaccine Conversations on Twitter in French, German and Italian. Proceedings of the International AAAI Conference on Web and Social Media16(1), 1236-1244. https://ojs.aaai.org/index.php/ICWSM/article/view/19374

EXP-Crowd: Gamified Crowdsourcing for AI Explainability

The spread of AI and black-box machine learning models makes it necessary to explain their behavior. Consequently, the research field of Explainable AI was born. The main objective of an Explainable AI system is to be understood by a human as the final beneficiary of the model.

In our research we just published on Frontiers in Artificial Intelligence, we frame the explainability problem from the crowd’s point of view and engage both users and AI researchers through a gamified crowdsourcing framework. We research whether it’s possible to improve the crowd’s understanding of black-box models and the quality of the crowdsourced content by engaging users in gamified activities through a crowdsourcing framework called EXP-Crowd. While users engage in such activities, AI researchers organize and share AI- and explainability-related knowledge to educate users.

The next diagram shows the interaction flows of researchers (dashed cyan arrows) and users (orange plain arrows) with the activities devised within our framework. Researchers organize users’ knowledge and set up activities to collect data. As users engage with such activities, they provide Content to researchers. In turn, researchers give the user feedback about the activity they performed. Such feedback aims to improve users’ understanding of the activity itself, the knowledge, and the context provided within it.

Interaction flows of researchers (dashed cyan arrows) and users (orange plain arrows) in the EXP-Crowd framework.

In our recent paper published on Frontiers in Artificial Intelligence, we present the preliminary design of a game with a purpose (G.W.A.P.) to collect features describing real-world entities which can be used for explainability purposes.

One of the crucial steps in the process is the questions and annotation challenge, where Player 1 asks yes/no questions about the entity to be explained. Player 2 answers such questions, and then is asked to complete a series of simple tasks to identify the guessed feature by answering questions and potentially annotating the picture as shown below.

Questioning and annotation steps within the explanation game.

If you are interested in more details, you can read the full EXP-Crowd paper on the journal site (full open access):

You can cite the paper as:

Tocchetti A., Corti L., Brambilla M., and Celino I. (2022). EXP-Crowd: A Gamified Crowdsourcing Framework for Explainability. Frontiers in Artificial Intelligence 5:826499. doi: 10.3389/frai.2022.826499

Analysis of Online Reviews for Evaluating the Quality of Cultural Tourism

Online reviews have long represented a valuable source for data analysis in the tourism field, but these data sources have been mostly studied in terms of the numerical ratings offered by the review platforms.

In a recent article (available as full open-access) and a respective blog post, we explored if social media and online review platforms can be a good source of quantitative evaluation of service quality of cultural venues, such as museums, theaters and so on. Our paper applies automatic analysis of online reviews, by comparing two different automated analysis approaches to evaluate which of the two is more adequate for assessing the quality dimensions. The analysis covers user-generated reviews over the top 100 Italian museums. 

Specifically, we compare two approaches:

  • a ‘top-down’ approach that is based on a supervised classification based upon strategic choices defined by policy makers’ guidelines at the national level; 
  • a ‘bottom-up’ approach that is based on an unsupervised topic model of the online words of reviewers.

The misalignment of the results of the ‘top-down’ strategic studies and ‘bottom-up’ data-driven approaches highlights how data science can offer an important contribution to decision making in cultural tourism.  Both the analysis approaches have been applied to the same dataset of 14,250 Italian reviews.

We identified five quality dimensions that follow the ‘top-down’ perspective: Ticketing and Welcoming, Space, Comfort, Activities, and Communication. Each of these dimensions has been considered as a class in a classification problem over user reviews. The top down approach allowed us to tag each review as descriptive of one of those 5 dimensions. Classification has been implemented both as a machine learning classification problem (using BERT, accuracy 88%) and as and keyword-based tagging (accuracy 80%).

The ‘bottom-up’ approach has been implemented through an unsupervised topic modelling approach, namely LDA (Latent Dirichlet Allocation), implemented and tuned over a range up to 30 topics. The best ‘bottom-up’ model we selected identifies 13 latent dimensions in review texts. We further integrated them in 3 main topics: Museum Cultural Heritage, Personal Experience and Museum Services.

The ‘top-down’ approach (based on a set of keywords defined from the standards issued by the policy maker) resulted in 63% of online reviews that did not fit into any of the predefined quality dimension.

63% of the reviews could not be assessed against the official top-down service quality categories.

The ‘bottom-up’ data-driven approach overcomes this limitation by searching for the aspects of interest using reviewers’ own words. Indeed, usually museum reviews discuss more about a museum’s cultural heritage aspects (46% average probability) and personal experiences (31% average probability) than the services offered by the museum (23% average probability).

Among the various quantitative findings of the study, I think the most important point is that the aspects considered as quality dimensions by the decision maker can be highly different from those aspects perceived as quality dimensions by museum visitors.

You can find out more about this analysis by reading the full article published online as open-access, or this longer blog post . The full reference to the paper is:

Agostino, D.; Brambilla, M.; Pavanetto, S.; Riva, P. The Contribution of Online Reviews for Quality Evaluation of Cultural Tourism Offers: The Experience of Italian Museums. Sustainability 2021, 13, 13340. https://doi.org/10.3390/su132313340

Large-Scale Analysis of On-line Conversation about Vaccines before COVID-19

Frequent words and co-occurrences used by pro-vaccination and anti-vaccination communities.

In this study, we map the Twitter discourse around vaccinations in English along four years, in order to:

  • discover the volumes and trends of the conversation;
  • compare the discussion on Twitter with newspapers’ content; and
  • classify people as pro- or anti- vaccination and explore how their behavior is different.

Datasets. We collected four years of Twitter data (January 2016 – January 2020) about vaccination, before the advent of the Covid-19 pandemic, using three keywords: ’vaccine’, ’vaccination’, and ’immunization’, obtaining around 6.5 MLN tweets. The collection has been analyzed across multiple dimensions and aspects. General

Analysis. The analysis shows that the number of tweets related to the topic in- creased through the years, peaking in 2019. Among others, we identified the 2019 measles outbreak as one of the main reasons for the growth, given the correlation of the tweets volume with CDC (Centers for Disease Control and Prevention) data on measles cases in the United States in 2019 and with the high number of newspaper articles on the topic, which both significantly increased in 2019. Other demographic, space-time, and content analysis have been performed too.

Subjects. Besides the general data analysis, we considered a number of specific topics often addressed within the vaccine conversation, such as the flu vaccine, hpv, polio, and others. We identified the temporal trends and performed specific analysis related to these subjects, also in connection with the respective media coverage.

News Sources. We analyzed the news sources most cited in the tweets, which include Youtube, NaturalNews (which is generally considered as a biased and fake news website) and Facebook. Overall, among the most cited sources, 32% can be labeled as reliable and 25% as conspiracy/fake news sources. Furthermore 32% of the references point to social networks (including Youtube). This analysis shows how social media and non-reliable sources of information frequently drive vaccine-related conversation on Twitter.

User Stance. We applied stance analysis on the authors of the tweets, to determine the user’s orientation toward a given (pre-chosen) target of interest. Our initial content analysis revealed that a large amount of the content is of satirical or derisive nature, causing a number of classification techniques to perform poorly on the dataset. Given that other studies considered the presence of stance-indicative hashtags as an effective way to discover polarized tweets and users, a rule-based classification was applied, based on a selection of 100+ hashtags that allowed to automatically classify a tweet as pro-vaccination or vaccination-skeptic, obtain- ing a total of 250,000+ classified tweets over the 4 years.

Share of pro- and anti- vaccine discourse in time. Pro-vaccine tweet volumes appear to be larger than anti-vaccine tweets and to increase over time.

The words used by the two groups of users to discuss of vaccine-related topics are profoundly different, as are the sources of information they refer to. Anti-vaccine users cited mostly fake news websites and very few reliable sources, which are instead largely cited by pro-vaccine users. Social media (primarily Youtube) represent a large portion of linked content in both cases.

Additionally, we performed demographics (age, gender, ethnicity) and spatial analysis over the two categories of users with the aim of understanding the features of the two communities. Our analysis also shows to which extent the different states are polarized pro or against vaccination in the U.S. on Twitter.

Stance of US states towards vaccination.

A video presenting our research is available on YouTube:

This work has been presented at the IC2S2 conference.

The cover image  by NIAID is licensed under CC BY 2.0.

Generation of Realistic Navigation Paths for Web Site Testing using RNNs and GANs

Weblogs represent the navigation activity generated by a specific amount of users on a given website. This type of data is fundamental because it contains information on the behaviour of users and how they interface with the company’s product itself (website or application). If a company could have a realistic weblog before the release of its product, it would have a significant advantage because it can use the techniques explained above to see the less navigated web pages or those to put in the foreground.

A large audience of users and typically a long time frame are needed to produce sensible and useful log data, making it an expensive task. 

To address this limit, we propose a method that focuses on the generation of REALISTIC NAVIGATIONAL PATHS, i.e., web logs .

Our approach is extremely relevant because it can at the same time tackle the problem of lack of publicly available data about web navigation logs, and also be adopted in industry for AUTOMATIC GENERATION OF REALISTIC TEST SETTINGS of Web sites yet to be deployed.

The generation has been implemented using deep learning methods for generating more realistic navigation activities, namely

  • Recurrent Neural Network, which are very well suited to temporally evolving data
  • Generative Adversarial Network: neural networks aimed at generating new data, such as images or text, very similar to the original ones and sometimes indistinguishable from them, that have become increasingly popular in recent years.

We run experiments using open data sets of weblogs as training, and we run tests for assessing the performance of the methods. Results in generating new weblog data are quite good, as reported in this summary table, with respect to the two evaluation metrics adopted (BLEU and Human evaluation).

Picture1

Comparison of performance of baseline statistical approach, RNN and GAN for generating realistic web logs. Evaluation is done using human assessments and BLEU metrics

 

Our study is described in detail in the paper published at ICWE 2020 – International Conference on Web Engineering with DOI: 10.1007/978-3-030-50578-3. It’s available online on the Springer Web site. and can be cited as:

Pavanetto S., Brambilla M. (2020) Generation of Realistic Navigation Paths for Web Site Testing Using Recurrent Neural Networks and Generative Adversarial Neural Networks. In: Bielikova M., Mikkonen T., Pautasso C. (eds) Web Engineering. ICWE 2020. Lecture Notes in Computer Science, vol 12128. Springer, Cham

The slides are online too:

Together with a short presentation video:

 

Are open source projects governed by rich clubs?

The network of collaborations in an open source project can reveal relevant emergent properties that influence its prospects of success.

In our recent joint work with the Open University of Catalunya / ICREA, we analyze open source projects to determine whether they exhibit a rich-club behavior, that is a phenomenon where contributors with a high number of collaborations (i.e., strongly connected within the collaboration network) are likely to cooperate with other well-connected individuals.

ownCloud-open-source-accessibilityThe presence or absence of a rich-club has an impact on the sustainability and robustness of the project. In fact, if a member of the rich club leaves the project, it is easier for other members of the rich club to take over. Less collaborations would require more effort from more users.

The work has been presented at OpenSym 2019, the 15th International Symposium on Open Collaboration, in Skövde (Sweden), on August 20-22, 2019.

The full paper is available on the conference Web Site (or locally here), and the slides presenting our results are available on Slideshare:

For this analysis, we build and study a dataset with the 100 most popular projects in GitHub, exploiting connectivity patterns in the graph structure of collaborations that arise from commits, issues and pull requests. Results show that rich-club behavior is present in all the projects, but only few of them have an evident club structure.

For instance, this network of contributors for the Materialize project seems to go against the open source paradigma. The project is “owned” by very  few users:

richclubEstablished in 2014 by a team of 4 developers, at the time of the analysis it featured 3,853 commits and 252 contributors. Nevertheless, the project only has two top contributors (with more than 1,000 commits), which belong to the original team, and no other frequent contributors.

For all the projects, we compute coefficients both for single source graphs and the overall interaction graph, showing that rich-club behavior varies across different layers of software development. We provide possible explanations of our results, as well as implications for further analysis.

Brand Community Analysis using Graph Representation Learning on Social Networks – with a Fashion Case

In a world more and more connected, new and complex interaction patterns can be extracted in the communication between people.

This is extremely valuable for brands that can better understand  the interests of users and the trends on social media to better target  their products. In this paper, we aim to analyze the communities that arise around commercial brands on social networks to understand the meaning of similarity, collaboration, and interaction among users.

We exploit the network that builds around the brands by encoding it into a graph model. We build a social network graph, considering user nodes and friendship relations; then we compare it with a heterogeneous graph model, where also posts and hashtags
are considered as nodes and connected to the different node types; we finally build also a reduced network, generated by inducing direct user-to-user connections through the intermediate nodes (posts and hashtags). These different variants are encoded using graph representation learning, which generates a numerical vector for each node. Machine learning techniques are applied to these vectors to extract valuable insights for each user and for the communities they belong to.

We report on our experiments performed on an emerging fashion brand on Instagram, and we show that our approach is able to discriminate potential customers for the brand, and to highlight meaningful sub-communities composed by users that share the same kind of content on social networks.

The use case is taken from a joint research project with the Fashion in Process group in the Design Department of Politecnico di Milano, within the framework of FAST (Fashion Sensing Technology).

This study has been published by Springer as part of ACM SAC 2019, Cyprus.

Here is the slideset presenting the idea:

The paper can be referenced as:

Marco Brambilla, Mattia Gasparini: Brand Community Analysis On Social Networks Using Graph Representation Learning. ACM Symposium on Applied Computing (SAC) 2019, pp. 2060-2069.

The link to the officially published paper in the ACM Library will be available shortly.

Possible Theses in Data Science

Here is a presentation that summarizes some of the relevant topics currently available for theses within the Data Science Lab under my supervision.

Feel free to get in touch in case you are interested.

Data Cleaning for Knowledge Extraction and Understanding on Social Media

 

Social media platforms let users share their opinions through textual or multimedia content. In many settings, this becomes a valuable source of knowledge that can be exploited for specific business objectives. Brands and companies often ask to monitor social media as sources for understanding the stance, opinion, and sentiment of their customers, audience and potential audience. This is crucial for them because it let them understand the trends and future commercial and marketing opportunities.

However, all this relies on a solid and reliable data collection phase, that grants that all the analyses, extractions and predictions are applied on clean, solid and focused data. Indeed, the typical topic-based collection of social media content performed through keyword-based search typically entails very noisy results.

We recently implemented a simple study aiming at cleaning the data collected from social content, within specific domains or related to given topics of interest.  We propose a basic method for data cleaning and removal of off-topic content based on supervised machine learning techniques, i.e. classification, over data collected from social media platforms based on keywords regarding a specific topic. We define a general method for this and then we validate it through an experiment of data extraction from Twitter, with respect to a set of famous cultural institutions in Italy, including theaters, museums, and other venues.

For this case, we collaborated with domain experts to label the dataset, and then we evaluated and compared the performance of classifiers that are trained with different feature extraction strategies.

The work has been presented at the KDWEB workshop at the ICWE 2018 conference.

A preprint of the paper can be downloaded and cited as reported here:

Emre Calisir, Marco Brambilla. The Problem of Data Cleaning for Knowledge Extraction from Social Media. KDWeb Workshop 2018, co-located with ICWE 2018, Caceres, Spain, June 2018.

The slides used in the workshop are available online here: