Automated news recommendation in front of adversarial examples and the technical limits of transparency in algorithmic accountability

Abstract

Algorithmic decision making is used in an increasing number of fields. Letting automated processes take decisions raises the question of their accountability. In the field of computational journalism, the algorithmic accountability framework proposed by Diakopoulos formalizes this challenge by considering algorithms as objects of human creation, with the goal of revealing the intent embedded into their implementation. A consequence of this definition is that ensuring accountability essentially boils down to a transparency question: given the appropriate reverse-engineering tools, it should be feasible to extract design criteria and to identify intentional biases. General limitations of this transparency ideal have been discussed by Ananny and Crawford (New Media Soc 20(3):973–989, 2018). We further focus on its technical limitations. We show that even if reverse-engineering concludes that the criteria embedded into an algorithm correspond to its publicized intent, it may be that adversarial behaviors make the algorithm deviate from its expected operation. We illustrate this issue with an automated news recommendation system, and show how the classification algorithms used in such systems can be fooled with hard-to-notice modifications of the articles to classify. We therefore suggest that robustness against adversarial behaviors should be taken into account in the definition of algorithmic accountability, to better capture the risks inherent to algorithmic decision making. We finally discuss the various challenges that this new technical limitation raises for journalism practice.

Appendices

Appendix A: details on the classifiers

The multinomial NB classifier was used directly on the histograms made of 20,000 words provided by the bag of words NLP. The only technical tweak was the use of a smoothing factor of 0.09 to deal with words having estimated probability zero.

For the MLP classifier, we selected the best parameters thanks to a grid search set to optimize the classifier’s accuracy, with one to four layers and number of neurons per layer ranging from 10 to 1,000). Several solutions led to similar results and we eventually selected a classifier with 3 layers: the first one uses 141 neurons (corresponding to the square root of the 20,000 words output by the bag of words NLP) with a relu activation function, the last one last one uses 7 neurons (i.e., our number of classes) with a tanh activation layer, and the hidden layer uses 31 neurons (which corresponds to the geometric mean between 141 and 7, i.e., $\sqrt{141\times 7}$) and a relu activation layer. Using more layers did not lead to significant concrete improvements in our case study.

Finally, we used different number and types of layers for the RNN: bidirectional, convolutional (combined with pooling), dense, LSTM (Long Short-Term Memory) with dropout. The one that provided the best results in our context used the next parameters:

Layer (type)	Output shape	# of params
Embedding_2	(None,100,200)	4,000,200
Bidirectional_2	(None,100,200)	180,600
Conv1d_2	(None,98,5000)	3,005,000
Global_max_pooling_1d_2	(None,5000)	0
Dense_3	(None,100)	500,100
Dropout_2	(None,100)	0
Dense_4	(None,7)	707

1. Total params: 7,686,607
2. Trainable params: 3,686,407
3. Non-trainable params: 4,000,200

The implementations of all the machine learning tools that we used are based on the scikit-learn library available at the address: https://scikit-learn.org/stable/. As for the NLP part of the tool, we used the SnowballStemmer library for the words stemming (https://kite.com/python/docs/nltk.SnowballStemmer), in which the stopwords removal can be activated as an option, and we used the Word2Vec models made available by Jean-Philippe Fauconnier for the words embedding (http://fauconnier.github.io/). We report the different learning curves of the three combinations of NLP and ML tools in the Fig. 4 As can be observed, the amount of profiling data (i.e., 4000 given that we estimate the accuracy with fivefold cross-validation) is sufficient for the NB and MLP classifiers to approach convergence, which confirms the amount of collected data is sufficient for those classifiers to provide meaningful outcomes. The RNN shows slightly worse results in our case, which may be due both to the simple (topic) feature that we aim to capture and to a lack of data for such a more data-demanding machine learning algorithm.

Appendix B: additional figures

See the Figs.

Fig. 4

Learning curves of the experimented NLP + ML tools

4 and Fig.

Fig. 5

MLP classifier: minimum article perturbation in function of the class distance

5.