Exploring and Mitigating Gender Bias in Recommender Systems with Explicit Feedback

Recommender systems influence a significant portion of our digital activity. They are responsible for keeping the user experience afresh by recommending varied items from a catalog of millions of items and also adapt their recommendations according to the personality and taste of the user. Therefore, a sound recommender system may go a long way in improving user experience quality, hence the user retentivity of a digital outlet.

Recommender systems have historically been judged on their accuracyHerlocker et al. (2004); Shani and Gunawardana (2011). When it is concerned with other factors such as novelty, user satisfaction, and diversityHurley and Zhang (2011); Ziegler et al. (2005a); Knijnenburg et al. (2012), the focus continues to be just on the satisfaction of the information needs of the users. Although of immense importance to the relevance of a recommender system, these criteria do not capture the complete picture. In recent years, the public and academic community have scrutinized artificial intelligence systems regarding their fairness. It has been observed that the results generated by various recommender systems reflect the social biases that exist in human stratumEkstrand et al. (2018); Shakespeare et al. (2020); Boratto et al. (2019). Scholars have focused on identifying, quantifying, and mitigating the bias present in the results generated by recommendation systems. Burke (2017) presents a taxonomy of classes for fair recommendation systems. The author suggests different recommendation settings with fairness requirements such as fairness for only users, fairness for only items, and fairness for both users and items. Our work falls into fairness for only items category where bias is shown by a particular set of users against a specific set of items in the dataset. In particular, we are interested in studying and eliminating users’ biasedness against the items associated with a specific gender in recommendation systems.

Bias prevention approaches can be classified according to the phase of the data mining process in which they operate: pre-processing, in-processing, and post-processing methods. Pre-processing methods aim to control distortion of the training set. In particular, they transform the training dataset so that the discriminatory biases contained in the dataset are smoothed, hampering the mining of unfair decision models from the transformed data. In-processing methods modify recommendation algorithms such that the resulting models do not entail unfair decisions by introducing a fairness constraint in the optimization problem. Lastly, post-processing methods act on the extracted data mining model results instead of the training data or algorithm. The method presented in our work is a hybrid of a pre-processing phase and a post-processing phase.

Two prominent studies have focused on gender bias in recommender systems. The work by Shakespeare et al. (2020) establishes the existence of bias in the results of the music recommender systems, and the work by Ekstrand et al. (2018) focuses on bias shown by Collaborative Filtering (CF) algorithms while recommending books written by women authors. Both the studies establish that the CF algorithms produced biased results after being fed the biased data from various socio-cultural factors. While both the works focus just on showing the existence of bias in the presence of the users’ implicit feedback, we also consider the explicit feedback ratings and the bias that may arise out of it. Thus, our model handles the case when the items associated with specific gender might have received worse feedback than they otherwise ought to achieve by a set of users. We go one step further and propose a model to mitigate these biases by quantifying a particular user’s bias and debiasing his or her feedback ratings. We theoretically show that the debiased ratings are unbiased estimators of the true preference of the user. Once the ratings are debiased, they are fed into the recommender algorithms as input to produce recommendations for the desired set of users. Since the recommender system is now fed with the debiased ratings, the resulting recommendations are free from the bias factor and avoid a self-perpetuating loop in the future.

The bias of an individual user reflects his or her taste. However, the KNN based algorithms produce recommendations based on similar characteristics between a set of users and naive implementation of these algorithms reflects the bias of one user in the recommendations produced for the other user. While not directly comparing the rating history of different users or items, Matrix Factorization algorithms rely on deriving latent factors, which depend on the rating history. Both the approaches make the system increasingly biased and homogenized after users interact with their biased recommendations and generate data for the next iteration. The above discussion suggests that though it is necessary to reflect the user’s preference in the recommendations produced for him or her to achieve accuracy, it is equally necessary to prevent the bias of one user from reflecting in the recommendations of another similar user. Our research focuses on this particular objective.

Our debiased ratings assure that the biases of one user do not affect other users; however, it may lead to loss of accuracy because of not reflecting the user’s own preferences. We introduce a new step called preference correction which injects the user’s preference parameter into his/her own debiased recommendation to maintain the accuracy of the system. On the publicly available Book-Crossing dataset (Ziegler et al. (2005b)) and Amazon Book Review dataset (Ni et al. (2019)), we empirically show that this approach retained the significant reduction in bias and had minimal effect on the accuracy of the system. The bias reflected in the recommendations produced by the UserKNN, ItemKNN, ALS, and SVD algorithms is reduced by as much as $42.39\%$, $37.65\%$, $26.51\%$, and $41.43\%$ respectively for the Amazon dataset and by $37.82\%$, $30.73\%$, $24.99\%$, and $32.34\%$ for the Book-Crossing dataset. When measured with respect to Root Mean Squared Error(RMSE), the final accuracy loss in the case of the Amazon dataset comes out to be $7.8\%$, $11.96\%$, $12.49\%$, and $10.38\%$ respectively for the four algorithms. In the case of the Book-Crossing dataset, the RMSE loss comes out to be $13.86\%$, $18.13\%$, $11.41\%$, and $12.89\%$ respectively. In particular, the following are our main contributions.

The problem of gender bias and discrimination has received lots of attention in recent works Hajian et al. (2016). Many proposals like Pedreschi et al. (2008), Pedreschi et al. (2009), Ruggieri et al. (2010), Thanh et al. (2011), Mancuhan and Clifton (2014), Ruggieri et al. (2014) are dedicated to detecting and measuring the existing biases in the datasets while other efforts Kamiran et al. (2010), Kamiran et al. (2012), Hajian and Domingo-Ferrer (2013), Hajian et al. (2014a), Hajian et al. (2014b), Dwork et al. (2011), Zemel et al. (2013) are focused on ensuring that data mining models do not produce discriminatory results even though the input data may be biased. Most of these works focus on the classical problem of classification. Amatriain et al. (2011) discuss the application of various classification methods like Support Vector Machines, Artificial Neural Networks, Bayesian classifiers, and decision trees in recommender systems. Their findings indicated that a more complex classifier need not give a better performance for recommender systems, and more exploration is needed in this direction.

When considering "fairness for only users" according to the taxonomy presented by Burke (2017), Boratto et al. (2019) and Tsintzou et al. (2018) discuss the bias with respect to the preferential recommending of certain items only to the users of a specific gender. While weighted regularization matrix factorization studied in Boratto et al. (2019) is only appropriate for implicit feedback, the Group Utility Loss Minimization proposed in Tsintzou et al. (2018) works only with respect to the UserKNN algorithm. Both the papers address the issue of gender bias by employing post-processing algorithms that work only in limited settings. Though Boratto et al. (2019) and Tsintzou et al. (2018) have addressed the issue of fairness of recommender systems with respect to gender, they have done so from the perspective of recommending certain items only to the users of a specific gender. The difference between their work and our study lies in the fact that we focus on the more direct issue of gender bias in recommendations shown to items associated with a specific gender.

Shakespeare et al. (2020) in their research highlight the artist gender bias in music recommendations produced by Collaborative Filtering algorithms. The work traces the causes of disparity to variations in input gender distributions and user-item preferences, highlighting the effect such configurations can have on user’s gender bias after recommendation generation. Mansoury et al. (2020) discuss the biases from the perspective of a specific group of individuals (for example, a particular gender) receiving less calibrated and hence unfair recommendations. Ekstrand et al. (2018) explores the gender bias present in the book rating dataset. Our work is different from the works by Shakespeare et al. (2020), Mansoury et al. (2020) and Ekstrand et al. (2018) in primarily two factors: (i) we consider explicit feedback as opposed to the implicit feedback, and (ii) we propose a principled approach to debias the ratings and theoretically show that the debiased ratings are unbiased estimators of true ratings.

The research by Leavy et al. (2020) focuses on algorithmic gender bias and proposes a framework whereby language-based data may be systematically evaluated to assess levels of gender bias prevalent in training data for machine learning systems. Our work is different from this study as this study is focused on evaluating gender bias in the language and textual data settings, while ours deals with gender bias in a more traditional user-item rating setting.

Consider a recommender system having $\mathcal{U}=\{1,2,\ldots,U\}$ users and $\mathcal{I}=\{1,2,\ldots,I\}$ items. Let $\mathbb{D}$ and $\mathbb{A}$ denote the set of items associated with disadvantaged group and advantaged group, respectively. For example, in a book recommender system, the books represent the items; $\mathbb{D}$ and $\mathbb{A}$ represent the set of books written by women and men authors respectively. With respect to book recommender system, researchers have already shown that the data is biased against female authors’ books Ekstrand et al. (2018).

Let $r_{ui}\in[1,R]$ denote the rating that user $u$ has given to the item $i$. As opposed to previous works, we consider explicit feedback wherein biases may not only arise from not giving the rating to the item but may also come from giving a bad rating to the item. The user profile $p_{u}=\{X_{u},R_{u}\}$ represents the set of books ($X_{u}$) and the ratings ($R_{u}=\{r_{ui}\}_{i\in X_{u}}$) that user $u$ has given to those items.

The proposed recommender system first pre-processes the data that: 1) finds the log-bias $\theta_{u}$ of each user $u$ and 2) generates the debiased rating $d_{ui}$ of each user $u$ and item $i$ using the computed bias in the first step. We then theoretically show that the debiased ratings generated are unbiased estimators of the true preferences of the user for the items rated by them. Thus, the debiased dataset can then be fed into various recommender algorithms to generate an unbiased predicted rating of a user $u$ for the item $i$, denoted by $\tilde{d}_{ui}$. This debiasing step ensures that the existing biases are not boosted further in the system. Our debiasing model is independent of any recommendation algorithm. We show the performance of our debiasing model on both K-nearest neighbors-based algorithms (UserKNN, ItemKNN) as well as matrix factorization-based algorithms (Alternating Least Square and Singular Value Decomposition) to produce the recommendations.

In the next step, we use preference corrector to reintroduce the preferences of a particular user $u$ to his/her own recommendations. This is achieved via producing a user specific rating $\tilde{r}_{ui}$ from the debiased rating $\tilde{d}_{ui}$. The recommendations are re-ranked according to the adjusted ratings, and the recommendations are presented to the user. This step ensures that the system does not lose accuracy for not considering the preferences of the users. Figure 1 shows the schematic diagram of our model. Consider that the ratings $r_{ui}$ are continuous values ranging from $1$ to $R$, then mathematically, a biased recommender system can be represented as follows:

1. 1.

   Each user $u$, while rating an item $i$, scales down the maximum rating $R$ by $e^{p_{ui}}$. $p_{ui}$ is a random variable, drawn from a distribution function $P_{u}(I)$, which has a mean value of $\alpha_{u}$. $p_{ui}$ represents the logarithm of the true preference of the user $u$ for the item $i$. For the sake of brevity, we call it log-preference of the user $u$ for the item $i$. Hence $e^{p_{ui}}$ is a representation of the true preference of user $u$ for the item $i$.

2. 2.

   In case the item is associated with the disadvantaged group, the user $u$ further scales down the rating of the item by a factor $e^{q_{ui}}$. $q_{ui}$ is a random variable, drawn from a distribution function $Q_{u}(I)$ having a mean value of $\beta_{u}$. $q_{ui}$ represents the logarithm of the biasedness of the user $u$ shown to the item $i$. For the sake of brevity, we call it the log-bias of the user $u$ for the book $i$. Hence $e^{q_{ui}}$ represents the biasedness of the user $u$ for the book $i$.

3. 3.

   For each user $u$, $\beta_{u}$ is sampled from the a distribution function $\Omega(x)$ which governs the global log-bias tendency of the users. We denote the mean value of $\Omega(x)$ by $\gamma$.

Thus, ratings $r_{ui}$ can be expressed as:

We now present a detailed description of each of the step.

To evaluate the proposed model, we run experiments on two publicly available book rating datasets, the Book-Crossing dataset, originally put together by Ziegler et al. (2005b) and the Amazon Book Review dataset, put together by Ni et al. (2019). We further process this dataset through the following stages:

We proposed a model to quantify and mitigate the bias in the explicit feedback given by the users to different items. We theoretically showed that the debiased ratings produced by our model are unbiased estimators of the true preference of the users for the books. With the help of comprehensive experiments on two publically available book datasets, we show a significant reduction in the bias (almost 40%) with just 10% decrease in accuracy using the UserKNN algorithm. Similar trends were observed for other algorithms such as ItemKNN, ALS, and SVD. Our model is independent of these algorithms’ choices and can be applied with any recommendation algorithm. We used book recommender system because we were able to generate the gender information from publicly available APIs. Our model is not restricted to book recommender system as long as protected attribute information about the items is known. We leave extension of the model to missing protected attribute as an interesting future work. It will be an interesting direction to see if the ideas from fair classification literature with missing protected attributes Coston et al. (2019) can be leveraged. We further did not address the bias originating from fewer ratings for a female-authored book than a male-authored one. We leave extending the model to the bias originating from lesser number of ratings and extensively studying the model for other recommender systems as the future directions.