background image

Crafting a Personalized Beer Wheat Icon Landscape

Analyzing User Preferences and Naming Impact for Guiding Targeted Recommendation

Siyuan Cheng

Zhixun Chen

Xinyi Ding

Yihan Wang

people Github link


Basic Statistics
Beer Style Similarities
Beer Popularity and Taste Preference
Beer Name, Beer Flavor
Let's Recommend!


Embark on a journey into the secret world of beer with our project! Dive deep into the trends and tastes of the beer world as we meticulously interpret vast datasets from BeerAdvocate and RateBeer. Our exploration uncovers not just the popularity of beer and the influence of naming on ratings, but also delves into the similarities between styles. Utilizing methods like K-means clustering, sentiment analysis, and word2vec, we analyze over a million beer ratings to comprehensively define beer similarities. From clustering regional preferences to exploring the sentiment and language in beer names, and studying textual similarities between beer styles, our project merges statistical methods with natural language processing techniques to offer a multi-faceted view of the beer domain. Our findings provide guidance for breweries and tailor recommendations for users, considering not just popularity but also quality attributes and regional taste variations. Join us on this fascinating journey of beer exploration!

Basic Statistics about the Beer Data

Our matched dataset, featuring 6,084 breweries and 45,640 beers from BeerAdvocate and RateBeer, includes over 1.2 million ratings. A 20-rating threshold refines this extensive dataset, optimizing processing speed and focusing on frequently reviewed beers. The matched data from both platforms ensures cross-platform consistency, enhancing our model's reliability and quality, making it ideal for developing a beer recommendation system.

Distribution comparison of rating features between two websites

The histogram charts presented appear to compare the distribution of beer ratings between two websites: RateBeer and BeerAdvocate. The data has been cleaned to remove all NaN values and then standardized using min-max scaling to ensure comparability across different rating scales. The histograms are divided into five categories: alcohol by volume (abv), appearance, aroma, palate, and taste. From a visual analysis, we can see that the general trends between RateBeer and BeerAdvocate are similar, suggesting a level of agreement in the ratings of the matched data provided by users on both platforms. The most notable patterns are:

  • Palate: The palate ratings show a more varied distribution, with RateBeer displaying a bimodal distribution, suggesting two common ratings at which beers are evaluated, whereas BeerAdvocate's distribution is more right-skewed.
  • Taste: Taste ratings are crucial, and here both sites show a right-skewed distribution, indicating a tendency towards higher ratings for taste.
  • ABV: The distribution for ABV appears to be left-skewed for both RateBeer and BeerAdvocate, indicating that most beers have a lower alcohol content.
  • Appearance: Both sites show a high frequency of beers with higher appearance ratings, with a sharp peak at the high end of the scale.
  • Aroma: The distributions for aroma ratings are quite similar between the two sites, with a slight right skew, suggesting that users generally rate the aroma of beers positively.

Beer feature and score correlations of the two websites

For RateBeer:

  • There is a moderate to strong positive correlation between taste and aroma, which is the highest correlation observed (0.74).
  • Appearance has a moderate positive correlation with palate (0.44) and taste (0.40).
  • ABV has the lowest correlations with other features, suggesting that the alcohol content does not strongly influence how users rate the other attributes of the beer.

For BeerAdvocate:

  • The correlations are generally higher compared to RateBeer, indicating a stronger relationship between the features on this platform.
  • Taste and aroma again have a strong positive correlation (0.76), which is consistent with RateBeer's ratings.
  • Appearance has a strong correlation with both palate (0.56) and taste (0.52), which is somewhat stronger than on RateBeer.

Overall, this heatmap suggests that certain attributes of beer, such as taste and aroma, are consistently perceived as related by users of both RateBeer and BeerAdvocate. The strength of the correlations varies slightly between the two platforms, but the general pattern remains the same, indicating similar rating behaviors among the users. This information can be particularly useful for brewers and marketers to understand what characteristics of beer are most influential in consumer ratings and perception.

Percentages about beer styles

The pie chart visualizes the popularity of various beer styles within our dataset, which includes 104 different styles. The pie chart lists first twelve beer styles which take up most of percentage in the dataset. "American IPA" is the most popular listed style, making up 12.1% of the dataset. This suggests that it's a prevalent style among beer varieties in your dataset. The next three styles in descending order of popularity are "American Pale Ale (APA)" at 6.77%, "Saison / Farmhouse Ale" at 5.89%, and "American Double / Imperial IPA" at 4.99%. These styles are evidently popular and significantly represented. The smaller slices, such as "Belgian Pale Ale" at 1.88%, indicate less common styles within the dataset

It's important to note that while a smaller percentage does not necessarily mean a style is unpopular globally, it does reflect its frequency within this specific collection of data. The pie chart effectively highlights the diversity within the beer styles and indicates which styles are most commonly represented in our dataset. This can be useful for understanding consumer preferences, production trends, or for making decisions about what beers to feature if the dataset reflects a particular market or beer selection.

Locations of Users and Breweries

There are 61 countries besides US in the dataset. There are 658 users and 4,323 breweries located outside of the United States. This indicates a significantly higher number of breweries compared to the number of users in these regions. It suggests that while there is a substantial brewing industry outside the US, the user base (who might be rating or using the dataset) is smaller in number.

There are 2,683 users and 3,958 breweries located in the United States. The number of users is considerably higher in the US compared to non-US countries. Additionally, the number of breweries in the US is slightly lower than in non-US countries. This could reflect a higher engagement level with the beer rating or dataset usage among US users, or it could indicate a more significant interest in the diverse range of beer styles available in the US.

The data reflects a robust international brewing industry, yet user engagement with this dataset is more pronounced in the US. This higher activity among American users could explain the predominance of American beer styles in the data. Although the number of breweries is comparable in the US and non-US countries, the user distribution suggests a dataset potentially skewed towards a US-centric audience.

Beer Style Similarities

Calculating the "distance" between different beer styles based on their textual descriptions involves using text similarity measures that can quantify how close or different the descriptions are.

We intend to calculate the similarity between beer styles, which are relatively abstract and not indicative of specific flavors. To address this, we have sourced detailed descriptions of beer characteristics from other websites. These descriptions cover the following aspects: Color, Clarity, Perceived Malt Aroma & Flavor, Perceived Hop Aroma & Flavor, Perceived Bitterness, Fermentation Characteristics, Body, Alcohol by Volume (ABV), International Bitterness Units (IBU), and Standard Reference Method (SRM) for color.

The selection of these dimensions to describe beer styles is deliberate, as they represent key aspects of a beer's sensory and chemical properties, which typically influence consumer preferences and choices. Among these, Alcohol_ABV, Bitterness_IBU, and Color_SRM are specific numerical ranges, while the others are textual descriptions. For the numarical ranges, we combine its mean value and its ranges to calculate similarities. For the textual parts, we compute similarity using word2vec to quantify the likeness between descriptors. Finally, we combine the numerical values and textual vectors to represent the flavor profile of a beer style.

Table for a subset beer style similarities

In the dataset, there are 104 beer styles with full explanation for their beer styles. Here we select 5 beer styles to analyze the similarity results

Analysis for a subset beer style similarities

English Brown Ale and American Brown Ale show the highest similarity (0.99), which is consistent with their close characteristics in terms of color (SRM values are overlapping), alcohol content (ABV), and malt flavor profiles, even though the American Brown Ale tends to have a higher bitterness (IBU) and more varied body.

Smoked Beer is markedly different from the other beer styles, as indicated by its lower similarity scores across the board (e.g., 0.55 with English Brown Ale). This is due to its broad range in every category, particularly the color (1-100 SRM) and bitterness (1-120 IBU), and the fact that the other sensory characteristics are variable.

American Pale Lager has moderate similarity with the American IPA (0.87) despite having a much lighter color (2-4 SRM compared to 6-12 SRM) and lower bitterness (5-19 IBU compared to 50-70 IBU). It shares a similar range of alcohol content with the IPAs, which may account for some of the similarity.

American IPA shows high similarity with the American Brown Ale (0.98), likely due to shared hop flavor characteristics and a similar range in alcohol content. However, the American IPA is distinct in its higher bitterness and the specific hop flavors it presents, such as floral, citrus-like, and piney notes, which contribute to its uniqueness.

Beer Popularity and Taste Preference

The beers were clustered based on ABV and several detailed ratings to investigate the similarity patterns among different beers.

The quantitative features 'ABV', 'Appearance', 'Aroma', 'Palate', and 'Taste' are utilized in this section. To mitigate the impact of ABV data's scale and distribution, it is divided into five intervals and values are reassigned. To better identify different types of beer, we aim to classify beers based on the relative relationships across various features. Therefore, we have chosen to explore this using k-means clustering based on cosine distance.
After careful selection of the optimal number of clusters using regression methods and incorporating the clustering results, the beers in our dataset have been divided into five distinct categories. We have analyzed and plotted the cluster centroids to represent the overarching characteristics of all the beers within each cluster. (All ratings have been normalized to a scale ranging from 0 to 1.)

From the radar charts, we can intuitively deduce that the alcohol by volume (ABV) level does not significantly impact the overall evaluation of beer, provided other features are well-balanced. Conversely, a beer's appearance and ABV level are not sufficient to compensate for poor taste and aroma. This suggests that while ABV and appearance contribute to a beer's appeal, taste and aroma are critical factors in determining its overall quality and reception. This observation is in line with the results obtained from ridge regression in each cluster.

By implementing ridge regression on each cluster, we generate a feature importance map based on the coefficients of features. These maps consistently highlight taste and aroma as the most influential factors in customer preferences across all clusters, while the third most significant feature tends to oscillates between palate and appearance. Meanwhile, attributes like ABV level and appearance remain a relatively minor impact, underscoring that taste, followed by aroma and palate, predominantly shapes customer evaluations of beers.

Let's now visualize some examples within each cluster. We select the top five beers with the highest ratings and the bottom five with the lowest ratings in each cluster, and use MDS to project them onto a 2D map. It's noticeable that although the ratings of beers vary widely, their characteristics consistently align with their respective clusters. For instance, red points (beers from the 'High ABV, Visually Pleasing Beer' cluster) all possess relatively high ABV and appearance values, typically accompanied by moderate or lower performance in other aspects. The green points (beers from the 'Mild ABV, Well-Rounded Beer' cluster) uniformly exhibit lower ABV values with well-balanced performance across other features.

Beer Name, Beer Flavor

Do beer names like "Tokyo" influence users' focus on oriental spices like coriander? We investigate how a beer's name, with its emotional tone (ranging from descriptive, like 'Co-op Wheat Beer,' to creative, such as 'Pheasantry Dancing Dragonfly') and language, may influence reviews.

The impact of beer taste on ratings is easy to understand and obvious. But at the same time, we are thinking, are there other hidden factors that will affect the rating?

The impact of different languages of beer on ratings

We first looked at the language of the beer names. Among beer names, English accounts for the largest proportion, reaching about 30%. we take beer names in English, French, German, Italian, Dutch, and Spanish as the main research objects.

We found that the language of a beer’s name had a significant impact on its rating. Among them, compared with other languages, the distribution of German and Dutch beer names can increase the beer score by 0.04, 0.05, at a 5% confidence level.

Interestingly, the two languages are considered highly similar. We speculate that this may be related to German beer culture and reputation. We found similar views in the literature. Barlow et al. (2018) found that knowledgeable consumers rated American lagers lower than other beers, including German and Czech style lagers, even though the taste and flavor attributes of these styles are very similar to those of American lagers.

The impact of sentiment conveyed by beer on ratings

In addition to language, we delve deeper into the linguistic sentiment conveyed by beer. We divided the linguistic sentiment of beer names into five levels from 1 to 5.

We found that, overall, the sentiment of positive beer names had a significant positive impact on their ratings. Each increase in emotional level can increase the beer rating by 0.01.

In subsamples from different users' locations, this positive impact of language sentiment is significant for users in the United States and Canada, but not significant in Europe.

Then, we replaced the explained variable (rating) with beer characteristics: abv, appearance, aroma, palate, and taste. We found that positive beer sentiments have a significant positive impact on appearance, aroma, palate, and taste, especially taste. (There is no doubt that abv is objectively unaffected.) A significant increase in flavor resulted in an increase in ratings. This result is consistent with our analysis in the previous section.

The impact of the sentiment conveyed by beer names in various languages on ratings

Furthermore, if we focus on both language and language sentiment, what will we find?

The intersection of sentiment and language shows that this positive effect of sentiment is enhanced in beers named in French and Spanish. In German and Italian, negative effects are shown.

Let's Recommend!

After a careful analysis on review data from beer rating platforms, we'd love to recommend beers for thirsty beer hunters based on personalized demands and different criteria. Hop on and see if you can find your dream beer with us!

Regional recommendations

Considering the wide geographic spread of both breweries and users, coupled with our discovery of the notable influence of beer name language on ratings, it's apparent that beer preferences can vary significantly across different regions. Therefore, in our recommendation system, we aim to go beyond merely evaluating various beer characteristics. Our goal is to also incorporate the intricate anthropological factors inherent in geographic locations, ensuring our recommendations resonate with the diverse tastes and cultural nuances of beer enthusiasts worldwide.

To achieve this, we first grouped the reviews entries into wider world regions based on user location. Recognizing that different platforms often used varying names for identical beers and styles, we undertook a mapping and normalization process to enable a unified analysis across these platforms. For each region, beers were ranked based on a combined score comprising their average ratings and the volume of reviews. The top three beers and styles for each region are prominently displayed. In our visual representation, larger bubbles signify a greater number of reviews from that region, while a darker hue indicates a higher proportion of reviews for the most popular beer in that region.

From the results, we can see that the top beers and styles in different regions are quite different. For instance, in the United States, Trappistes Rochefort 10 is a familiar favorite, while those traveling to South America might find Falke Tripel Monasterium an appealing local taste to explore. However, it's important to note that the distribution of reviews by location is uneven, which could affect the representativeness of the data for certain regions. A case in point is Africa, where we have only one review entry. Consequently, these findings are reflective of the trends within our specific dataset and should be interpreted with caution, as they might not fully encapsulate broader regional preferences.

Qualitative and similarity recommendations

In addition to the regional recommendation, we also want to provide a more personalized recommendation based on the users' preferred beer quality. We extracted the top 10 descriptive words from previous reviews for each beer and style, and this kind of recommendation is based on containment relationships of this adjective set to the qualititive query. Some examples from our recommendation are displayed below. Furthermore, we enhance user experience by visualizing a word cloud of descriptive terms for each recommended beer. This approach allows users to gain a nuanced understanding of the beer's characteristics before making their choice. Additionally, for those who have enjoyed a particular beer style and wish to explore similar options, our tool facilitates this exploration through an analysis of beer similarities previously conducted. This feature is designed to further personalize and refine the recommendation process.


The dataset, merging BeerAdvocate and RateBeer, includes 6,084 breweries, 45,640 beers, and 1.2 million ratings. Ratings align between platforms, and correlations highlight taste and aroma agreement. American IPA prevails in diverse beer styles, reflecting global brewing diversity with heightened US user engagement.

Analyzing beer style similarities using textual descriptors and numerical ranges, we found that English Brown Ale and American Brown Ale exhibit high similarity, while Smoked Beer stands out as notably different. Specific characteristics contribute to each style's uniqueness.

Then we apply clustering techniques to categorize beers based on key attributes like taste and aroma, etc. By integrating this with regression analysis, we effectively determine beer preferences and the factors driving ratings. This approach enables us to offer targeted recommendations and insights into regional beer tastes.

Continuing, our attention turns to the names of the beers. A noteworthy rise of around 0.05 points is evident for German and Dutch names. Positive emotional sentiments in names notably boost ratings, particularly for US and Canadian users, impacting taste, aroma, and palate. French and Spanish names also show a significant positive impact from emotional sentiment.

Finally, we develop a recommendation system that provides personalized suggestions incorporating regional and qualitative criteria, which is also capable of performing association recommendation based on style similarity. Descriptive word clouds are also visualized for each beer and style to enhance user experience.

Sources & Reference

RateBeer; Beer Adovocate; Craft Beer; Brewers Association;

Barlow, M. A., Verhaal, J. C., & Hoskins, J. D. (2018). Guilty by association: Product-level category stigma and audience expectations in the US craft beer industry. Journal of Management, 44(7), 2934-2960.

Wish you a merry Christmas and a hoppy drink!