Through a Gender Lens: Learning Usage Patterns of Emojis from Large-Scale Android Users

1 INTRODUCTION

• Gender differences matter to user modeling because ignoring them could compromise personalized information services and contribute to inequality in expression and information access.The paper notes that gender is sometimes inferred from user information to improve profiling and personalization, with potential privacy concerns.Source: §1 · p. 1
• Women are more likely than men to use emojis, and the genders differ in emoji preferences and sentiment expression.Some preference differences align with common beliefs, while some sentiment-expression differences are surprisingly different from them.Source: §1 · pp. 1–2
• The study presents the first large-scale analysis of gendered emoji use, addressing a topic not systematically examined in prior emoji research.Earlier work compared emoji use across apps, platforms, and cultures.Source: §1 · p. 1
• 81.1% overall accuracy was achieved when inferring gender from emoji use, comparable to reported English text-based models and generalizable to non-English users.The model uses emojis without textual or other contextual information and is described as more robust to privacy threats than text-based models.Source: §1 · p. 2

2 RELATED WORK

• Research has examined emojis’ general popularity and variation across apps, platforms, countries, and cultures, as well as their communicative functions as content-word replacements, text decoration, tone adjustment, and emotional or situational information.Source: §2 · p. 2
• A prior Twitter study found more females tweeting with emojis than males, but gendered emoji use had not been studied systematically.The authors frame their work as a systematic study of gender-specific emoji use.Source: §2 · p. 2
• Research on gender differences documents variation in facial activity and emoticon use, alongside gendered patterns in topics and language choices.The emoticon findings have been linked to evidence about non-verbal expressivity.Source: §2 · p. 2
• Gender inference has used names, images, interactions, and text, while non-English approaches have been limited and English-based models have not worked well for some languages.The paper uses emojis as a cross-language indicator and compares emoji-based inference with text-based methods.Source: §2 · p. 2

3 THE DATA SET

• The dataset contains 134,419 active users with self-reported gender and 401 million messages collected from December 4, 2016, to February 28, 2017.The data were collected through Kika Keyboard, an Android input-method app supporting 82 languages.Source: §3 · p. 2
• 1,356 emoji types appear in the dataset, and 83.9% of users used emojis at least once.Kika supports emojis released by the Unicode Standard.Source: §3 · p. 3
• The dataset covers 58 languages and 183 countries and regions, with binary self-reported gender labels; 53% of users are female and 47% male.Kika runs at the system level and collects timestamped messages across apps, enabling analysis beyond social-media contexts.Source: §3 · p. 3
• Text content is used only to infer users’ languages and reproduce a state-of-the-art text-based gender-inference model for comparison.The data were anonymized before delivery, and the study received institutional ethics approval and used restricted, HIPPA-compliant cloud processing.Source: §3 · p. 3

4.1 Emoji Popularity

• 7.96% of female users’ messages contain emojis versus 7.02% of male users’ messages (%emoji-msg; p-value≪0.01).Figure 1 also shows female users include emojis in a larger proportion of their messages.Source: §4.1 · p. 3, Figure 1 · p. 3
• 43.9% of female users, versus 29.2% of male users, use emojis in more than 5% of their messages; the pattern holds across all three months.Source: §4.1 · p. 3

4.2 Emoji Preference

• The top-ten lists overlap in eight emojis, but face with tears of joy accounts for 22.1% of female versus 18.9% of male emoji use, making women’s distribution more skewed.Source: §4.2 · p. 3
• The ten most informative emojis are all classified as female emojis under the study’s p(Male|e) > 0.47 threshold.The paper uses mutual information to measure emoji–gender dependence; higher mutual information indicates a more informative emoji for distinguishing genders.Source: §4.2 · p. 4, Table 1 · p. 4
• Co-use communities pair basketball with trophies for men and bathtubs for women; women also co-use shopping bags with clothing-, shoe-, and bag-related emojis.Source: §4.2 · p. 4

4.3 Sentiment Expression

• Women use both positive (50.87% vs 50.25%) and negative emojis (10.11% vs 9.42%) more often than men, with p-value≪0.01 for both comparisons.The study scores emojis using LIWC and Unicode names and annotations, then applies z-tests with Bonferroni correction for multiple tests.Source: §4.3 · p. 5
• Women use face-related emojis more (58.17% vs 56.11%), while men use heart-related emojis more (19.41% vs 17.62%), contrary to cited expectations about real-life love expression.The authors report that some gendered preferences for conveying sentiments through emojis differ from common interpretations.Source: §4.3 · p. 5

5.1 Prediction Set-up

• The binary task uses 39,372 users with at least 100 emoji messages (31,872 train; 7,500 test), measuring accuracy and gender-specific precision against a majority-guess baseline.The two precision metrics count true male or female users among users predicted to be in that gender group.Source: §5.1 · p. 5

5.2 Feature Extraction

• Emoji-frequency features capture average, maximum, and median emojis per emoji message and proportions of emoji-only, single, multiple, consecutive, and repeated patterns.Frequency is one of three feature sets among 1,370 total features.Source: §5.2 · pp. 5–6
• Emoji preference contributes 1,356 per-user emoji proportions, while five sentiment features track emoji and message positivity, negativity, and mixed-valence messages.The model excludes emoji co-usage patterns to avoid overfitting.Source: §5.2 · p. 6

5.3 Model Evaluation

• GBC reaches 0.811 accuracy, 24% above baseline, with balanced gender precision despite class imbalance; all five default models beat baseline on accuracy and both precisions.The test set contains 65.3% female and 34.7% male users.Source: Table 3 · p. 6, §5.3 · p. 6

5.4 Generalizing to Specific Languages

• Across English and 10 selected non-English language tests, GBC beats majority baseline; Italian accuracy is 0.841, 27% above baseline, with balanced precision across languages.The selected non-English languages represent four language families.Source: Table 4 · p. 7, §5.4 · p. 7
• Arabic users show the highest male precision (0.854, 54% above baseline) but the lowest female precision (0.690); cultural self-presentation is offered only as a possible explanation.Source: §5.4 · p. 7

5.5 Discussion

• Accuracy declines slightly as emoji-message counts fall, yet the model beats baseline in every group, including users with only 1–20 emoji messages.That lowest-message group contains 43,309 users.Source: Table 5 · p. 7, §5.5 · p. 7

• The comparison uses 4,156 English users with at least 50 emoji messages, split into 3,306 training and 850 test users, and evaluates released-text, retrained-text, emoji, and semantic-emoji models.The semantic-emoji model uses language-specific emoji sentiment scores derived from nearest-neighbor words; gender inference remains based on emojis.Source: §6 · p. 8
• Released text-model accuracy is 0.800 versus 0.718 for the retrained text model under defaults; the authors suggest its larger training corpus and data-distribution differences may explain the gap.Source: §6 · p. 8
• With default settings on the same data, the emoji model and retrained text model reach similar accuracy (0.736 vs. 0.718); the semantic emoji model performs better.The comparison table reports default settings and tuned results in parentheses.Source: Table 6 · p. 8, §6 · p. 8

• Keyboard developers can make emoji layouts gender-aware and improve recommendations for words or emojis users may type next.The paper also notes that these patterns may apply to emoji input on PCs.Source: §7 · p. 8
• Emoji-only models carry much less risk of accessing or leaking private information than free-text models while sacrificing little inference accuracy, for contexts where gender inference is necessary.Source: §7 · p. 9

8 LIMITATIONS

• Only female and male genders were collected, leaving generalization to other genders unclear and possible self-selection bias.A cited example reports that Facebook offers 71 gender options.Source: §8 · p. 9
• The analysis covers active Kika users, and the findings may not generalize to users of other emoji apps.Source: §8 · p. 9
• Android-only data leave the influence of platform-specific emoji renderings on interpretations and emoji choices untested.Prior work reports that rendering differences across platforms can produce diverse interpretations.Source: §8 · p. 9

9 CONCLUSION

• The data cover 134,419 active users in 183 countries, with 401 million messages in 58 languages collected over three months.Source: §9 · p. 9
• The analysis finds considerable female–male differences across emoji frequency, emoji preference, and sentiment expression.Source: §9 · p. 9
• Emoji-only gender inference is sufficiently accurate to be comparable to models built on free-text features, with much lower privacy risk.Source: §9 · p. 9
• Emoji-based inference is language-independent, and its competitive performance generalizes to non-English users.Source: §9 · p. 9