Senators → Chi-square=9.600, p=0.143, dof=6
			votes_cat     Low  Mid-Low  Mid-High  High
			polarity_cat                              
			Negative        3        1         0     1
			Neutral         0        1         0     1
			Positive        0        1         3     1 
		

			Partylists → Chi-square=3.281, p=0.773, dof=6
			votes_cat     Low  Mid-Low  Mid-High  High
			polarity_cat                              
			Negative        0        1         1     1
			Neutral         1        0         0     1
			Positive        5        4         4     3 
		

			All → Chi-square=11.330, p=0.079, dof=6
			votes_cat     Low  Mid-Low  Mid-High  High
			polarity_cat                              
			Negative        1        1         5     1
			Neutral         1        0         1     2
			Positive        7        7         2     5 
		

Correlation Tests

Conclusion

Pearson correlation coefficient (polarity vs votes): -0.2784966690117171

Spearman correlation coefficient (polarity vs votes): -0.3678864219767173

Which is interpreted as a negative moderate correlation, therefore, we reject the null hypothesis. (Research Gate, 2018)

heatmaps on polarity

emotions are conditional

The heatmap in Figure 5 shows the emotional tones of tweets mentioning senatorial candidates. Neutral is most common, especially for Kiko Pangilinan and Bam Aquino, who also appear with enjoyment. Imee Marcos has fewer mentions, showing neutral along with some anger and contempt. Other candidates are linked with scattered tones, with fewer appearances of the other tones.

tones are relative

Figure 6 presents a heatmap of emotional tones in tweets mentioning partylists. For Duterte Youth, anger and contempt are most common, with some neutral and enjoyment also appearing. Makabayan, ML, Akbayan, BBM, Kabataan, Bayan Muna, Magdalo, PBBM, Patrol, ACT-CIS, and ACT Teachers are mostly associated with neutral and enjoyment tones, showing fewer of the other tones. The remaining mentioned partylists are characterized primarily by neutral tones.

What are the critical proponents that shape the decisions of Filipino voters over support for a candidate?

same words, different emotions

Figure 8 showcases words parsed by the AI model that features a correlation with respect to a mentioned candidate. The word 'bisaya' had differing polarities for each candidate, with arlene brosas obtaining negative polarity and positive polarity for Ben Tulfo. These findings suggest factors relating to the candidate's reputation and past media records, which further impact how people online perceive the individuals.

1 partylist = 1 word

The graph showcases similar context from Figure 8 but for comments featuring mentions of partylists. For figure 9, 'legacy' had negative polarities for duterte youth whilst being positive for mamamayang liberal (ml) partylist. The word 'problema' also had differing polarities for bayan muna and makabayan partylist, having negative and positive polarities respectively. Other polarities had harmonious collerations with respect to the words they are associated with. The graph implies that collective impressions from the media may have been used under different contexts, factors such as word usage as nouns, adverbs, or adjectives, being used to ascertain associated words, reflecting different polarities for each organization.

halalan 2025?!!

Figure 7 forecasts a word cloud in correspondence of the most frequent words mentioned by users. Candidates including kiko pangilinan, bam aquino, and duterte having larger prevalence over most candidates in the graph.

How much did online social media sentiments influence actions
over the elections?

kiko x bam

Figure 10 showcases the count of mentions in perceived judgement towards a candidate. For most representatives, frequency lies at around 20-40, with the exception of kiko pangilinan and bam aquino gaining 100-120 counts of perceived judgements ruled in favor. Imee marcos also gained significant traction of opposition, ranging from 40-60 mentions for netizens. The contrasting favors toward candidates merited higher frequencies, implying a stronger connection to high emotions being a strong driver for engagement; whether positive or negative.

the plot thickens

Figure 11 features perceived judgement for partylists in relation to the amount of mentions based on a user's judgement. Makabayan, duterte youth, and akbayan partylists earned high mentions of judgement towards favor with noticeable oppositions for the pbbm partylist. This graph showed significant contrast towards the sentiments on Figure 9, which could further imply the correlations on negative judgement over favorability.

candidate mentions

(The correlation tests done in the hypothesis testing earlier may also be used
for this research question.)

Are we able to predict the influence of social media on nationwide elections?

Scroll to find out ;)

Now knowing that some categories have a correlation with a senatorial candidates' votes, let's try to create a model for predicting a candidate's votes using the categories from the Sentiment Analysis. Considering that partylists dont have a correlation with a candidate's votes, a model will not be created.

How did it perform?

R2 Score of

0.6376646934020613

RMSE of

2759968.114456545

What did we use?

Alpha of

1.0

Model of

Ridge Regression

How did we make it?

Knowing that some features have a correlation with a senatorial candidate's votes, we used a RidgeRegression model to fit our training data, settting X to all the features, and setting y to be the total votes of each candidate. We split our data into test and training sets to determine the metrics of the model later on. After fitting the model, with alpha an alpha set to 1.0, the model was used to predict the votes of each candidate in the test set, which gave the scores and errors seen on the left.

Ridge regression was chosen for its ability to consider multicollinearity between features, which is more than apparent with our nutshell plot. And in terms of context, this 'makes sense' as Anger would be naturally correlated with a Negative Polarity, and an Infavor judgement would be negatively correlated with an OpposedTo judgement. With this, Ridge Regression was seen to be a better 'fit' for a model.

Prior to picking our model, we had also tried a Lasso Model, as well as tried different alphas for our Ridge Regression, along with feature selection through RFE on sklearn. It was seen that the Ridge Regression with an alpha of 1.0 performed the best out of all the configurations and models that were tried.

Modelling notebook

The model was made and tested in this colab notebook.
Google Colab

Negative Prevails

After fitting the model, the following coefficents are now seen, with Anger having the highest positive coefficient, which closely matches what was seen in the nutshell plot, where Anger had the strongest positive correlation with a candidate's votes. Following that trend, Unsure also had a high positive coefficient. Suprisingly, OpposedTo has a negative coefficient, despite having a strong positive correlation earlier, this may be attributed to the train-test split, where a set of candidates with low OpposedTo values may have had low votes, leading to a negative coefficient. With Polarities, it is still seen that a candidate's average polarity has a negative coefficient, matching what was seen with the correlation in the nutshell plot.

How does it perform?

After the model was fitted, the test sets were fed onto it, and its predicted values from the X test set were compared to the actual y test values.

This gave an R^2 value of ~0.63766469, showing that the model has a moderate fit to the data, indicating that the features used have an effect on a candidate's votes, and 'closely' matches the training data.
It may be seen that there are only 11 test points, attributed to the small set of candidates with social media prescence from the initial scraping done. In the future, a more thorough scrape shall be done to increase the dataset to have a better test for the model.

This also have an RMSE of ~2759968.114456545, showing that there is a large error in the model's prediction in the millions, indicating that more data is still needed to train the model to do a better fit.
This high error may also be attributed to outliers in the data, especially with some candidates dominating with votes. This may also be a cause of the votes being in the millions, causing a similar error scale to be seen.

Implications and What can we do better?

With the model developed, we can perform insights on scraped twitter data with the exact features used to train the model, and use these values to 'predict' the votes of each candidate as a whole from all the tweets scraped or gathered.

As an inherent limitation, this model is only suitable for a sample size of 1000 tweets, with only the specific featured used being the tones, judgements, and polarities extracted, ideally, using the same LLAMA4MAVERICK Model to extract these features. Additionally, the model developed is only fit for senatorial candidates, as the data and features gathered for partylists showed no correlation to their votes,

In the future, more scraping should be scraped, and a diversity of platforms and media types is suggested, expanding upon comments, and performing OCR in Images and Videos to fit the more diverse nature of social media. Additionally, more features should be extracted or reduced to determine more factors that affect a candidate's votes. Especially using specific platforms, demographics, or other fields that affect a candidate's potential votes.

In finality, our data shows that there exists significant correlations over online user sentiment with regards to vote turnouts in the Philippine midterm elections of 2025; having higher positive coefficients based on anger proportional to candidate votes, as based on our model. Neutral tonality was also found to correlate with higher polarity; favorability had interesting outputs despite correlating negative words to certain organizations still merited favorability. Unexpectedly, particular proponents in relation to candidates gave interesting insights over how words had contrasting polarities on varying partylists and candidates.

Amongst the insights derived over the AI model, the researchers recognize the possibility over the limitations of the training process of the data, as the general pool of comments that were gathered were fixed onto a single platform limiting the diversity of inputs.

For further enhancement over the exploration of these implications, we recommend integrating data across multiple social media applications to diversify the input, enhancing the features added to examine each comment or post.

In conclusion, this study examined the role of social media discourse in shaping electoral outcomes during the 2025 Philippine midterm elections. The regression analysis highlighted that certain emotional tones and patterns of online sentiment exert stronger influence on candidate visibility and voter support than others, while some expressions play a weaker or even negative role.

These findings suggest that the dynamics of social media are complex, with sentiment shaping electoral behavior in ways that go beyond simple positive or negative reactions. This reinforces the study’s objective of understanding how emotions and judgments expressed online connect with political outcomes.

The results underscore the growing significance of digital platforms in democratic processes, where public discourse can shape perceptions and ultimately influence electoral futures. Future research should expand data sources and refine analytical approaches to capture the full spectrum of online sentiment, offering deeper insights into how social media continues to transform political engagement.