Exploring Food Waste Conversations on Social Media: A Sentiment, Emotion, and Topic Analysis of Twitter Data

Abstract

Food waste is a complex issue requiring novel approaches to understand and identify areas that could be leveraged for food waste reduction. Data science techniques such as sentiment analysis, emotion analysis, and topic modelling could be used to explore big-picture themes of food waste discussions. This paper aimed to examine food waste discussions on Twitter and identify priority areas for future food waste communication campaigns and interventions. Australian tweets containing food-waste-related search terms were extracted from the Twitter Application Programming Interface from 2019–2021 and analysed using sentiment and emotion engines. Topic modelling was conducted using Latent Dirichlet Allocation. Engagement was calculated as the sum of likes, retweets, replies, and quotes. There were 39,449 tweets collected over three years. Tweets were mostly negative in sentiment and angry in emotion. The topic model identified 13 key topics such as eating to save food waste, morals, economics, and packaging. Engagement was higher for tweets with polarising sentiments and negative emotions. Overall, our interdisciplinary analysis highlighted the negative discourse surrounding food waste discussions and identified priority areas for food waste communication. Data science techniques should be used in the future to monitor public perceptions and understand priority areas for food waste reduction.

1. Introduction

Food waste is an urgent global issue that has detrimental impacts across many different sectors of society. Environmentally, food waste constitutes 8–10% of global greenhouse gas emissions [1]. Economically, food waste has been estimated to cost the global economy USD one-trillion [2]. It is also well known that global food production must substantially increase to feed the growing population [3]. With the rising cost of living and many struggling to access food [4], wasting food is simply not sustainable.

To address food waste and minimise its negative impacts on the climate, the United Nations developed Sustainable Development Goal 12, Responsible Production and Consumption [5]. Goal 12 aims to halve global food waste by 2030 [5]. Australia has committed to achieving this target and has set up a National Food Waste Baseline to consistently monitor food waste across the supply chain [6]. Australia’s food waste landscape is similar to that of other higher-income countries, with most food wasted at the consumer level [7]. To achieve the target of halving food waste within the next ten years, substantial change is required, and novel approaches are required to identify priority areas for food waste communication campaigns and interventions [8].

Social media is an ever-growing communication platform: nearly 60% of the global population comprises active social media users [9], and in some regions, such as the United States of America, usage is as high as 72% of the population (see Supplementary File S1 for a glossary of terms) [10]. People spend 2.5 h daily on social media [9] for many reasons, such as social connection, entertainment, and information gathering [11]. Social media data analysis for research has been a useful way to understand the attention that certain issues get, the public sentiment towards them, and whether there are trends in topic discussions that could inform interested stakeholders [12]. Changes in online conversations can indicate where change is needed in policy [13]. In contrast, these conversations can also be considered when a policy change is made to analyse how the public perceives it [13]. However, current social media research has been small-scale, including methods such as manual content analysis, which typically analyses less than 5000 posts [14] (n = 220 [15], n = 423 [16], and n = 1000 [17]) due to limitations such as the time taken to code the data manually [18,19]. Furthermore, the topics discussed on social media may differ from what a participant would openly share with a researcher in traditional research settings such as focus groups [20]. Hence, social media gives access to rich data, which are often freely available via an Application Programming Interface (API) (see Supplementary File S1 for a glossary of terms) [18]. However, social media (and data gathered from social media) is under-utilised in food waste research, despite being found to have a positive impact on food waste reduction [8]. For the limited number of food waste campaigns and interventions that have used social media, they have focused on traditional engagement metrics such as likes and comments as a measure of success, rather than considering the broader conversations occurring [8].

Current food waste research heavily focuses on the individual and their undesirable behaviours [21], often ignoring bigger-picture themes within society. When taking an ecological approach to consumer behaviour, different systems within society can be examined to understand how outside factors affect individuals [22,23]. In the Socio-Cultural Ecological Systems (SCES) model, often used for understanding behavioural influences, social media falls within the “exosystem” (Figure 1). In the exosystem and on social media platforms, data are now so “big” that we must develop new ways to manage and analyse them. At a broad scale, we need to use tools such as social listening (see Supplementary File S1 for glossary of terms) [24] to understand a big-picture view of food waste conversations. Understanding social media as an indicator of what is happening in the exosystem allows for a deeper understanding of food waste and the bidirectional influences at play. For example, individuals contribute to social media and create user-generated content, but the content on social media also influences individuals. This is crucial to understanding how we can change the food waste landscape.

Natural Language Processing (NLP) is a common tool used for social listening, defined as a branch of artificial intelligence that can use machine learning to help computers understand, interpret, and produce human language (see Supplementary File S1 for a glossary of terms) [25]. NLP analyses social media data based on the language used, rather than engagement alone. Once established, social media analysis techniques such as NLP can automate data collection and analysis [18] to provide a broad picture of the conversation occurring online [20] and provide insight into topics such as food waste, informing key stakeholders of public opinion. One key application of NLP, which helps to understand large amounts of data, is subjectivity analytics (see Supplementary File S1 for a glossary of terms), defined as the task of detecting, extracting, and classifying opinions, sentiments, and attitudes concerning different topics, as expressed in textual input [26]. Sentiment and emotion analysis classifies text into categories such as “positive”, “negative”, or “neutral” and can be performed through various approaches, such as by using a rule-based or lexicon approach [27,28], a machine learning algorithm that is trained to classify sentiment from a test dataset [29], or a mixture of the two methods [30]. Similarly, NLP techniques such as topic modelling (see Supplementary File S1 for a glossary of terms) can be used to explore predominant topics discussed and how they change over time [31]. Topic modelling can also be performed through a variety of approaches, such as non-probabilistic techniques such as Latent Semantic Analysis and Non-Negative Matrix Factorisation and probabilistic techniques such as Latent Dirichlet Allocation and Probabilistic Latent Semantic Analysis [32]. NLP has been used in many disciplines to assess public opinion towards climate change [33,34,35,36], food and nutrition [14], product reviews [37], and reactions to advances in artificial intelligence [38] and monitor various health issues such as Mpox [39], COVID-19 [40,41,42], or public opinion towards vaccination [43,44]. Data science techniques have also been used to explore food supply chain optimisation [45]. However, these techniques are novel when assessing the public’s perceptions of food waste, presenting a unique opportunity to understand the food waste landscape and inform future food waste policy, research, and practice.

Aims

It is currently unclear how large-scale social media data could help to identify priority areas for food waste communication campaigns and interventions, research, policy, and practice. This paper combined the disciplines of data science (D.L.), communication (L.B.), and nutrition (E.L.J., A.M., T.A.M.) to provide a nuanced understanding of food waste social media data. The importance of interdisciplinary teams has been emphasised in prior research, which found that data science and subject-specific experts should work together to ensure the methods are rigorous, accurate, and relevant to the field of interest [14]. Therefore, this exploratory and data-driven research aimed to analyse food waste discussions on Twitter from an interdisciplinary perspective to identify how food waste is discussed in one part of the exosystem and explore how the conversation evolves over time. A secondary aim was to explore engagement in relation to the sentiment, emotion, and topics of the tweets.

2. Materials and Methods

2.1. Procedure

2.1.1. Search Term Development

Twitter was chosen as the platform to explore food waste conversations because it is text-based and has publicly available data through the Application Programming Interface (API). When the first year of focus was defined as 2019, Twitter was among the most-popular social networking platforms [46]. Since 2019, Twitter has struggled to maintain its ranking amongst other platforms, such as TikTok and Instagram [47]. However, it remains one of the global population’s favourite platforms [47]. Twitter has also had a mass exodus of Twitter employees and users from the platform since its sale to Elon Musk [48]. Our data collection had finished in December 2021 before this controversy began.

Developing the search strategy was an iterative process. Test data were collected through the Twitter Advanced Search API by the team’s data scientist (D.L.) between each iteration, with manual checking of 100 randomly sampled tweets from the dataset to assess relevance to food waste. Based on the manual checking, terms were adjusted, added, or removed to improve relevance. Initial search terms were developed by topic expert E.L.J. and discussed with all other authors. E.L.J. explored relevant literature and social media posts to further understand the language describing food waste. Initial terms included “food waste”, “food loss”, “food surplus”, “food spoilage”, “saving food”, “wasted food”, and “save food” with no hashtags. These terms extracted tweets that were 30% relevant to food waste. Final food-waste-related search terms used to extract tweets from the Twitter Advanced Search API were: “food waste”, “food wastage”, “wasted food”, “wasting food”, and “food spoilage”. These phrases were used to avoid bringing up all results related to food or all results related to waste (e.g., general rubbish). When searching for “food waste”, the tweet must include both “food” and “waste”, but the terms did not necessarily have to be next to each other to be collected. Hashtags searched included #foodwaste, #zerofoodwaste, #stopfoodwaste, and #nofoodwaste. There were five iterations before arriving at these final terms, which extracted approximately 80% relevant tweets.

2.1.2. Data Collection

For this exploratory study, tweets from users in Australia were collected to thoroughly explore the immediate food waste environment of the authors. By focusing on Australian tweets only, nuances specific to the Australian food system and the COVID-19 lockdown periods could be examined. User’s location was determined through the location description in their Twitter profile, as the geolocation data were not available for every tweet. There are a range of locations that people can list in their Twitter profile; for example, SYD refers to Sydney, Australia. Therefore, tweets were run through a filter created by the data scientist (D.L.) to determine the tweet location for inclusion in the dataset. The filter assessed the location field in the text. Geolocations outside of Australia were manually identified by D.L. and filtered out. A list of Australian geolocation names and their common acronyms (sourced from the Australian Bureau of Statistics [49]) were used to identify Australian tweets, which were retained for analysis. A flowchart further describing this process is detailed in Supplementary File S2 (Page 1). Other inclusion criteria included: tweets in English; tweets from January through to December in 2019, 2020, and 2021 to obtain an overview of food waste conversation in multiple years; and tweets relevant to the specified search terms. The Twitter Advanced Search API (rather than scraping) was used to retrieve tweets (original, retweets, quotes, and replies) that met the inclusion criteria and included at least one of the search terms. Data were collected by the team’s data scientist (D.L.) retrospectively over a one-month period to comply with the rate limits of the API. There were 500 tweets collected at a time, with a wait time of two seconds between API calls. A flowchart describing this process is detailed in Supplementary File S2 (Page 2). Retweets were included in the data collection to accurately represent the conversation online, whereby some tweets would have a higher “weighting” than others because they were more frequently shared.

Cross-sectional data were extracted by the team’s data scientist (D.L.) from the Twitter API into a data lake in JSON format containing information related to the tweets such as the date of the tweet, the username, the user description, the user location, the date the user was created, the source of the tweet, the number of retweets, replies, likes, and quotes, geolocation data (if available), tweet media (photo, video, Graphics Interchange Format (GIF), text only), and tweet text (Figure 2). These data were extracted from the JSON format into a database using a metadata- and content-extraction engine developed by Monash Data Futures Institute. The extracted data were then used for further analysis.

2.1.3. Data Pre-Processing

Pre-processing for the emotion engine included the team’s data scientist (D.L.) removing the “retweet” label, symbols (@, #), stop words based on the Natural Language Toolkit (NLTK) default stop word list (so that only meaningful words would be analysed for emotion) [50], punctuation, non-American Standard Code for Information Interchange (ASCII) characters, and hyperlinks. The cleaned text was processed through the emotion-prediction engine. Pre-processing for the sentiment engine involved removing stop words (using the NLTK list) and removing punctuation, non-ASCII characters, and hyperlinks before processing through the engine. The same process was used to create a cleaned text for input into the topic model software in Python (Version 3.9) and BigML (https://bigml.com/; accessed 15 September 2022). Data were also processed through an engine developed by Monash Data Futures Institute (https://www.monash.edu/data-futures-institute; accessed 30 July 2023) that detects non-human tweets (i.e., fake tweets from bots). There was no pre-processing required for the bot engine. Bot tweets were not removed, as we wanted the analysis to be reflective of real-time Twitter feeds, which typically include bots. For further details on the analysis types, such as the sentiment analysis, see the glossary in Supplementary File S1. For further details on the pre-processing steps, see Supplementary File S3.

2.2. Data Analysis and Visualisation

2.2.1. Sentiment Analysis

The rule-based Valence Aware Dictionary and Sentiment Reasoner (VADER) engine was used to classify sentiment [27]. VADER is an open-source engine specially developed for social media data that can process slang and emoticons/emojis [27]. Machine learning classifiers often require extensive training data, and accuracy typically depends on the size of the training dataset and whether it represents a wide range of topics [39]. The VADER engine was the most-appropriate for this dataset as it has been extensively used in social media research (construct validity), and the rule-based approach overcomes limitations related to the training dataset size and availability [14]. In the VADER lexicon, each word is assigned a sentiment value ranging from −4 (very negative) to +4 (very positive) based on the average polarity ranking from 10 independent human assessors involved in the engine development [27]. Then for each tweet, the polarity score of each word is combined and normalised to classify an overall sentiment score for the text between −1 and +1, while also considering heuristics such as punctuation, capitalisation, and degree modifiers [27]. VADER has five different sentiments ranging from very negative to very positive (rather than three, which is sometimes used) to ensure sentiment is measured accurately (content validity).

The VADER lexicon was reviewed by the topic expert (E.L.J.) before analysis to ensure that the most-frequently occurring terms in the food waste tweets had an appropriate sentiment in the context of the topic. To accurately represent food waste as a neutral topic and avoid a negative skew of the tweets, the sentiment of the words “waste”, “wastage”, and “wasted” was amended from negative to neutral. This allowed the sentiment of the terms outside of “food waste” to be considered without a predisposition towards a negative sentiment. Throughout the analysis, manual checking (face validity) was conducted to ensure that the sentiment classification reflected the topic expert’s opinion.

2.2.2. Emotion Analysis

Emotion analysis extracts and examines the predominant emotion (e.g., joy, surprise, anger) expressed in a text based on the language used within the text (see Supplementary File S1 for a glossary of terms) [51]. The emotion engines that classify the predominant emotions within a text typically use an algorithm developed from a training dataset [51], which is used to train the machine learning algorithm to predict the desired outcome (i.e., identifying an emotion). The emotion engine used to classify food waste tweets is an open-source engine, “Emotion English DistilRoBERTa-base”, trained with approximately 20,000 data points from diverse datasets [52]. The emotion categories were based on Ekman’s six basic emotions described in depth throughout the literature: anger, fear, disgust, joy, sadness, and surprise, and had an additional neutral category (content validity) [52]. Each food waste tweet was processed through this engine and assigned a predominant emotion. Emojis/emoticons were removed and replaced with a text description (e.g., 😊 replaced with “happy”). Similar to the approach for sentiment, manual checking (face validity) was conducted to “sense-check” the emotion classifications.

2.2.3. Topic Modelling

Topic modelling is an unsupervised machine learning method (see Supplementary File S1 for a glossary of terms), whereby a probabilistic algorithm identifies patterns of word occurrence in a document [31] to produce a set of topics that consist of clusters or words that co-occur in the text [53]. Coherence testing was conducted to determine the optimal number of topics to represent the dataset of food waste tweets. Topic coherence analyses the semantic similarity of the words within a topic and provides an overall coherence score, which determines the ideal number of topics in a dataset due to their meaning, rather than statistical inference alone [54,55]. Coherence measures have been shown to correspond well to human qualitative interpretation of topics [56]. Coherence testing was conducted in Python using the LDA method with the Gensim implementation [54], based on the topic coherence pipeline detailed in the literature [55]. The aim of coherence testing is to create meaningful topics through a data-driven approach combined with expert knowledge. Coherence testing indicated that 13 topics (coherence score = 0.476), 14 topics (coherence score = 0.469), or 7 topics (coherence score = 0.460) were the most-representative of the dataset. These topic models were created and assessed for semantic coherence by the topic expert (E.L.J.) by looking at the topic distribution, keywords, and how similar the topics were to each other (see Supplementary File S4 to view the three models with the highest coherence score). The topic expert (E.L.J.) deemed 13 topics the most-semantically meaningful for the dataset and discussed the decision with the other topic experts (T.A.M., L.B.) to confirm that the 13 topics were the most-appropriate.

The final topic model was formed on the online BigML platform based on expert guidance from the team’s data scientist (D.L.). BigML uses an unsupervised probabilistic Latent Dirichlet Allocation (LDA) (see Supplementary File S1 for a glossary of terms) algorithm to define topics. Due to the interdisciplinary nature of the research team, visualisation of the data was important, and BigML provides a user-friendly interface with good visualisation features. BigML uses the well-established method of LDA topic modelling, which has been widely used in other research [32]. The topic model configuration included specifying 13 topics, bigrams (so food waste could be seen as a term), and filtering HTML key terms and non-language characters. Emojis/emoticons were removed from the text prior to uploading the data to BigML. The topic model produces a “topic probability”, which refers to the average probability of a given topic in an instance of the dataset. A term probability is also included, which refers to the relevance of a certain term (e.g., food waste) to the topic (i.e., a higher probability equates to a higher relevance). After BigML generated the topics, the human interpretation of the topics was important to understand the key theme of the topic and what it means in the context of food waste. The topics were independently named based on their keywords by three topic experts (E.L.J., T.A.M., L.B.), who then collaborated to triangulate the topic names and discuss until a consensus was reached. The topic model batch distribution was then exported to Python. Tweets were assigned a probability value for each of the 13 topics (i.e., for a given tweet, how likely it is to be classified under each of the 13 topics). Further analysis, such as sentiment and emotion classification and engagement analysis, was conducted based on assigning tweets to the topic with the highest probability value.

2.2.4. Sentiment, Emotion, and Topic Evolution

Python library Matplotlib (Version 3.5.1) was used for the data visualisation. For a flowchart detailing the data analysis steps, see Supplementary File S5. In summary, to analyse how the conversation changed across time, the sentiment and emotion of the tweets were graphed over the three-year period. Python libraries pandas (Version 1.4.2) and NumPy (Version 1.21.5) were used to determine the proportion of tweets with each sentiment/emotion in each month, which was then presented as a line graph. The topic evolution was conducted to analyse how the topics of conversation have changed [57]. The BigML topic distribution was used to determine the topic with the highest probability for each tweet, which was then aggregated to present the most-commonly discussed topics for each quarter across the three-year period. To understand the broader context of the topics, news headlines were analysed from the time periods of interest using Wikipedia Portal [58].

As the COVID-19 pandemic became prevalent mid-way through the data-collection period (2020) and affected the food habits of Australians, lockdown periods were visualised on the graphs to help interpret the results. Lockdown dates were sourced from the Australian Bureau of Statistics [59]. Lockdown periods were shaded onto the sentiment and emotion evolution graphs in light grey when any State in Australia was in lockdown for more than three days. This cut-off was chosen to ensure an accurate representation of the climate, i.e., one State in a three-day lockdown is not likely to be representative of the whole of Australia or likely to represent the sentiment/emotion of a whole month of tweets. It is important to note that each State in Australia had different levels of lockdown restrictions (e.g., Melbourne and Victoria enforced 5 km travel restrictions, whereas other States did not). However, the dataset presents an aggregate view of food waste conversations in Australia; therefore, all lockdowns were grouped similarly.

2.2.5. Engagement Analysis

A measure of engagement was created to assess how the public interacted with tweets of varying topics, sentiments, and emotions. The total engagement score was calculated as the sum of likes, replies, retweets, and quotes. When assessing the number of retweets, it should be noted that the same tweet could appear multiple times in the dataset (if it met the inclusion criteria). Each retweet had its own engagement data (e.g., number of likes), but had the same retweet score as the other occurrences because the number of retweets is cumulative.

2.2.6. Statistical Testing

Normality testing was conducted using the Kolmogorov–Smirnov test, histograms, and QQ plots in IBM SPSS statistics (Version 28). The distribution was non-parametric; therefore, the median and 25th and 75th percentiles are reported, and non-parametric statistical tests were used. IBM SPSS Statistics and Python libraries SciPy (Version 1.7.3) and scikit-posthocs (Version 0.7.0) were used to test for statistical differences between years. The Chi-squared test of independence was used for categorical data to assess if there was a significant difference across time. The Kruskal–Wallis test was used to explore differences in continuous data, and if significant (p < 0.05), the post hoc Dunn test was conducted. The Bonferroni correction was applied when conducting post hoc tests related to differences between years (3 comparisons, p < 0.016), sentiment (5 comparisons, significance set at p < 0.010), emotion (7 comparisons, significance set at p < 0.007), and topics (13 comparisons, significance set at p < 0.004).

2.3. Ethics

Ethical approval was granted by the Monash University Human Research Ethics Committee (Project ID: 27376) before the data collection. The data collection adhered to Twitter’s privacy policy, terms of use, and terms of conditions. Identifiable data were stored on a password-protected database that only the authors had access to. All tweets collected in this study were aggregated to summarise key themes and ensure users remained anonymous.

3. Results

3.1. Descriptives

There were 39,449 tweets collected from 17,655 unique users across the three-year period (

Table 1. Descriptive statistics of tweets included in the sample.

Variable	Total	2019	2020	2021	p-Value
N (n (%))	39,449	15,280 (38.73)	12,691 (32.17)	11,478 (29.10)	<0.001 1
Tweet engagement
Total number of likes (n (%))	432,385	84,541 (19.55) a	89,540 (20.71) ab	258,304 (59.74) b	0.003 1
Total number of retweets (n (%))	147,321,472	42,715,607 (28.99) a	58,099,268 (39.44) b	46,506,597 (31.57) c	<0.001 1
Total number of quote tweets (n (%))	12,697	4083 (32.15)	4557 (35.90)	4057 (31.95)	0.170 1
Total number of replies (n (%))	29,403	8528 (29.01) a	10,107 (34.37) b	10,768 (36.62) b	<0.001 1
Tweet source
Twitter Web App (n (%))	8275 (20.98)	1617 (10.58)	3241 (25.54)	3417 (29.77)	N/A 2
Twitter Web Client (n (%))	1840 (4.66)	1698 (11.11)	142 (1.12)	0 (0)
Twitter for iPhone (n (%))	12318 (31.23)	4689 (30.69)	4045 (31.87)	3584 (31.22)
Twitter for Android (n (%))	7056 (17.89)	2564 (16.78)	2245 (17.69)	2247 (19.58)
Twitter for iPad (n (%))	1896 (4.81)	688 (4.50)	625 (4.92)	583 (5.08)
Other (n (%)) c	8064 (20.44)	4024 (26.34)	2393 (18.86)	1647 (14.35)
Tweet format
Text only (n (%))	33,520 (84.97)	12,831 (83.97)	10,882 (85.75)	9807 (85.44)	<0.001 3
Photo (n (%))	5404 (13.7)	2182 (14.28)	1665 (13.12)	1557 (13.57)
Video (n (%))	236 (0.60)	156 (1.02)	30 (0.24)	50 (0.44)
Animated GIF (n (%))	289 (0.73)	111 (0.73)	114 (0.90)	64 (0.56)
Tweet type
Original (n (%))	9632 (24.42)	4414 (28.89)	3014 (23.75)	2204 (19.20)	<0.001 4
Retweet (n (%))	20,361 (51.61)	7575 (49.57)	6610 (52.08)	6176 (53.81)
Quote tweet (n (%))	1961 (4.97)	763 (4.99)	628 (4.95)	570 (4.97)
Reply tweet (n (%))	7495 (19.00)	2528 (16.54)	2439 (19.22)	2528 (22.02)
Tweet authenticity
Number of bot tweets (n (%))	3596 (9.12)	746 (4.88)	644 (5.07)	2206 (19.22)	<0.001 4

GIF: Graphics Interchange Format. ¹ Kruskal–Wallis test (significance at p < 0.05) for differences between years. Values with different superscript letters are significantly different based on results from the post hoc Dunn test with Bonferroni correction. Bonferroni correction for 3 comparisons, significance = p < 0.016. ² Statistical test could not be performed as there were too many categories to run, which violates the assumption of counts less than 5. ³ Other is a combination of many less-common sources that people use to tweet from, e.g., Hootsuite, Instagram, LinkedIn. ⁴ Chi-squared test of independence (significance at p < 0.05) for differences between years.

). The number of tweets was the highest in 2019 and reduced in subsequent years. When considering devices people tweet from, Twitter for iPhone was more commonly used than the Web App or Twitter for Android. Text-only tweets were the most-common across the three-year period, followed by tweets including photos. Videos and animated GIFs were not common, each constituting less than 1% of the collected tweets. Overall, a high proportion of tweets was classified as retweets (approximately 50%), with the proportion increasing year after year. In contrast, the proportion of original tweets constituted 29% of the dataset in 2019 and decreased year on year. The number of bot tweets significantly increased across the three years, representing nearly 20% of tweets by the end of 2021. Engagement metrics varied, with the total number of retweets surpassing 40 million yearly. The total number of likes was particularly high in 2021, whereas the total number of quotes and replies remained consistent across the three years.

3.2. Sentiment

Overall, negative tweets related to food waste were predominant (46%), followed by positive (27%) and neutral (23%;

Table 2. Sentiment of food waste tweets from 2019–2021 using the Valence Aware Dictionary and Sentiment Reasoner (VADER) engine.

Sentiment	Total (n = 39,449)	2019 (n = 15,280)	2020 (n = 12,691)	2021 (n = 11,748)
Very negative (n (%))	797 (2.02)	241 (1.58)	337 (2.66)	219 (1.91)
Negative (n (%))	18,273 (46.32)	7127 (46.64)	5924 (46.68)	5222 (45.50)
Neutral (n (%))	9002 (22.82)	3625 (23.72)	2759 (21.74)	2618 (22.81)
Positive (n (%))	10,574 (26.80)	4037 (26.42)	3425 (26.99)	3112 (27.11)
Very positive (n (%))	803 (2.04)	250 (1.64)	246 (1.94)	307 (2.67)

p value < 0.001 from Chi-squared test of independence (significance at p < 0.05) for differences between years.

). When considering the sentiment categories across the years, the proportion of negative, positive, and neutral tweets remained relatively stable from 2019 to 2021 (Table 2). Extreme ends of the sentiment spectrum also remained low across the three years, averaging 1–2% of the tweets.

Sentiment Evolution by Month

There were substantial variations in sentiment across the months (Figure 3). There was a spike in the proportion of very negative tweets in January 2020, coinciding with a drop in the proportion of negative tweets. The proportion of negative tweets increased in March 2020 to 51.4% (Figure 3). Similarly, August 2020 had a high proportion of negative tweets (54%). The proportion of negative tweets sharply decreased to 35% in February 2021, which coincided with an increase in the proportion of neutral tweets. The proportion of positive tweets was the highest in January 2021 (30.8%) and December 2021 (41%).

3.3. Emotion

Overall, anger was the predominant emotion of food waste tweets (27%), followed by fear (19%) and neutral (17%). Disgust (5%) and surprise (3%) were the least-commonly identified emotions in food waste tweets (

Table 3. Predominant emotion of food waste tweets across years.

Emotion	Total (n = 39,449)	2019 (n = 15,280)	2020 (n = 12,691)	2021 (n = 11,748)
Anger (n (%))	10,617 (26.91)	4043 (26.46)	3809 (30.01)	2765 (24.09)
Disgust (n (%))	2030 (5.15)	828 (5.42)	482 (3.8)	720 (6.27)
Fear (n (%))	7690 (19.49)	3248 (21.26)	2476 (19.51)	1966 (17.13)
Joy (n (%))	6099 (15.46)	2665 (17.44)	1640 (12.92)	1794 (15.63)
Neutral (n (%))	6650 (16.86)	2290 (14.99)	2307 (18.18)	2053 (17.89)
Sadness (n (%))	5277 (13.38)	1848 (12.09)	1624 (12.8)	1805 (15.73)
Surprise (n (%))	1086 (2.75)	358 (2.34)	353 (2.78)	375 (3.27)

p value < 0.001 from Chi-squared test of independence (significance at p < 0.05) for differences between years.

). In 2021, the proportion of angry tweets reduced to below the 2019 proportions, and the proportion of joy tweets increased, but remained below the 2019 levels. Sadness, surprise, disgust, and fear remained relatively constant from 2019–2020 (Table 3).

Emotion Evolution by Month

The proportion of angry tweets about food waste sharply increased at the beginning of March 2020 (Figure 4). This coincided with a decrease in joy tweets (17% to 13%). A similar trend of angry tweets about food waste was observed in the shaded sections of the graph; however, there was a high amount of variability. In 2021, the proportion of angry tweets reduced to below 2019 proportions. The proportion of sad tweets increased from 13% in 2020 to 16% in 2021. Joy tweets sharply increased at the end of 2021, whereas other emotions, such as sadness, fear, and anger, reduced. The proportion of surprise and disgust tweets remained relatively low across the three years. However, there were spikes in the proportion of disgust tweets at various time points, such as June 2019, January 2021, and August 2021.

3.4. Topic Modelling

Modelling identified topics across all SCES levels, from the individual to macro, with most directly related to food waste. Other topics related to themes such as climate change, packaging, and economics, highlighting that many Twitter users recognise the inextricable links between these topics and food waste. The complexity of the food waste environment is evident when seeing the range of relevant SCES levels that topics were allocated to (

Table 4. Results of LDA topic model, with topics listed in order of topic probability.

Topic Name	SCES Level	Topic Probability	8 Key Terms within Topic, in Order of Probability	No. of Tweets a Total (n = 39,449) (n (%))	No. of Tweets a 2019 (n = 15,280) (n (%))	No. of Tweets a 2020 (n = 12,691) (n (%))	No of Tweets a 2021 (n = 11,748) (n (%))
Eating to save food waste	Individual, micro	0.96%	Eat, food, day, wasting food, life, free, live, family	4890 (12.40)	1579 (10.33)	1601 (12.62)	1710 (14.56)
Supply chain—food waste	Meso, exo	0.88%	Food waste, via, fight, chain, supply, share	3970 (10.06)	1713 (11.21)	1202 (9.47)	1055 (8.98)
Food waste—morals	Individual, micro	0.84%	Food, waste, people, waste food, leftovers, handling, f*ck, hate	2925 (7.41)	679 (4.44)	1412 (11.13)	834 (7.1)
Food waste—research	Meso, exo	0.80%	Food, food waste, waste, Australian, research, read, support, innovation	3400 (8.62)	1402 (9.18)	1027 (8.09)	971 (8.27)
Actions to reduce food waste	Individual, micro	0.80%	Reduce, food, help, waste, reduce food, food waste, buy, cut	3240 (8.21)	1443 (9.44)	982 (7.74)	815 (6.94)
Supply chain—production	Meso, exo	0.78%	Production, water, feed, farm, energy, grow, plant, planet	3591 (9.10)	1494 (9.78)	1049 (8.27)	1048 (8.92)
Food waste—governance	Meso, exo	0.77%	Food, waste, start, meal, government, plan, restaurant, supermarkets	2884 (7.31)	896 (5.86)	1061 (8.36)	927 (7.89)
Food waste—management	Meso, exo	0.75%	Waste, food waste, food, landfill, business, environment, learn, management	2921 (7.40)	1122 (7.34)	917 (7.23)	882 (7.51)
Food waste—Australian context	Meso, exo, macro	0.72%	#foodwaste, sustainability, recycling, australia, waste, join, sydney, melbourne	2218 (5.62)	1342 (8.78)	501 (3.95)	375 (3.19)
Food waste—Global context	Meso, exo, macro	0.69%	Food, waste, produce, world, global, billion, million, country	2791 (7.07)	1099 (7.19)	868 (6.84)	824 (7.01)
Food waste—economics	Individual, micro, meso, exo	0.68%	Waste, food, money, stop, change, climate, stop wasting, human	1667 (4.23)	755 (4.94)	530 (4.18)	382 (3.25)
Food waste—disposal	Individual, micro, meso, exo	0.68%	Waste, bin, compost, food, home, week, organic, green	2873 (7.28)	876 (5.73)	832 (6.56)	1165 (9.92)
Food waste—packaging	Individual, micro, meso, exo	0.66%	Plastic, waste, food, packaging, looking, cooking, donate, produce	2079 (5.27)	880 (5.76)	709 (5.59)	490 (4.17)

^a Tweets were assigned a probability value for each of the 13 topics (i.e., for a given tweet, how likely it was to be classified under each of the 13 topics). The number of tweets calculation was based on counting the tweets under their assigned topic with the highest probability value. SCES refers to the Socio-Cultural Ecological Systems model [23].

).

The topics can be visualised by their distance from each other (closer topics are more closely related), their colour (similar colours are related), and their probability (a larger circle represents a larger probability that a given tweet will be about that topic). As shown in Figure 5, there were distinct topic groups that were quite distant from each other. For example, “Food waste—Australian context” was distant from other topics and included keywords such as sustainability, Melbourne, Australia (refer to Table 4 for more details). In contrast, “Food waste—morals” (keywords, e.g., people, waste food, f*ck, hate) and “Food waste—Governance” (keywords, e.g., meal, government, plan) were closely related and more central to the other topics such as “Food waste—global context” (keywords, e.g., waste, produce, global), “Food waste—economics” (keywords, e.g., money, waste, climate), and “Food waste—packaging” (keywords, e.g., plastic, waste, packaging).

“Eating to save food waste” (keywords, e.g., eat, wasting food) was the most-predominant topic in the dataset and had the highest probability score (Table 4). Tweets about this topic discussed themes of overeating to reduce food waste and the annoyance of people who do not eat their whole meal (hence, resulting in food waste).

3.4.1. Topic Evolution by Quarter

The topic with the highest number of assigned tweets overall was “Eating to save food waste”. “Supply chain—food waste”, “Food waste—research”, and “Supply chain—production” were also common (Table 4). These topics did not necessarily correspond with the highest topic probabilities of the 13 topics. In 2019 specifically, “Supply chain—food waste” was the most-commonly assigned topic. In 2020 and 2021, “Eating to save food waste” was the most-commonly assigned topic. “Food waste—economics” and “Food waste—packaging” had the lowest number of tweets assigned to them across the three-year period and had lower probabilities than most topics.

When looking at the most-common topics across quarters, current events appeared to affect what was discussed on Twitter (Figure 6). “Food waste—economics” and “Food waste—Australian context” were the most discussed in Q2 2019 when the Australian federal election was held. In 2020, the start of the COVID-19 pandemic affected the distribution of the topics. “Food waste—morals”, which had assigned tweets with a predominantly negative sentiment and angry emotion, became the most-popular topic of food waste tweets in Q1 2020 (when COVID-19 became commonly talked about in Australia based on news headlines). This was also a time when news headlines indicated that panic buying was extreme, with many people not having access to sufficient food.

In Q4 of 2020 and 2021, “Eating to save food waste” was the most-talked about topic. This trend may have been related to the end-of-year holiday season, where many celebrations involve food. Similarly, “Food waste—research” was also discussed in the end-of-year period (Q4 of 2019 and 2020).

“Food waste—disposal” became a more-common topic in 2021 compared to 2019 and 2020. “Supply chain—production” was highly discussed in Q4 2019 and Q3 2021. “Food waste—governance”, “Food waste—global context”, and “Food waste—management” remained relatively stable across the three years.

3.4.2. Sentiment of Tweets in Assigned Topics

Next, we assessed the sentiments of tweets in the most-commonly assigned topics. Most of the tweets within the 13 topics had a predominantly negative sentiment despite “waste” being changed to a neutral term in the pre-processing stage of the data collection. Topics including particularly negative tweets were “Food waste—global context” (62% negative), which discussed worldwide food waste, “Food waste—economics” (61% negative, 10% very negative), which focused on the extensive cost of food waste to individuals and the wider economy, and “Food waste—morals” (55% negative), which discussed people’s food-waste-related actions (Figure 7). “Food waste—Australian context” referred to localised discussions of food waste within Australia and had more-balanced sentiment proportions, with 37% neutral, 35% negative, and 24% positive tweets. Topics with a higher proportion of positive tweets were “Food waste—packaging” (38% positive), which was related to food packaging and its positive and negative effects, and “Food waste—research” (38% positive, 6% very positive,) which discussed advancements in technology related to food waste reduction.

3.4.3. Emotion of Tweets in Assigned Topics

We also assessed the main emotions of tweets in the assigned topics. Of the 13 topics, seven topics had tweets with a predominant emotion of anger, including “Actions to reduce food waste”, “Food waste—disposal”, “Food waste—economics”, “Food waste—global context”, “Food waste—management”, “Food waste—morals”, and “Supply chain—production” (Figure 8). Other negative emotions were predominant in the tweets alongside anger, including sadness (particularly within “Eating to save food waste” and “Food waste—global context”) and fear (“Food waste—Australian context” and “Food waste—research”). Disgust remained relatively low except for tweets in “Food waste—economics”, which had 24% disgust tweets. “Food waste—packaging” and “Food waste research” had the highest proportion of joy tweets, similar to their predominantly positive sentiments.

3.5. Engagement

When considering engagement, 81% of tweets had some form of engagement (i.e., retweets, likes, replies, quotes), whereas 19% of tweets had no engagement. In terms of total engagement, very negative had significantly higher engagement than all other sentiment categories, with a median engagement score of 11 (p < 0.01;

Table 5. Median (25th and 75th percentiles) engagement of food waste tweets across sentiment, emotion, and assigned topic categories.

	Total	2019	2020	2021
Engagement score 1	3 (1; 20)	3 (1; 11)	3 (1; 25)	4 (1; 39)
Sentiment 2,3
Very negative	11 (1; 331) a	6 (1; 331) a	42 (2; 40,336) a	7 (1; 116) abc
Negative	3 (1; 33) b	3 (1; 20) ab	4 (1; 35) b	4 (1; 48) b
Neutral	3 (1; 12) c	2 (1; 9) c	3 (1; 12) c	3 (1; 21) c
Positive	3 (1; 12) c	2 (1; 7) c	3 (1; 16) d	4 (1; 24) bc
Very positive	4 (1; 140) b	3 (1; 6) bc	4 (1; 140) bd	11 (2; 258) a
Emotion 2,4
Anger	3 (1; 25) a	3 (1; 9) ab	4 (1; 368) a	4 (1; 33) a
Disgust	58 (3; 21,550) b	118 (2; 100,981) c	8 (2; 71) a	116 (8; 27,371) b
Fear	3 (1; 10) c	3 (1; 8) a	3 (1; 10) bd	4 (1; 13) c
Joy	2 (1; 8) d	2 (0; 6) d	2 (1; 5) c	4 (1; 33) a
Neutral	4 (1; 72) a	3 (0; 29) be	7 (1; 140) a	5 (1; 64) a
Sadness	2 (1; 18) c	3 (1; 47) e	2 (1; 10) d	2 (1; 15) c
Surprise	2 (1; 13) cd	2 (0; 8) ad	2 (1; 10) cd	3 (1; 24) ac
Topics 2,5
Actions to reduce food waste	3 (1; 16) ab	3 (1; 32) ab	3 (1; 9) a	3 (1; 19) ab
Eating to save food waste	3 (1; 22) bc	3 (1; 30) a	2 (1; 11) a	3 (1; 27) a
Food waste—Australian context	2 (0; 5) d	1 (0; 4) c	2 (0; 7) a	2 (0; 8) c
Food waste—disposal	3 (1; 14) be	2 (1; 6) de	2 (1; 7) a	7 (1; 258) def
Food waste—economics	11 (1; 40,336) f	198 (1; 100,981) f	9.5 (1; 24,595) bc	4 (1; 22) aghi
Food waste—global context	5 (1; 61) g	8 (1; 177) g	4 (1; 32) bd	6 (1; 35) adg
Food waste—governance	4 (1; 39) ch	3 (1; 39) ahij	4 (1; 39) d	6 (1; 44) dgj
Food waste—management	2 (1; 7) i	2 (0; 5) cd	2 (1; 7) a	3 (1; 9) bch
Food waste—morals	23 (1; 17,917) f	2 (0; 17) bdh	2003 (3; 51,456) e	15 (1; 409) e
Food waste—packaging	2 (0; 591) bej	0 (0; 2) k	20 (1; 591) c	24 (2; 75,787) k
Food waste—research	3 (1; 7) aij	3 (1; 7) bej	3 (1; 8) a	3 (1; 7) bci
Supply chain—food waste	3 (1; 8) ae	2 (1; 6) dj	2 (1; 6) a	4 (1; 58) afj
Supply chain—production	4 (1; 25) h	4 (1; 22) ai	4 (1; 25) d	5 (1; 30) aj

¹ Values presented as the median (25th, 75th percentiles) unless otherwise stated. Total engagement score = sum of number of retweets + likes + quotes + replies. Engagement data rounded up to the nearest whole number. ² p < 0.001 from the Kruskal–Wallis test for differences in total engagement within sentiment categories, emotion categories, and topics. p < 0.001 from the Kruskal–Wallis test for differences in total engagement within sentiment categories, emotion categories, and topics within each year. Values with different superscript letters are significantly different based on the results from the post hoc Dunn test with Bonferroni correction. ³ Bonferroni correction for 5 comparisons, significance = p < 0.010. ⁴ Bonferroni correction for 7 comparisons, significance = p < 0.007. ⁵ Bonferroni correction for 13 comparisons, significance = p < 0.004. Tweets were assigned a probability value for each of the 13 topics with analysis conducted based on assigning tweets to the topic with the highest probability value.

). When the pandemic was officially declared in Australia in 2020, very negative tweets had the highest engagement, corresponding with the increase of tweets in the topic “Food waste—morals”. Disgust elicited the most engagement of the emotion categories, with a median engagement score of 58 compared to other categories, such as neutral, anger, and fear, which had significantly less. Tweets in “Food waste—morals” and “Food waste—economics” had significantly higher engagement than other topics, but were not significantly different from each other. Tweets in “Food waste—morals” had particularly high engagement in 2020 when there was much negativity in the media surrounding the COVID-19 pandemic and panic buying. Tweets assigned to “Food waste—packaging” had low engagement in 2019, but had the second-highest engagement of all topics in 2020 and the highest in 2021. Topics that included tweets with low engagement across the three-year period included “Food waste—Australian context”, “Food waste—management”, and “Food waste—research”.

4. Discussion

4.1. Summary of Findings

To the authors’ knowledge, this study was the first to explore food waste tweets using the data science techniques of subjectivity analytics and topic modelling. The three-year data collection period allowed us to explore how food waste was discussed in one aspect of the exosystem and how the conversation changed over time throughout large-scale events such as the COVID-19 pandemic. The overall sentiment of food waste tweets was predominantly negative, despite “waste” being changed to a neutral term in the sentiment analysis lexicon. Very negative, negative, and very positive tweets were most engaged with, indicating that the extreme ends of the sentiment caused people to react on Twitter. Tweets were mostly classified as “anger”, “fear”, and “neutral” in the emotion analysis. Strong negative emotions such as disgust received the highest engagement. Key topics from the topic modelling (in order of probability) included: “Eating to save food waste”, “Supply chain—food waste”, “Food waste—morals”, and “Food waste—research”. However, an interesting finding was that the most-probable topic identified in the topic modelling, “Eating to save food waste”, did not equate to the most engaged with. The most-engaging topics related to morals, economics, and packaging, all of which were polarising topics in these analyses. Collectively, these findings can inform stakeholders on how food waste is discussed online and the key priority areas to direct future communication efforts in food waste campaigns and interventions.

4.2. The Urgency of Food Waste Reduction

Food waste was discussed negatively on Twitter, highlighting wider public perceptions of the topic. Consumer perceptions and attitudes towards food waste are complex and dependent on many factors, such as their upbringing, social environment, culture, and values [60]. The negative framing of this topic may reflect how urgent food waste reduction is and the necessity to change consumer behaviour to slow climate change. Furthermore, even though Australian tweets were analysed in the current study, the topic “Food waste—global context” was an engaging topic, particularly in 2019. The identification of this topic highlights that the world needs to work together to make positive changes to reduce food waste and halt climate change. While food waste has not been explored in the literature using the methods employed in the current study, climate change research has employed sentiment analysis to explore certain topics. One paper exploring the opinions on #worldenvironmentday found that neutral sentiment was predominant, followed by negative, then positive [61]. They also found that climate change as its own topic was discussed negatively [61]. Furthermore, in a study looking at the happiness of climate change tweets, tweets discussing climate change were less happy than other tweets [35]. These results highlight the negative public opinion about environmental topics such as food waste and climate change and demonstrate the need for urgent political and social action to save the environment.

4.3. The Virality of Negativity

Some (negative) emotions, such as disgust, had extremely high median engagement scores from a large number of retweets. While this reflects the urgency of food waste reduction, it also reflects people “jumping on the bandwagon” and sharing their opinion on a polarising topic. It is well known that outrage tactics can cause engagement as negativity is more contagious than positivity [62]. Protection Motivation Theory details how people are motivated to protect themselves against potential harm by assessing the threat and their ability to cope with it [63]. On social media, people may be exposed to a perceived threat to their own values or beliefs, which motivates them to engage in negative behaviours such as online arguments. Within this echo chamber (see Supplementary File S1 for a glossary of terms) of negativity, the positive side to social media can be lost. One paper found that users on Weibo (similar to Twitter) with negative emotions were more strongly connected than users spreading joy [64]. Simply put, bad news spreads far and wide and connects users more than good news. Our study also showed that food waste tweets with neutral emotions had more engagement than other emotions, such as sadness, joy, and surprise. Similarly, one study found that people prefer climate change messages framed without emotion (i.e., neutral) compared to messages with fear, sadness, and anger [65]. However, this changed based on the perceived persuasiveness, feelings, and impressions of the person sharing the message; in some cases, emotional messages may be more persuasive. This mix of research highlights the complex nature of emotional messaging and behaviour change, creating a challenging environment for food waste communicators.

4.4. The Impact of COVID-19 on the Food Waste Landscape

It is well known that the COVID-19 pandemic significantly affected the food supply chain [66,67] and impacted self-reported food waste in many different countries [68]. One study that surveyed households in Thailand found that food waste increased due to the increased reliance on food delivery services [69]. Some studies in countries such as Italy [70], Iran [71], the USA [72], and Australia [73] found in self-reported surveys that less food was being wasted despite shopping behaviours changing. While these studies have attempted to quantify food waste and provide an indication of food waste behaviour, they do not provide definitive evidence of food waste quantities. Due to the various restrictions during the COVID-19 pandemic, food waste research at this time heavily relied on self-reported food waste via surveys, which is subject to social desirability bias and is typically under-estimated [68,74,75]. Furthermore, research has shown that self-reported food waste is only weakly correlated with measured food waste [76]. Therefore, reports of food waste quantification during the pandemic should be interpreted with caution.

In terms of online behaviour, many people were struggling with their mental health during the COVID-19 pandemic. This led to extreme online negativity, which was evident in this study during the lockdown periods. One topic that generated much engagement and peaked early in the COVID-19 pandemic in Australia was “Food waste—morals”. The keywords within “Food waste—morals” were highly emotional, including the profanity f*ck, hate, and people. This topic focused on the individual and microsystem, blaming others with tweets that included an angry emotion and negative sentiment. In 2020, panic buying was out of control, with Australian supermarkets forced to enforce product restrictions on canned food, meats, rice, pasta, and flour to ensure enough food for everyone [77]. The emergence of this topic indicates that people were angry at others who were over-purchasing food and then letting it go to waste. Within this topic, people were focusing on problem-based language and blaming other people for the issues that food waste causes. However, strong emotions can be important in changing attitudes, beliefs, and behaviours [78]. Environmental research has shown that both positive and negative emotions are drivers of changing pro-environmental behaviour [79]. Therefore, there is the potential that the negativity surrounding the moral aspect of food waste may drive consumers to improve their food-waste-related behaviours in the future.

This negativity related to the morality of food waste was similar to the sentiment and emotion of the topic of “Food waste—economics”. Both topics had high engagement and are inextricably linked: many struggled financially through the pandemic, leading to negative feelings towards others who threw out food. While studies have found that the cost of food waste is a deterrent against consumers wasting food [80,81], the economic impact of food waste is widespread for households, businesses, and the economy as a whole [82], so it is unsurprising that this topic generated outrage. The rising cost of living, unemployment, and inflation during the COVID-19 pandemic also saw many people have limited money to spend on essentials such as food [83]. The emergence of economics in the topic modelling provides evidence to stakeholders that more must be done to assist consumers with the cost of food and further education about how consumers can reduce their costs by reducing food waste is warranted.

4.5. The Positive Aspects of Packaging for Food Waste Reduction

While many negative topics had high engagement, “Food waste—packaging” had the lowest probability of all topics, but had high median engagement compared to other topics. This topic differed from others because the predominant emotions were neutral and joy, and the sentiment was more positive than negative. Packaging in the context of food waste is a complex and inflammatory topic as it is bad for the environment for many reasons, one being its contribution to greenhouse gas production. However, if it saves food (such as by providing more-appropriate portion sizes), then packaging can benefit the environment [60,84]. Consumer perceptions on the issue are complex [85], with some recognising that packaging has positive features and others aiming to completely eradicate packaging, potentially at the demise of the food. Research has shown that consumers perceive packaging (particularly plastic) as harming the environment [86]. Packaging is also seen as a more-serious environmental issue compared to food waste [87], which is why it is surprising that tweets assigned to the packaging topic had a high proportion of positive tweets. The predominantly positive framing of this topic indicates that the tweets may have recognised this connection and be potentially focused on opportunities to use packaging to reduce food waste and benefit the environment. Other research on packaging utilising Twitter data with similar methods as this paper found that packaging is often tied to the themes of the environment, zero waste, and general sustainability [88]. This research did not alter the VADER lexicon as we did in the current study and found mostly neutral and positive tweets about plastic waste [88]. They highlighted that over one-third of the tweets were solution-focused, discussing behaviours with a call to action surrounding reducing plastic waste [88]. Overall, plastic packaging is an up-and-coming research area for food waste management. Our findings indicate that the sentiment around packaging is more positive than negative, suggesting that, as other research has found, the tweets may be solution-based rather than problem-based. This is a promising finding for food industry stakeholders as the benefits of packaging for food waste reduction may be becoming more recognised, leading to greater public awareness and greater potential to reduce food waste.

4.6. The Emergence of Metabolic Food Waste

In contrast to the packaging topic, which had the lowest probability of all topics, “Eating to save food waste” had the highest probability of all topics. The topic was in the individual and microsystem in the SCES model with keywords related to eating and living with family, highlighting that meals are a communal event for many people. Tweets assigned to this topic discussed the desire to eat the whole meal to avoid waste, even if it meant overeating. Overeating to save food waste is known as metabolic food waste. Metabolic food waste is typically a problem in society’s middle and higher-income sectors, where overconsumption is common [89]. Attitudes towards food waste can influence metabolic food waste, i.e., feeling guilty about food waste can lead to overeating without people realising that this is technically wasting food [89]. If people knew that regular overconsumption also wastes the resources that went into food production and puts them at risk of health issues, overconsumption to reduce food waste may be less common [89]. The idea of overconsumption is often ignored in efforts to fight food waste [89] despite being a key contributor to consumer food waste. One observational study investigating metabolic food waste found that the average amount of metabolic food waste of participants classified as overweight was 63 kg per person, and for participants classified as obese, it was 127 kg per year [90]. The existence and popularity of this topic from 2019 to 2021 highlights that further consumer education on metabolic food waste is warranted to reduce consumer food waste. Stakeholders should prioritise reducing overconsumption as an aspect of food waste reduction in the future.

4.7. Strengths and Limitations

A key strength of this study was the utilisation of an interdisciplinary team, including experts from data science, communication, and public health nutrition. Interdisciplinary work is essential when addressing complex issues such as food waste. Having a data scientist on the team is important to assist with developing comprehensive data collection and analysis, and topic experts are important when interpreting and presenting the data. Another strength was using hashtags in conjunction with keywords to collect a dataset as representative of the food conversation as possible. Hashtags have become common for collecting social media data, meaning researchers may miss key information [91]. For example, using hashtags in isolation can lead to missing key data and excluding users who may not be familiar with using hashtags [91]. Therefore, using keywords in addition to hashtags enabled a rich dataset to be collected.

Twitter has substantially changed since our data collection finished, including a recent rebrand to “X”. Social media is ever-changing, and our methods must evolve to suit new platforms and new forms of media. The novelty in our work is not the use of Twitter itself, but rather using data science methods within an interdisciplinary team to gather insights about food waste. With this in mind, some limitations were identified throughout the study. Firstly, Twitter does not collect socio-demographic information about users on their platform, which means that age, gender, or ethnicity could not be analysed in this study. Twitter has been reported to attract users of higher socio-economic status, and in our study, only Australian tweets were analysed. Additionally, collecting data about Twitter users’ location remains challenging. We relied on the users’ profile description to determine whether they were located in Australia because geolocation data were unavailable. Therefore, the location used to classify tweets as Australian might not be accurate and our results are not representative of the general population or all social media platforms. Secondly, when collecting large amounts of data through the API, some tweets are irrelevant to the topic area. To minimise the effect of this, we tested multiple search strategies on the API until the relevancy was up to 80% of the dataset. The relevancy of LDA topic modelling has been questioned when used in short texts [92]. However, in this study, the interdisciplinary nature led us to prioritise visualisation, and therefore, BigML (which uses LDA) was chosen as the tool to conduct topic modelling. Lastly, to our knowledge, there are no publicly available sentiment and emotion engines that are specific to food waste. Therefore, general engines were used, which may limit the ability of the engine to classify sentiment and emotion accurately.

4.8. Future Directions

What this research means for food waste communication campaigns and interventions is not clear-cut as our analysis identified influences across multiple levels of the SCES model. Social media is not a panacea to food waste communication; it is fraught with negativity, trolls, and fake user accounts. However, it can be leveraged to share important messages with those who are receptive to the information. We found that negativity garners more engagement, but this does not necessarily mean that all food waste communication should be problem-focused. The topics that were most engaged with on Twitter in this study provide an overview of the most-important issues to consumers: economics, the morals of food waste, and packaging. Stakeholders can use this information to align their priorities and develop messaging targeting these topics to assist consumers in reducing their food waste or addressing other priority areas. Furthermore, food waste stakeholders should be aware that morals and cost are inflammatory topics that can generate engagement. More research is warranted with consumers to explore the type of messaging that inspires consumers to act, as not all engagement is necessarily positive. It is also important to note that engagement is not the be-all and end-all. Many people use social media without engaging and are happy to scroll by without providing their opinions [93,94]. This means part of the story is missing, and impacts can occur beyond social media engagement; however, more research is needed to explore this further. Overall, food waste communicators must be aware of the negativity bias on social media, but continue to utilise social media to share information and facilitate two-way communication with consumers. Future research could benefit from collecting and analysing global tweets about food waste to allow a deeper insight into the nuances between different regions and how food waste could be addressed globally. Whilst outside the scope of this study, future research could also consider assessing human opinions of the tweets to compare the human sentiment and emotion classification with the machine-based classifications. This would provide insight into the nuances between human opinions and machine-based algorithms and could be used to train the classification models and improve the accuracy of classification. It would be interesting to explore the differences in sentiment, emotion, and topics with sarcastic tweets removed. Furthermore, comparing different topic modelling methods such as LDA with BERTopic or Latent Semantic Analysis could be beneficial in exploring the differences in the topics generated for food waste data.

5. Conclusions

This study explored sentiment, emotion, and topic analysis to understand the wider conversation based on food waste tweets. These analyses enabled us to develop a procedure for searching, listening to, and making meaning of food waste conversations on social media. We identified some key priority areas for food waste campaigns and interventions based on the online discussion: the economic impact of food waste, metabolic food waste, the morality of food waste, and packaging to save food waste. Food waste was negatively discussed with many angry tweets, particularly during the COVID-19 lockdowns. We found that negativity garners the most engagement, but this does not mean that all food waste communication should be problem-focused. “Eating to save food waste” was the most-commonly discussed topic in the dataset. However, this topic was not the most engaged with, highlighting that engagement does not always equal the most common. Smaller topics such as “Food waste—packaging” had the lowest probability overall, but had high engagement in 2020 and 2021. To create collective engagement and action to reduce food waste, there is a need to connect different people talking about the range of areas of food waste. Overall, we found data science techniques such as sentiment analysis, emotion analysis, and topic modelling useful to monitor food waste over time. It would be helpful to set up a data-collection and -analysis pipeline for certain topics to continuously see how the conversation changes over time. Future research should investigate other sustainability topics (e.g., plastics, healthy and sustainable diets) alongside food waste to explore wider linkages between and across systems in the SCES model. Taking an interdisciplinary approach and exploring food waste beyond the individual is essential in the fight against food waste.