Organized Event Participant Prediction Enhanced by Social Media Retweeting Data

We find that the event participant prediction can be solved by recommendation techniques. Similar to the user and item interaction in a recommendation problem, event participation can be also treated as the interaction between users and events. After considering several options, we choose the state-of-the-art cold-start recommendation model proposed by Wang et al. [24]. It is a generalization of a neural matrix factorization (NeuMF) model [17] which originally used one-hot representation for users and items.

We aim to use the model to learn the following function:

where $l(u)$ and $r(e)$ are the learned embeddings for user $u$ and event $e$. NeuMF ensembles two recommendation models, called generalized matrix factorization (GMF) and multi-layer perceptron (MLP). Specifically, it makes prediction:

where $\sigma$ is a linear mapping function, and $\bigoplus$ is a concatenation operation.

Since the dataset usually contains only observed interactions, i.e., user purchase records of items, when training the model, it is necessary to bring up some negative samples, for example, by randomly choosing some user-item (user-event) pairs that have no interaction. The loss function for participant prediction is defined as the following:

where $y_{ue}=1$ if user $u$ participated in event $e$, and 0 otherwise. $\mathcal{Y}$ denotes observed interactions and $\mathcal{Y}^{-}$ denotes negative samples.

We have acquired in the previous section joint user embeddings, $l^{J}(u)$, from the entity-connected graph. Note that we can apply the same graph technique to learn embeddings in single domains as well, denoted as $l^{S}(u)$ and $l^{T}(u)$ respectively for the retweeting data and target domain. From problem formulation, we also have base user embedding for the target domain $l^{B}(u)$. A problem is that the graph embeddings $l^{J}(u)$ and $l^{T}(u)$ are only available for a small number of target domain users, because they are learned from limited participation data. When we predict participants in future events, we need to consider the majority of users who have not participated in past events. These users have base embeddings $l^{B}(u)$ but not graph embeddings $l^{J}(u)$ and $l^{T}(u)$.

We need to map base embedding $l^{B}(u)$ to the embedding space of $l^{J}(u)$ when making the prediction. As some previous works proposed, this can be done through linear latent space mapping [25]. Essentially it is to find a transfer matrix $M$ so that $M\times U^{s}_{i}$ approximates $U^{t}_{i}$, and $M$ can be found by solving the following optimization problem

where $\mathcal{L}(.,.)$ is the loss function and $\Omega(M)$ is the regularization. After obtaining $M$ from users who have both base embeddings and graph embeddings, we can map the base user embedding to graph user embedding $l^{J^{\prime}}(u)=M\times l^{B}(u)$ for those users who have no graph embedding.

An alternative solution would be using the base user embedding as the input for training the model. This would then require us to map graph user embedding to target domain base user embedding. Unlike mapping base embedding to graph embedding, where some target domain users have both embeddings, we do not have social media users with base embeddings. So the mapping requires a different technique. We solve it by finding the most similar target domain users for a social media user, and using their embeddings as the social media user base embedding. More specifically, we pick $k$ most similar target domain users according to the graph embedding, and take the average of their base embedding:

where $U^{K}$ is top-k target domain users most similar to the social media user $u$ according to their graph embeddings.

We have shown two ways to create joint training data by mapping graph embeddings to base embeddings, and by mapping base embeddings to graph embeddings. Both embedding spaces have their advantages. The graph embeddings are taken from the interaction data, thus contain information useful for predicting participation. The base embedding contains user context obtained from the target domain, which can supply extra information. While it is possible to use the two types of embeddings separately, we would like to propose a fusion unit that leverages the advantages of both embedding spaces. We call the method base and graph embedding fusion (BGF).

After obtaining training data for two types of embeddings, we train two prediction models separately for them using the NeuMF model. The input event embeddings $r(e)$ are the same for both models. The input user embeddings are selected depending on whether the user has a graph embedding available or not. More specifically, for graph embedding space, the input $l(u)$ is set to $l^{J}(u)$ if user $u$ has graph embedding, and otherwise it is set to the mapped embedding $l^{J^{\prime}}(u)$. Similarly we do for the base embedding space, and select either $l^{B}(u)$ or $l^{B^{\prime}}(u)$ depending on the availability. Then, instead of output predictions, we take the concatenation layers of two NeuMF models, produced by the concatenation in Equation (2) and concatenate them together. The prediction is made on the output of this large concatenation layer.

Following a recent trend in deep learning research, we use an attention module [26] to further refine the output of the model. An attention module is generally effective when we need to choose more important information from inputs. Since after running two prediction models, we have a large number of information units, it is suitable to apply the attention module.

The idea of attention is to use a vector query to assign weights to a matrix so the more important factors can be emphasized. The query is compared with keys, a reference source, to produce a set of weights, which is then used to combine the candidate embeddings. For the current scenario, we use the concatenated output of NeuMF as the key and the event embedding as the query. The output of the attention module is a context vector $c_{i}$ for event $i$

where $a_{ij}$ is attention weights, and $s_{j}$ is the key. We transform the concatenated output of NeuMF into a matrix with the same number of columns as the query dimensions, and use it as the key $s_{j}$. The attention weights can be obtained using the $general$ attention score [27].

We insert the attention module after the output of two prediction models and use the event embedding as the query to select the more important information. Empirically, we do find adding the attention module improves overall prediction accuracy.

We note that BGF can be used with a single domain. We can construct the graph for a single domain without the bridging relations, i.e., only keeping the word co-occurrence and the user participation relations. Using the described above, we can have two sets of embedding generated, from the base embedding and from the graph, and on them the BGF unit can be applied. In the empirical study to be presented later, the single domain BGF is shown to have achieved relatively high prediction accuracy.

We have integrated social media retweeting into the event participation data of a target domain using the method described above. Now we can simply combine the retweeting data with the event participant data, treating them as a single dataset. However, there are better ways to train the model across domains, as proposed by recent studies in transfer learning. Here we will introduce a transfer learning technique that can be used to further improve our method.

The technique is called knowledge distillation (KD) [28]. It has been shown that, when model learning is shifted from one task to another task, this technique can be used to distill knowledge learned in the previous task. The distilled knowledge becomes accessible through the KL-divergence, a measure of the difference between prediction results using the new model and the old model. Specifically, we set up a loss through KL-Divergence:

where $\hat{Y}_{new}$ and $\hat{Y}_{old}$ are predictions made with the model learned in the new domain and the old domain, respectively, and $D_{KL}$ is the point-wise KL-Divergence.

We first train the model using the retweeting data, and then shift to the target domain participation data. The single domain loss and the KD loss can be counted together in the cross-domain model learning, as

We use a publicly available dataset⁴⁴4https://ieee-dataport.org/documents/meetup-dataset. The dataset was collected for the purpose of analyzing information flow in event-based social networks [12]. On Meetup, users can participate in social events, which are only active in a limited period, or they can join groups, which do not have time restriction. Events and users are also associated with tags, which are associated with descriptive English keywords. Popular event examples are language study gatherings, jogging and hiking sessions, and wine tasting workshops. The dataset contains relations between several thousands of users, events, and groups. Our interest is mostly in the user-event relation.

We prepare a corresponding Twitter retweet dataset. We monitor Twitter for tweets authored by users with the keyword “she/her” and “he/him” in their profile description, which results in more than two million tweets. While these tweets covers many topics, they are more or less gender-aware given the author profiles. We construct retweet clusters from these retweets, and obtain several thousands of retweet clusters, each retweeted at least ten times by users in the dataset.

Since our objective is to investigate the effect of adding retweets when the target domain has limited data, we generate datasets of different sizes. Specifically, we select three sizes of datasets, containing 100, 200, and 500 events. To balance the retweets with event data, we use the same number of tweets as the events. The events are randomly selected, and the tweets are also randomly selected with the restriction that their texts have common words with the event descriptions. The number of events, users, participation, tweets, Twitter users, and retweets are shown in Table I.

We use pre-trained embeddings to represent event descriptions and tweets of the same language. Specifically, we use the Spacy⁵⁵5https://spacy.io/ package, which provides word embeddings trained on Web data, and a pipeline to transform sentences into embeddings. For our approach, we also need to provide base user embeddings. For the Meetup dataset, the users are associated with tags, which are associated with text keywords. We again use Spacy to transform user tags to embeddings, and use them as the base user embedding.

We set up two test cases, based on whether or not test data contain events in the training data. In the case where test data contain events in the training data, which is called warm test, we randomly pick up one user from each event, adding it to the test data and removing it from the training data. In the case where test data contain no event in the training data, which is called cold test, we use all data shown in Table I as the training data, and use additional 1,000 events as the test data.

We create the training dataset by random negative sampling. For every interaction entry $(u,e)$ in the training dataset, which is labeled as positive, we randomly pick four users who have not participated in the event, and label the pairs as negative. The testing is done by event. For each event $e$ in the test dataset, we label all users who participated in the event $U^{+}$ as positive. Then, for the purpose of consistent measurement, we pick $n-|U^{+}|$ users, labeled as negative, so that the total candidate is $n$, which is set to 100. For the warm test, $|U^{+}|=1$, while for the cold test, $|U^{+}|$ varies from event to event.

We predict the user preference score for all the $n$ users, rank them by the score, and measure the prediction accuracy based on top $k$ users in the rank. We measure $Recall@10$ and $Precision@5$.

We compare our method with three baselines in the existing literature, in addition to variations of our own approaches. The baselines include:

• •

  base, which runs the recommendation model on target domain base embeddings.

• •

  BGF, the base and graph fusion model we introduced. In this variation, it is used with only the target domain data.

• •

  MIX, a variation of our approach without the knowledge distillation component. Instead, it mixes target domain participation data and the retweets as a single training dataset.

• •

  BPRMF [29], a single domain matrix factorization-based recommendation model, known for its effectiveness in implicit recommendation.

• •

  CKE [30], a knowledge graph-based recommendation model. It can be used for cross-domain prediction if the supporting domain is transformed into a knowledge graph.

• •

  KGAT [31], a state-of-the-art knowledge graph-based recommendation model. It can be used for cross-domain prediction like CKE. However, it does not deal with cold items so we skip it for the cold test.

We implement our approach and all baselines in Python and Tensorflow. We set 200 as the latent factor embedding size where it is needed.

The experimental results are shown in Tables II, respectively. Single-domain methods are indicated by (SD) and cross-domain methods are indicated by (CD). The best results in each test are highlighted in bold font.

First we look at the warm test. We can see that the proposed method has a clear advantage over other methods, achieving the best accuracy for both scenarios and for all training data sizes. Particularly, we see that it steadily outperforms MIX method, validating the effectiveness of knowledge distillation. The second best cross-domain method is KGAT, especially for smaller training data sizes. But its performance deteriorates as training data sizes increase. The best single-domain method, BGF, outperformed cross-domain methods like KGAT and CKE when training data size is large, showing the strength of fusing graph embedding and base embedding together. The proposed method, utilizes BGF and knowledge distillation, outperformed single domain BGF by up to 66%.

Next we look at the cold test. We see that the result is more complex in the cold test. When the training data size is smaller, the proposed method generally shows some advantages. For example, when the training data size is 100, it achieves 2.4% higher Recall@10 than BGF. When the training data size is 500, the base model achieves the best accuracy.

Comparing warm and cold tests, we see that cross-domain methods have an advantage for the former, but a disadvantage for the latter. The reason is that when we already have some participant data for an event, it is easier to use external knowledge to enhance the information. However, when there is no data for a new event, the useful information is mostly from the target domain itself, and retweeting data can only add limited useful information to the model, if not noises, especially when the target domain has sufficient training data.