A social network user depression symptom multivariate time series feature extraction method

Abstract

The application relates to the technical field of depression detection, and specifically discloses a social network user depression symptom multivariate time sequence feature extraction method, which comprises the following steps: firstly, a data set is constructed for a network platform; then, according to the depression language features of users on the social network, multiple depression symptom text descriptions are given; subsequently, an algorithm is proposed to calculate the similarity between each tweet of a user and the descriptive text of the depression symptom; the values are arranged in time sequence to obtain the user depression symptom multivariate time sequence feature; finally, the effectiveness, interpretability and contribution of each feature to classification of the extracted feature are verified through several classic multivariate time sequence classifiers. The method can effectively detect depression users on the network platform, well depict the dynamic change process of the depression state of the user, effectively mine the hidden features of the user and is easy to visualize.

Description

Technical Field

[0001] This invention relates to the field of depression detection technology, specifically a method for extracting multivariate time-series features of depressive symptoms from social network users. Background Technology

[0002] Depression, also known as melancholia, is a mental illness characterized primarily by prolonged low mood and loss of pleasure. A recent study indicates that over 350 million people worldwide suffer from depression. Furthermore, major depressive disorder can lead to suicide, and the World Health Organization has listed it as one of the leading causes of suicide. Despite the availability of many effective treatments for depression, over 70% of early-stage depression patients fail to seek timely medical attention and treatment due to shame or a lack of awareness of their condition.

[0003] More and more people are using online social networks to record their lives and feelings. This content not only reflects users' daily lives but also their mental state. This makes it possible to detect depression through online social networks. Several studies have already utilized online social networks for depression detection. Most of these studies extract features such as language and emotion, images, behavior, and demographic characteristics for depression detection. However, these features are not temporally structured and cannot reflect changes in an individual's mental state. Furthermore, the mental state of people with depression may constantly shift between depressive and normal states. Such shifts can be reflected in tweets posted on social networks. Therefore, dynamically detecting depression through each tweet a user posts may provide new insights into understanding the state of people with depression on online social networks.

[0004] However, current research faces three main challenges: First, time-series-based methods for detecting depression on social networks require the collection of users' tweet history as completely as possible, while maintaining temporal continuity. However, there is currently no open-source online social network dataset for detecting depression. Second, tweets on online social networks are typically short and use casual language, significantly increasing the difficulty of identifying expressions suggestive of depression within a single tweet. Third, the characteristics of depressed individuals on online social networks are diverse and often differ considerably from those observed in clinical medicine. Therefore, establishing a link between depressive characteristics on online social networks and clinical depressive symptoms remains a challenge.

[0005] In recent years, research on online social network depression detection has made significant progress. A review and summary of these studies reveals that the features used can be mainly categorized into three types: textual features, statistical features, and image features.

[0006] Existing work has confirmed that depression can significantly impact a person's language abilities, and depressed individuals typically express more negative emotions in everyday conversations. Therefore, a large body of research focuses on extracting linguistic and emotional features from tweet text. Recently, Trotzek et al. demonstrated that word vectors and user-level language metadata features can effectively distinguish between depressed and non-depressed users; this study achieved state-of-the-art performance in early depression detection tasks at the time. Ricard et al. used ANEW and LabMT to extract linguistic and emotional features from user tweet comments, finding that this community-generated data can effectively supplement user-generated data. Another commonly used feature type is statistical features. These features are often used in conjunction with textual features. They mainly include demographic features and user behavior features, such as age, gender, posting frequency in the early morning, and the number of mentions in tweets.

[0007] Beyond textual and statistical features, recent research has found that image information in user tweets is also important for depression detection tasks. Reece et al. conducted color analysis, metadata analysis, and face detection on nearly 44,000 Instagram photos, finding that images in user tweets can indeed serve as a valid indication of whether a person suffers from depression. Since then, multimodal methods that synergistically utilize textual and image data from user tweets have emerged. Gui et al. found that tweets on online social networks are often short and lack contextual information, making it difficult to determine the true emotional polarity of a tweet using only its textual content. This problem can often be solved by combining textual and image information.

[0008] In summary, the aforementioned features are effective for detecting depression in online social networks because they reflect a user's mental state over a period of time. However, depressed individuals are not constantly in a depressed state, so a single, comprehensive observation of a user's partial tweet history is not representative.

[0009] Recent studies have used time series or time-dependent variables to understand the dynamic process of depression in online social networks. Reece et al. used state-space timeline analysis of a hidden Markov model to track the onset and recovery of depression in users. Their study found that the model typically predicted a lower probability of depression in users who received treatment, demonstrating the effectiveness of the treatment. Rissola et al. conducted a timeline analysis of tweets from some depressed users, showing significant differences even among depressed users in the time course of depressive mood changes, indicating substantial differences in their behavioral manifestations. Furthermore, the work of Seabrook et al. confirmed that the fluctuating characteristics of daily emotional expression play an important role in screening for depression in online social networks, although these characteristics may be highly dependent on the social network platform.

[0010] These findings suggest that it is possible to detect depression using time-dependent characteristics of users' mental states. Furthermore, dynamic analysis of the mental states of depressed users can deepen our understanding of the behavior of depressed users on online social networks. Nevertheless, there is currently no work that utilizes time-series features for depression detection on online social networks. Summary of the Invention

[0011] To address the aforementioned problems, the present invention aims to provide a method for extracting multivariate time-series features of depressive symptoms from social network users. This method maps user tweets into multivariate time series data. Each tweet is mapped to several numerical points by calculating the textual similarity between the tweet and multiple descriptive texts about depressive symptoms. This method not only effectively uncovers latent user features but is also easily visualized. The technical solution is as follows:

[0012] A method for extracting multivariate time-series features of depressive symptoms in social network users includes the following steps:

[0013] Step 1: Build a well-labeled dataset using the collected network platform data and open-source it;

[0014] Step 2: Combining the keywords of depressive symptoms summarized during the annotation process and the characteristics of tweets on online platforms, a set of descriptive texts of depressive symptoms was proposed;

[0015] Step 3: Calculate the similarity value between each of the user's tweets and the descriptive text of depressive symptoms, i.e. the depressive symptom value of the tweet, and arrange the depressive symptom values ​​in chronological order to obtain the multivariate time series features of depressive symptoms that reflect the dynamic changes of the user's depressive symptoms;

[0016] Step 4: Validate the effectiveness, interpretability, and contribution of each feature to the classification by using a multivariate time series classifier.

[0017] Furthermore, the specific steps for constructing the dataset in step 1 are as follows:

[0018] Step 1.1: Use web crawlers to collect user IDs (identifiers) from several online platforms, and perform preliminary screening by setting collection keywords related to depression and collection time periods. Then, build another web crawler based on the APIs (Application Programming Interfaces) of the online platform developers to collect user resumes and tweets, thereby constructing a candidate dataset of depressed users and a candidate dataset of non-depressed users.

[0019] Step 1.2: Annotate the candidate depressed user dataset and the candidate non-depressed user dataset using an annotation system that supports multi-user annotation, to obtain the depressed user dataset and the non-depressed user dataset.

[0020] Furthermore, step 1.1 specifically includes: constructing a candidate depressed user dataset: first, collecting a large number of tweets under the topic of "depression" on the network platform; then, identifying the most frequent and distinctive depression words in the text as seed search terms; next, using these seed search terms to collect tweets within a specific time period, and parsing candidate depressed user IDs from these tweets; finally, collecting the resumes and tweets of these users to obtain several candidate depressed users and their tweets.

[0021] Construct a candidate non-depressed user dataset: First, collect user IDs under the "Daily" topic on the network platform, then collect the resumes of these users and all tweets within the specified time period to obtain a number of candidate non-depressed users and their tweets.

[0022] Furthermore, the depressive symptom text matrix composed of descriptive texts of depressive symptoms in step 2 is denoted as... T s :

[0023]

[0024] in, S This refers to descriptive text representing the symptoms of depression. s i This refers to descriptive text representing a specific depressive symptom. m This indicates the number of depressive symptoms.

[0025] Furthermore, step 3, calculating the similarity value between each user's tweet and the descriptive text of depressive symptoms, specifically includes:

[0026] Step 3.1: Extract the embedding of each depressive symptom text and construct the depressive symptom text matrix. T s Transformed into a depressive symptom embedding matrix Es :

[0027]

[0028] in, Text indicating depressive symptoms s i Embedded, f This represents a pre-trained sentence embedding model;

[0029] Step 3.2: For each tweet text from a user t Similarly, a pre-trained sentence embedding model is used. f Extract its text embedding t e ;

[0030] Step 3.3: Utilizing the embedding matrix of depressive symptoms E s Calculate the depression symptom score of the tweet S d :

[0031]

[0032] in, θ i It is the tweet text. t Depression symptom scores, f s This represents the function for calculating text similarity.

[0033] Furthermore, the text similarity calculation function is a cosine similarity function, denoted as: cosine-sim ;

[0034]

[0035] in, u , v These represent the user's tweet embedding vector and the depressive symptom text embedding vector, respectively. u i ,v i These are the discrete values ​​in these two vectors. n Indicates dimension.

[0036] Furthermore, the multivariate time-series features of the depressive symptoms MTS d Represented as:

[0037]

[0038] in, m Indicates the number of depressive symptoms. n This indicates the number of tweets by the user. It represents the set of natural numbers.

[0039] The beneficial effects of this invention are:

[0040] 1. This invention constructs a well-labeled dataset for detecting depression. To construct this dataset, embodiments of this invention first collected 64,513 candidate users with depression and 31,085 candidate users without depression using a well-designed data collection strategy. Then, based on the Diagnostic and Statistical Manual of Mental Illnesses, Fifth Edition (DSM-5) and some recent research in the medical field, a labeling rule for depressed users on online social networks was designed. After rigorous labeling, 3,711 users with depression and 19,526 users without depression were finally obtained, and their tweets were collected as completely as possible.

[0041] 2. This invention proposes a method for extracting multivariate time series features of depressive symptoms from social network users. This method maps user tweets into multivariate time series. Each tweet is mapped to several numerical points by calculating the textual similarity between the tweet and multiple descriptive texts of depressive symptoms. Experiments show that this feature extraction method can not only effectively uncover users' latent features but is also easy to visualize.

[0042] 3. This invention reveals the clinical medical characteristics of features and their importance in online depression detection. By visualizing and comparing each dimension of multivariate time-series features of depressed and non-depressed users, this paper finds that depressed users experience more frequent and severe fluctuations in depressive symptoms. Furthermore, the visualized features can also identify the depressive phases of depressed users. In addition, experimental results on the contribution of features to classification show that the importance of different depressive symptoms in online depression detection differs significantly from their importance in clinical medical diagnosis. Attached Figure Description

[0043] Figure 1 A schematic diagram illustrating a method for extracting features of depressive symptoms from online social network users.

[0044] Figure 2The following are the characteristics contributing to the overall score: Sad (Sadness), Int (Interest loss), App (Appetite loss), Ins (Insomnia), Agi (Agitation), Ene (Energy), Sel (Self-blame), Con (Concentration Problem), Sui (Suicidal ideation), Sym (Sympathetic arousal), Pan (Panic), and All (All Symptoms).

[0045] Figures 3(a) to 3(k) visualize the multivariate time series characteristics of depressed and non-depressed users; each dimension is compared separately, with an observation window of 110 time steps; among them, Figure 3(a) sadness; Figure 3(b) decreased interest; Figure 3(c) appetite problems; Figure 3(d) insomnia; Figure 3(e) extremism; Figure 3(f) reduced energy; Figure 3(g) self-blame; Figure 3(h) decreased attention; Figure 3(i) suicidal tendencies; Figure 3(j) sympathetic arousal; Figure 3(k) panic. Detailed Implementation

[0046] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0047] This invention aims to explore the extraction of dynamic features of user depressive signals. First, a well-labeled dataset was constructed and made open-source using data collected from the Sina Weibo platform. This dataset includes 3,711 depressed users and 19,526 non-depressed users. Next, a feature extraction method is proposed, which can reflect the dynamic changes of users' depressive symptoms in the form of time series data. Furthermore, the contribution of the features to classification and their interpretability are explored. Experimental results show that the method of this invention is effective, and the extracted multivariate time series features can well characterize the dynamic changes of users' depressive states. This invention also provides a method and approach for analyzing the dynamic features of depressed users on social networks. The specific method is as follows:

[0048] 1. Dataset Construction

[0049] This invention aims to identify indicative patterns of depressive episodes from dynamic multivariate time series of user depressive symptoms. Therefore, user tweets should be collected as completely as possible. Wang et al. constructed the first large-scale dataset for detecting depression among Sina Weibo users. However, this dataset set a maximum of 100 user tweets, resulting in incomplete data collection. For this reason, this invention decided to construct its own dataset to ensure complete collection of user tweets.

[0050] (1) Candidate Sina Weibo Depression Dataset (C-SWDD)

[0051] To collect data from the Sina Weibo platform, a web crawler was first used to collect a large number of user IDs, and these were initially filtered by setting collection keywords and time periods. Next, another web crawler was built based on the Sina Weibo developer APIs to collect user profiles and tweets.

[0052] Candidate Depressed User Dataset. First, a large number of tweets were collected from the Sina Weibo topic "Depression". Then, the most frequent and distinctive words related to depression in these texts were selected as seed search terms. Next, all tweets from 2014 to 2021 were collected using these seed candidate terms, and candidate depressed user IDs were extracted from these tweets. Finally, the resumes and tweets of these users were collected, resulting in 64,513 candidate depressed users and 11,610,155 tweets.

[0053] Candidate Non-Depressed User Dataset. First, user IDs were collected under the "Daily Life" topic on Sina Weibo. Then, the resumes of these users and all their tweets from 2014 to 2021 were collected, resulting in 31,085 candidate non-depressed users and 5,144,688 tweets.

[0054] (2) Sina Weibo Depression Dataset (SWDD)

[0055] Next, the C-SWDD dataset needs to be labeled. The labeling rules are based on the Diagnostic and Statistical Manual of Mental Illnesses, Fifth Edition (DSM-5) and the work of Fried et al.:

[0056] ① Users who report having depression and provide a hospital diagnosis certificate are directly labeled as depressed; otherwise, they must meet condition ②.

[0057] ② The user exhibits at least five of the symptoms listed in Table 1 within a two-week period, with at least one of the symptoms being either the first or second symptom listed in the table.

[0058] Table 1. Symptoms of Depression

[0059]

[0060]

[0061] To ensure that annotation rules are strictly enforced while improving annotation efficiency, this invention employs a corresponding annotation system that supports multiple users.

[0062] Depression User Dataset: Despite the development of an annotation system, the annotation process was time-consuming because, according to the annotation rules, annotators needed to review a user's entire tweet history and find sufficient evidence to determine whether a user was depressed. Therefore, a total of 3,711 depressed users, comprising 785,689 tweets, were ultimately identified from the candidate depression dataset as the depression user dataset.

[0063] Non-depressed user dataset: To ensure that the non-depressed user dataset does not include depressed users, this embodiment designed a corresponding strategy to remove users whose tweets contained depressive words. In this way, 19,526 non-depressed users and a total of 4,068,732 tweets were obtained as the non-depressed user dataset.

[0064] The statistics for the C-SWDD and SWDD datasets are shown in Table 2. These datasets have been open-sourced on GitHub.

[0065] Table 2. Dataset Statistics

[0066]

[0067] 2. Feature Extraction

[0068] To extract a multivariate time series of user depressive symptoms, the most typical depressive symptoms and their descriptive text must first be identified. Next, an algorithm is designed to calculate the "discrete values ​​of depressive symptoms" for each user's tweets. These discrete values ​​are then assigned timestamps and arranged chronologically to obtain the "user's depressive symptom time series." In this way, each dimension of the multivariate time series corresponds one-to-one with a depressive symptom (such as loss of interest, anhedonia, etc.), representing the trend of a particular depressive symptom changing over time for that user. The feature extraction method of this invention is as follows: Figure 1 As shown, the flow of the feature extraction algorithm is described in Algorithm 1.

[0069] (1) Descriptive texts of depressive symptoms

[0070] This invention references the clinical definition of depression in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), and combines it with keywords of depressive symptoms summarized during the annotation process and characteristics of Weibo posts to propose a set of descriptive texts of depressive symptoms, denoted as... S The text matrix of depressive symptoms is denoted as T s :

[0071]

[0072] in, S This refers to descriptive text representing the symptoms of depression. s i This refers to descriptive text representing a specific depressive symptom.m This indicates the number of depressive symptoms. The specific descriptive text for depressive symptoms is shown in Table 3.

[0073] Table 3. Descriptive texts of depressive symptoms

[0074]

[0075] (2) Calculation of depression symptom score in tweets

[0076] This invention utilizes depressive symptom text to calculate the depressive symptom score for each user's tweet. Specifically, it first extracts the embedding of each depressive symptom text, and then constructs a depressive symptom text matrix. T s Transformed into a depressive symptom embedding matrix E s :

[0077]

[0078] in, Text indicating depressive symptoms s i Embedded, f This represents the pre-trained sentence embedding model.

[0079] Next, for each user's tweet text t Similarly, a pre-trained sentence embedding model is used. f Extract its text embedding t e Then, the depression symptom embedding matrix was used. E s Depression symptom score of the tweet S d :

[0080]

[0081] in, θ i It is the tweet text. t Depression symptom scores, f s This represents the function for calculating text similarity.

[0082] This invention uses cosine similarity as the text similarity calculation function, denoted as . cosine-sim :

[0083]

[0084] in, u,v These represent the user's tweet embedding vector and the depressive symptom text embedding vector, respectively. u i , vi These are the discrete values ​​in these two vectors. n It is their dimension.

[0085] (3) Time series of user depressive symptoms

[0086] For each user's tweet, a score representing their depressive symptoms was calculated, and these scores were then sorted and concatenated chronologically. Ultimately, this yielded a multivariate time-series feature profile of the user's depressive symptoms. MTS d :

[0087]

[0088] in, m Indicates the number of depressive symptoms. n This indicates the number of tweets by the user. It represents the set of natural numbers.

[0089] (4) Algorithm flow:

[0090] The algorithm description of the method proposed in this invention is as follows:

[0091] Algorithm 1: Extraction of Temporal Features of User Depression Symptoms

[0092] enter:

[0093] T S Depressive symptoms text matrix;

[0094] T u User tweets sorted by time;

[0095] f c Text cleaning functions;

[0096] f e , embedding extraction functions;

[0097] f s Text similarity calculation function;

[0098] Output:

[0099] MTS d Multivariate time series of user depressive symptoms;

[0100] .

[0101] 3. Experimental setup

[0102] (1) Experimental environment

[0103] All experiments in this embodiment were performed on a workstation equipped with an Intel Core™ i7-10875H CPU, an NVIDIA GeForce RTX 2060 GPU, and 32GB of RAM. Each experiment was run independently 10 times, and the results were averaged. All models were implemented using Keras and Tensorflow. The global settings for model training are shown in Table 4.

[0104] Table 4 Global Settings for Model Training

[0105]

[0106] (2) Experimental Dataset

[0107] Since SWDD is an imbalanced dataset (see Table 2), this embodiment randomly selected 3,500 depressed users and 3,500 non-depressed users as the experimental dataset, labeled as SWDD-7K (see Table 5).

[0108] In this embodiment, the dataset is divided into three subsets: 60% for training, 20% for validation, and 20% for testing. The number of depressed and non-depressed users is the same in each subset.

[0109] Table 5 Statistical information of the experimental dataset

[0110]

[0111] (3) Multivariate time series classifier

[0112] The primary objective of this invention is to verify the effectiveness of multivariate time-series features in detecting depression on online social networks. Therefore, the downstream model needs to be a classifier that supports multivariate time-series classification. This embodiment selects three state-of-the-art (SOTA) machine learning models and six deep learning models for verification and comparison. Note that these classifiers are all domain-independent end-to-end deep learning models. This is mainly based on the following two considerations: First, the accuracy of deep learning models largely depends on the quality of the extracted features. Second, domain-independent end-to-end methods are more general and better adaptable to other domains. Furthermore, models with different architectures are selected for experimental evaluation to explore which architecture is most suitable for the method.

[0113] All classifiers in this embodiment are described below:

[0114] TDE (Temporal Dictionary Ensemble) is an ensemble model designed specifically for time series classification (TSC) tasks. It combines an adaptive Gaussian processing model with a dictionary-based classifier.

[0115] ROCKET: Random convolutional kernel transform (ROCKET) is a highly efficient linear kernel TSC classifier.

[0116] TSF: Time Series Forest (TSC) is a tree ensemble TSC classifier. It combines entropy gain and distance metric, making it computationally efficient.

[0117] MLP*: A TSC baseline model based on an improved multilayer perceptron (MLP) architecture, which adds ReLU and dropout layers to the traditional MLP.

[0118] FCN*: Fully convolutional neural network (FCN) is a highly efficient model originating from the field of image semantic segmentation. It has been modified and applied to the TSC task.

[0119] MCNN: Multi-scale convolutional neural network (MCNN) is a highly efficient TSC model that can automatically extract time series features of different scales and frequencies.

[0120] MC-DCNN: Multi-channel deep convolutional neural network (MC-DCNN) is a model specifically designed for multivariate time series classification (MTSC) tasks. It can effectively learn the features of univariate time series in each channel.

[0121] MALSTM-FCN: Multivariate attention long short term memory fully convolutional network (MALSTM-FCN) is an MTSC model derived from action recognition. It is efficient and lightweight enough.

[0122] TWIESN: The time-warping invariant echo state network (TWIESN) is a powerful model for the MTSC task, capable of easily learning to classify multivariate time series data of different lengths.

[0123] 4. Method effectiveness

[0124] To verify the effectiveness of the proposed feature extraction method in online depression detection, some models were trained and tested on the SWDD-7K dataset according to the settings described above. The results are shown in Table 6 (note that since the test dataset is balanced, the precision and recall are the same).

[0125] Table 6 Classifier Performance Indicators

[0126]

[0127] * indicates that the model has been fine-tuned for a multivariate time series classification task.

[0128] The table above shows that the TSF and TWIESN models achieved the best performance in the machine learning and deep learning model groups, respectively. The TSF model even outperformed the TWIESN model slightly. This is likely because the TSF model extracts features such as the mean, standard deviation, and slope of time series changes. These features can effectively capture the differences in the time series fluctuations of depressive symptoms between depressed and non-depressed users (this difference will be analyzed in detail in Section 4.4). Besides TSF, deep learning models generally outperformed machine learning models. MLP* is a strong baseline model, achieving an F1 score of 86.88%. FCN* and MALSTM-FCN were the worst-performing deep learning models, possibly because the FCN architecture is not suitable for learning the features extracted in this invention.

[0129] 5. Feature contribution

[0130] Multivariate time series is composed of multiple univariate time series. Each dimension of the multivariate time series features in this invention can be viewed as an independent feature that can interpret changes in a user's specific depressive symptom. Therefore, to explore the role of each dimension of the depressive symptom change time series features in online depression detection, this paper uses the TWIESN model to train on each dimension of the time series features.

[0131] like Figure 2 As shown in the rightmost bar of the graph, the model performs best when all features are used. Sadness and self-blame are the two most discriminative features, indicating that these are typical symptoms of depression among users on Sina Weibo. Agitation, energy loss, concentration problem, suicidal ideation, and panic are also effective features in distinguishing between depressed and non-depressed users. Interest loss, insomnia, and sympathetic arousal can distinguish between depressed and non-depressed users, but the effect is not particularly good. It can be inferred that non-depressed users may sometimes also exhibit these symptoms. The results of this invention's experiments are basically consistent with the findings of Fried et al., in which they explored the importance of various depressive symptoms for clinical diagnosis. In addition, appetite problem is not an effective symptom in distinguishing between depressed and non-depressed users on social networks, although this symptom is important for clinical diagnosis.

[0132] 6. Feature interpretability

[0133] Time series features have the advantage of easy visualization. Visualizing and comparing the extracted features of a large number of depressed and non-depressed users revealed that the severity of depressive symptoms changes continuously over time, with depressed users experiencing more frequent and drastic changes than non-depressed users. This finding is consistent with a study by Seabrook et al. Furthermore, this invention found that certain specific time series shapes can effectively reflect and explain a user's emotional and cognitive state during that period. For ease of explanation, this embodiment randomly selected one depressed user and one non-depressed user, and compared their time series features in each dimension separately, as shown in Figures 3(a) to 3(k).

[0134] As can be seen, the curves for non-depressed users are generally flat with little fluctuation, and are mostly below the symptom value of 0.2, while the curves for depressed users show relatively drastic changes.

[0135] Furthermore, Figures 3(a) and 3(g) show dense “peaks” higher than 0.5, indicating that the user may have experienced episodes of sadness and self-blame during these times. Additionally, the two “spikes” exceeding 0.6 in Figure 3(i) and the one exceeding 0.7 in Figure 3(c) respectively indicate that the user may have experienced strong suicidal tendencies and severe appetite problems during these times.

Claims

1. A method for extracting multivariate time series features of depression symptom changes of online social network users, characterized in that, The method comprises the following steps: Step 1: constructing a well-labeled dataset and opening source by using collected network platform data; Step 2: combining the key words of depressive symptoms summarized in the labeling process and the characteristics of network platform tweets, and proposing a set of depressive symptom description texts; Step 3: calculating the similarity value between each tweet of the user and the depressive symptom description text, i.e. the depressive symptom value of the tweet, and arranging the depressive symptom values in chronological order to obtain the depressive symptom multivariate time series feature reflecting the dynamic change of the depressive symptoms of the user; Step 4: verifying the effectiveness, interpretability and contribution of each feature to classification through a multivariate time series classifier; The similarity value between each tweet of the user and the depressive symptom description text in step 3 comprises: Step 3.1 : Extracting the embeddings of each depression symptom text, transforming the depression symptom text matrix into a depression symptom embeddings matrix Step 3.2: Extracting the embeddings of each depression symptom text, transforming the depression symptom text matrix into a depression symptom embeddings matrix : wherein, represents a depression symptom text embedding, represents a pre-trained sentence embedding model; Step 3.2: Text of each tweet for the user Again, using a pre-trained sentence embedding model extract its text embedding ; Step 3.3: Utilizing depression symptom embedding matrix Computing depression symptom values for tweets : wherein, is a depression symptom value of a tweet denotes a similarity computation function;​ The depressive symptom multivariate time series feature is represented as: wherein, denotes the number of depressive symptoms, denotes the number of user tweets, denotes the set of natural numbers.

2. The online social network user depressive symptom change multivariate time series feature extraction method according to claim 1, characterized in that, The dataset constructed in step 1 is specifically: Step 1.1: collecting user IDs of several network platforms by using a crawler, and preliminarily screening by setting the collection keywords and collection time period related to depression, and then constructing another crawler based on the APIs of the developer of the network platform to collect user resumes and tweets, so as to construct a candidate depressive user dataset and a candidate non-depressive user dataset; Step 1.2: labeling the candidate depressive user dataset and the candidate non-depressive user dataset through a labeling system supporting multi-person labeling to obtain a depressive user dataset and a non-depressive user dataset.

3. The online social network user depressive symptom change multivariate time series feature extraction method according to claim 2, characterized in that, The step 1.1 specifically comprises: Constructing a candidate depressive user dataset: first, collect a large number of tweets under the "depression" topic of the network platform; then, count the depressive words with the highest frequency and the most discriminative degree in these texts as seed retrieval words; then, use these seed retrieval words to collect tweets in a specific time period, and parse the candidate depressive user IDs from these tweets; finally, collect the resumes and tweets of these users to obtain several candidate depressive users and their tweets; Constructing a candidate non-depressive user dataset: first, collect user IDs under the "daily" topic of the network platform, then collect the resumes and all tweets of these users in the specific time period to obtain several candidate non-depressive users and their tweets.

4. The online social network user depressive symptom change multivariate time series feature extraction method according to claim 1, characterized in that, The depression symptom text matrix composed of the depression symptom description texts in Step 2 is denoted as : wherein, represents the set of depressive symptom description texts, represents the descriptive text of a certain depressive symptom, represents the number of depressive symptoms.

5. The online social network user depressive symptom change multivariate time series feature extraction method according to claim 1, characterized in that, The similarity calculation function is a cosine similarity function, denoted as: ; wherein, respectively represent the user tweet embedding vector and the depression symptom text embedding vector, are discrete values in these two vectors, respectively.

Images