Optimal hybrid classification model for event recommendation system

Abstract

There is a growing need for recommender systems and other ML-based systems as an abundance of data is now available across all industries. Various industries are currently using recommender systems in slightly different ways. These programs utilize algorithms to propose appropriate products to consumers based on their prior choices and interactions. Moreover, Systems for recommending events to users suggest pertinent happenings that they might find interesting. As opposed to an object recommender that suggests books or movies; event-based recommender systems typically require distinct algorithms. A developed event recommendation method is introduced which includes two stages: feature extraction and recommendation. In stage, I, a Set of features like personal willingness, community willingness, informative content, edge weight, and node interest degree are extracted. Stage II of the event recommendation system performs a hybrid classification by combining LSTM and CNN. In the LSTM classifier, optimal tuning is done by Improvised Cat and Mouse optimization (ICMO) algorithm. The results of the ICMO technique at an 80% training percentage have the maximum sensitivity value of 95.19%, whereas those of the existing approaches SSA, DINGO, BOA, and CMBO have values of 93.89%, 93.35%, 92.36%, and 92.24%. Finally, the best result is then determined by evaluating the whole performance.

Keywords

Event Recommendation feature extraction LSTM CNN recommendation system Improvised Cat and Mouse optimization (ICMO)

Nomenclature

Description

three-tier IoT-edge-cloud

attention-based context-aware group event recommendation model

Artificial neural network

Collaborative filtering

Cat and Mouse Optimization Algorithm

convolutional neural network

content-venue-aware topic model

event-based social networks

Graph-based Context-Aware Recommendation Systems

Group Event Recommendation Framework

IoT data and collaborative filtering-based recommendation

Long term short memory

learning-to-rank

machine learning

Restricted Boltzmann Machines

Rectified linear unit

software-as-a-service

Spatial-Temporal Topic Model

1. Introduction

Technology has improved our quality of life and given us access to a wide range of information, allowing us to search for what we need [28]. Combining online and physical user engagement has become a new development trend in recent years as a result of the advancement of the mobile Internet, which has made user interaction easier than ever. Process behavior analysis [7] is the focus of the data science discipline known as process mining. Recent years have seen the emergence of EBSN [17,29,32], which combines offline real events with a person’s virtual online social ties. In EBSN, people typically take part in events in groups. Examples include running marathons with buddies, watching films with families, attending conferences and seminars with coworkers, etc. In EBSN, there seem to be a lot of customers, organizations, and happenings. Social networking sites have recently grown quickly and gained recognition from both the business and academic worlds. EBSN is a complicated social network that connects offline activities and internet community groups. For an instance, Meetup1 has more than 16 million subscribers and hosts more than 300,000 events per month. In the financial markets, security analysts [30] play a crucial role by offering knowledgeable recommendations on the investments of the equities they cover. Helping users select appropriate and customized events to attend is a common duty in EBSNs.

Recent years have seen a rise in interest in personalized recommendations. It’s been demonstrated that enhancing recommendation systems with topology data, community structure, and semantic meaning can increase their effectiveness. EBSNs [20] have two components, as opposed to standard social networks: internet community relationships and offline events involvement connections [6,21]. Offline relationships are created by attending similar offline activities, whereas online relationships are created by joining similar online lobby groups. In conventional social networking sites, the majority of recommendation work focuses on assessing users’ social power from their online relationships. Nevertheless, users are also influenced by things that happen in the actual world. In addition to the users’ interest, the geographical location and timing of the events have a significant impact on whether users can participate. In order to indicate if users are prepared to engage in an event, the geographical and temporal elements in EBSNs are just as significant as the social feature. Systems for event recommendations utilize algorithms to present people with relevant choices based on their prior choices and expertise. Due to the event recommendation mechanism, this is the case doesn’t employ the conventional recommendation system scenario. We attempted to adopt a hybrid approach in this work that combines sentiment analysis and collaborative filtering [15]. Instances of recommendation systems in use include Netflix, YouTube, Tinder, and Amazon. For instance, an e-commerce program might propose new things that a consumer would be keen to buy, or the YouTube recommender system [13] might advise films to watch next. Relying on their selections, the programs tempt consumers with pertinent offers. For online newspapers, recommender systems can improve user – experience. An e-commerce website can employ a recommender system for a variety of purposes. It may market products on our website, create customized emails, and make product suggestions. There are many benefits to using this Saas platform for such an e-commerce company. Among the most widely used recommendation techniques are CF and its variants. To develop their own personalized film recommendation engine [16,31], for instance, as a resume work, even novice data scientists can utilize it.

The main objectives are:

Feature extraction and recommendation are the two phases of a brand-new event recommendation system.

A hybrid classification is done by combining LSTM and CNN.

In the LSTM classifier, optimal tuning is done by Improvised Cat and Mouse optimization (ICMO) algorithm.

The structure is as follows: Section 1 presents the Introduction. Related papers are discussed in Section 2. Proposed event recommendation system: Section 3 provides an architectural description. The description of feature extraction and suggestion is covered in Section 4. Section 5: Merged LSTM and CNN cover the suggested hybrid classifier. Section 6 discusses the Improvised Cat and Mouse Optimization (ICMO) algorithm. Section 6 discusses the conclusion.

2. Literature review

2.1. Literature review

In 2021, Guoqiong Liao et al. [22] developed an ACGER model in EICPSs. Multilayer neural networks are used by ACGER to simulate the complex, nonlinear impact of settings on people, organizations, and activities. In particular, a novel attentiveness technique is intended to allow context-influence values on users to alter dynamically in response to the events in question. A group’s choice is determined from multiple viewpoints: indirect choice and direct choice, taking into account that groups may exhibit radically distinct behavioral traits from team members.

In 2020, Yulu Du et al. [10] suggested forthcoming activities to a user group. The ranking issue of group recommendation is defined, and a GERF-oriented LTR method is suggested. To be more precise, we examine several contextual factors that may affect a user’s decision to attend events and then extract the user’s preference for the event taking into account each contextual factor. Then, in simulating the interests of the group, the choice ratings of the users inside the group are used as the characteristics for LTR.

In 2020, Yulu Du et al. [9] learn the relationship between both the event planner and the textual information, i.e., how the content of events hosted by the same planner tends to be increasingly similar. CVTM is given to record collective interests in activity from multiple viewpoints: content and venue, relying on this discovery. To reduce the lack of textual material, the connection between the planner and the topic is represented in CVTM, allowing us to more precisely derive group interests from an event’s text. Lastly, a CVTM-based technique for group event suggestions is suggested.

In 2020, Ruichang Li et al. [19] suggest a new hybrid deep architecture to resolve the cold-start issue of cooperative event suggestion. Both conditional RBM and multiple RBM are included in our system. From individual and group comments, the former derives high specific group preferences. Based on descriptive data, including event venue and organizer, the latter derives hidden event attributes. In order to address the cold-start problem, the hybridized deep structure can make use of customer reviews and contextual knowledge about occurrences.

In 2020, Ruichang Li et al. [18] suggest an STTM for recommending a cold-start event and examining the connections between time, place, and planner in EBSNs. Users have a varied event and venue subject dispersion in STTM at different periods and STTM can record user preferences on information and geographical location altering over time.

In 2021, Pratibha Mahajan and Pankaj Deep Kaur [23] suggested an event recommendation system. The proposed method blends IoT data, CF methodology, and social influence to recommend an event where there is a high likelihood that users will participate. For every candidate occurrence, the predictive value for both IoT-oriented characteristics and CF regarding social influencers is first determined. The customer is then shown the main event with the highest predictive rating.

In 2021, Pratibha Mahajan and Pankaj Deep Kaur [24] suggested a 3T-IEC-related solution for real-time context-aware event recommendation issues. In addition to an event suggestion engine in the cloud layer, the 3T-IEC provides an edge computing layer wherein IoT data is analyzed to derive and contextualize IoT-based location data. Additionally, the events have been filtered using contextual data such as the user’s present location, climate, and temporal viability. Additionally, selected events are modeled after grouping, class, and economic factors in order to rank them using a multi-criteria decision-making process.

In 2021, Lan Zhang et al. [33] offer a novel GCAR system with a knowledge graph to analyze and forecast users’ behavior, i.e., to provide suggestions based on past occurrences and their implicit correlations. The model combines contextual data that has been gleaned from users’ past actions as well as relationships between occurrences, where the settings have been represented as knowledge graphs. Events interconnections and their nuanced relationships can be built by reaping the benefits of the knowledge graph’s benefits, and they have already been incorporated into the recommendation system.

In 2021, Kudori et al. [15] Collaborative filtering is one of the techniques that can be applied when creating event recommendation systems. The event recommendation system stands out from other suggestion systems because of character. This is so because the event recommendation system doesn’t employ the conventional recommendation system scenario.

In 2022, Gupta et al. [12] have presented the user is mapped with the process in the work and the selected item is mapped with the symbol. An improved similarity metric that takes into account sequential preferences is produced by matching the EMISSION matrix of users’ item selection processes. Using the Movie Lens dataset and a number of evaluation measures, the analysis of proposed method is determined for various configurations. Table 1 displays the challenges of existing methods.

Table 1
features and challenges of conventional methods

Author [citation] Methods Features Challenges

Guoqiong Liao et al. [22] ACGER model Via multilayered neural nets, ACGER simulates the deep, nonlinear effects of settings on individuals, organizations, and events. To properly produce recommendations, a variety of data is required.

Yulu Du et al. [10] LTR technique It is a group of algorithmic methods that use supervised ML to address relevancy ranking issues in search. Ranking people when many users are being evaluated is challenging.

Yulu Du et al. [9] CVTM To gauge audience interest in a gathering from both the event’s content and setting aspects. Significant configuration is required.

Ruichang Li et al. [19] Hybrid model (RBM + conditional RBM) To get around the cold-start issue, it uses customer insights and relevant data about occurrences. Because it is challenging to determine the Energy gradient equation, training is more challenging.

Ruichang Li et al. [18] STTM It records users’ evolving interests in both content and geographic location. It’s hard to capture user engagement.

Pratibha Mahajan and Pankaj Deep Kaur [23] IoTCFR method IoT and CF data are merged to recommend an event where there is a high likelihood that users will participate. Adding side capabilities for the item/query is difficult.

Pratibha Mahajan and Pankaj Deep Kaur [24] 3T-IEC system By using proximity networking, the edges tier collects information from the network edge. Although costs for cloud storage are reduced, there is a local end cost.

Lan Zhang et al. [33] GCAR system It examines and forecasts user behavior. Needs significant spending.

Kudori et al. [15] Collaborative filtering Increased accuracy of the proposed model Computation time is more complex.

Gupta et al. [12] HMM-based approach selection process provides an improved similarity measure Larger datasets are needed

Author [citation]	Methods	Features	Challenges
Guoqiong Liao et al. [22]	ACGER model	Via multilayered neural nets, ACGER simulates the deep, nonlinear effects of settings on individuals, organizations, and events.	To properly produce recommendations, a variety of data is required.
Yulu Du et al. [10]	LTR technique	It is a group of algorithmic methods that use supervised ML to address relevancy ranking issues in search.	Ranking people when many users are being evaluated is challenging.
Yulu Du et al. [9]	CVTM	To gauge audience interest in a gathering from both the event’s content and setting aspects.	Significant configuration is required.
Ruichang Li et al. [19]	Hybrid model (RBM + conditional RBM)	To get around the cold-start issue, it uses customer insights and relevant data about occurrences.	Because it is challenging to determine the Energy gradient equation, training is more challenging.
Ruichang Li et al. [18]	STTM	It records users’ evolving interests in both content and geographic location.	It’s hard to capture user engagement.
Pratibha Mahajan and Pankaj Deep Kaur [23]	IoTCFR method	IoT and CF data are merged to recommend an event where there is a high likelihood that users will participate.	Adding side capabilities for the item/query is difficult.
Pratibha Mahajan and Pankaj Deep Kaur [24]	3T-IEC system	By using proximity networking, the edges tier collects information from the network edge.	Although costs for cloud storage are reduced, there is a local end cost.
Lan Zhang et al. [33]	GCAR system	It examines and forecasts user behavior.	Needs significant spending.
Kudori et al. [15]	Collaborative filtering	Increased accuracy of the proposed model	Computation time is more complex.
Gupta et al. [12]	HMM-based approach	selection process provides an improved similarity measure	Larger datasets are needed

3. Proposed event recommendation system: Architectural description

In this work, an event-based recommendation system is presented. The group of events is assumed as $event = {_{1}, E_{2}, \dots, E_{n}}$ and the user group is indicated as $v = {V_{1}, V_{2}, \dots, V_{u}}$ . Additionally, the group set is expressed as $g = {G_{1}, G_{2}, \dots, G_{m}}$ . Here, $n, u, m$ were the total user count, events, and groups. Therefore, each one of the occurrences is referred to as a tuple that provides data about its date, time, and class. Additionally, each user has the option of joining several groups, and each organization has the ability to organize events for its membership. In the proposed event recommendation system, the model work with respect to different constraints related to the user’s perspective. Thereby, the system consists of 2 phases: Feature extraction and Recommendation.

Feature extraction: In this stage, a set of features like personal willingness, community willingness, informative content, edge weight, and node interest degree are extracted.

Recommendation: In this stage, a hybrid classification is performed for the event recommendation system. LSTM and CNN models are combined to optimize weight in order to improve classification performance. Figure 1 depicts the proposed event-based recommendation system.

Fig. 1.

Proposed event recommendation system.

4. Description on feature extraction and recommendation

In the feature extraction stage, a set of features like personal willingness, community willingness, informative content, edge weight, and node interest degree were extracted.

Personal willingness [ 11 ]: The user’s subjective effort to share data with the outside community is referred to as personal willingness. The level of personal willingness that a user selects when they join a community is depicted as $ψ_{u}$ .

Where, $ψ_{u} \in [0, 1]$

The user may select his preferred willingness value based on his needs. The user is fully exposed to the outer world when $ψ_{u} = 1$ . The identity of the user is kept secret from the public when $ψ_{u} = 0$ .

Community willingness [ 11 ]: Community willingness is expressed as $ψ_{c}$ . It serves as a gauge for determining the level of interaction within the community in relation to other groups. It serves as a gauge for determining the level of communication within the community in relation to other communities.

Where, $ψ_{c} \in [0, 1]$

Every community is free to set its community willingness in accordance with its own needs, where each community member must meet the requirement of communal willingness.

Informative content [ 11 ]: Informative content feature is depicted as $cont = {{cont}^{1}, {cont}^{2}, \dots, {cont}^{k}}$ where ${cont}^{k}$ is the feature of kth the message piece.

Proposed edge weight: Edge closeness and node degree are used to calculate edge weight. Whereas node willingness $ψ_{i, j}$ is included in the calculation of $b_{i, j}^{ψ}$ node willingness is also taken into consideration when calculating edge weight. The conventional edge weight [11] is outlined as follows in Eq. (1): $\begin{array}{l} (1) & a_{i, j} = \partial . b_{i, j}^{ψ} + (1 - \partial) . \sqrt{\frac{A_{i, j}}{h_{i}^{out} . h_{j}^{in}}} \end{array}$

Here $b_{i, j}^{ψ}$ is determined as $b_{i, j}^{ψ} = \frac{λ_{i, j}^{ψ}}{\sqrt{q_{i}^{ψ} - q_{j}^{ψ}}}$ and $A_{i, j} = {\begin{matrix} 1, if node i and j coupled \\ 0, otherwise \end{matrix}$

The information passing between two users was evaluated by σ calculation which acts as the weighting function in the proposed evaluation. As per the proposed edge weight is outlined in Eq. (2): $\begin{array}{l} (2) & a_{i, j} = \partial . b_{i, j}^{ψ} + (1 - \partial) . \sqrt{\frac{A_{i, j}}{h_{i}^{out} . h_{j}^{in}}} \times σ \end{array}$

The standard deviation is depicted in Eq. (3).

Where, $\begin{array}{l} (3) & σ = \sqrt{\frac{1}{N} \sum_{i = 1}^{N} {(x_{i} - μ)}^{2}} \end{array}$

Here,

$σ =$ standard deviation

$μ =$ population average

$N =$ population size

$x_{i} =$ each population-based value

$\partial \in [0, 1]$ = impact factor

$b_{i, j}^{ψ} =$ edge intimacy degree

$h^{out} =$ out-degree of node i

$h^{in} =$ in the degree of node j

Only a topology framework or community structure is used to divide communities, when $\partial = 0$ . Edge intimacy degree, which describes the frequency of communication or communication between two users is the single factor that influences edge weight, when $\partial = 1$ .

Proposed node interest degree: Users with similar interests are more able to engage in the same communities and share chats of worth. Users can specify the labels for their own preferences when they joined a community. For the message ${cont}_{i}$ , it is depicted I Eq. (4). $\begin{array}{l} (4) & {cont}_{i} = {(f_{1}, a_{1}), (f_{2}, a_{2}), \dots, (f_{N}, a_{N}) \end{array}$

Where,

$a_{i} =$ weight

$f_{i} = i^{t h}$ keyword in the message ${cont}_{i}$

In conventional node interest degree, depending on message structure ${cont}_{i}$ , Eq. (5) shows the definition identifies the interesting similarities between the nodes i and j. $\begin{array}{l} (5) & Intr (i, j) = simi (i, j) = \frac{\sum_{n = 1}^{h} a_{i, n} . a_{j, n}}{\sqrt{\sum_{n = 1}^{h} a_{i, n}^{2}} . \sqrt{\sum_{n = 1}^{h} a_{j, n}^{2}}} . ψ_{i, j} \end{array}$

As per the proposed evaluation, the user’s mutual information is averaged in the proposed node interest degree which is depicted in Eq. (6). $\begin{array}{l} (6) & Intr (i, j) = simi (i, j) = \frac{\sum_{n = 1}^{h} a_{i, n} . a_{j, n}}{\sqrt{\sum_{n = 1}^{h} a_{i, n}^{2}} . \sqrt{\sum_{n = 1}^{h} a_{j, n}^{2}}} . ψ_{i, j} \times H M_{a} \end{array}$

Here $H M_{a}$ is evaluated in Eq. (7). $\begin{array}{l} (7) & H M_{a} = \frac{\sum_{i = 1}^{n} a_{i}}{\sum_{i = 1}^{n} \frac{a_{i}}{x_{i}}} \end{array}$

Where, $H M_{a} =$ weighted harmonic mean

$a =$ weight function

$h =$ public keywords count

$a_{i, n} = n^{t h}$ public keyword weight of message from ith node

$a_{j, n} = n^{t h}$ public keyword weight of message from jth node

$ψ_{i, j} =$ node willingness

$Intr =$ interest

Finally, the obtained feature set F is defined in Eq. (8) $\begin{array}{l} (8) & F = (ψ_{u}, ψ_{c}, cont, a_{i, j}, Intr (i, j)) \end{array}$

5. Proposed hybrid classifier: Combining optimized LSTM and CNN

Hybrid classification is done by combining LSTM and CNN. For model induction and data preprocessing, the notion of hybrid classification makes use of fundamental classification techniques. The better outcome is obtained by averaging the combined LSTM and CNN classifiers. In the LSTM classifier, the weight is optimally tuned by the developed optimization algorithm.

5.1. Optimized LSTM classifier

The feature set F will be the input to optimized LSTM [5].LSTM is used to preserve pretty long series data reliance and accomplish high precision forecasting, the sequence data must be learned and memorized by the special gated units. The forget gate, input gate, and output gate are the three gates components that make up the majority of the LSTM neural network. The analysis is acquired and remembered by the specific gate units in order to maintain long-period series information dependence and achieve large predictions. The forget gate controls how well the current cell can remember prior information. The output gate represents the neuron’s output. Let’s say the input data is $(r_1, r_2, \dots, r_t)$ at time t. The below Eq. (9), Eq. (10), Eq. (11), Eq. (12), Eq. (13) are the computation formula for each LSTM: $\begin{array}{l} (9) & l_{t} = p (a_{l} \times [k_{t - 1}, e_{t}] + u_{l}) \\ (10) & m_{t} = p (a_{m} \times [k_{t - 1}, e_{t}] + u_{m}) \\ (11) & n_{t} = p (a_{n} \times [o_{t}, k_{t - 1}, e_{t}] + u_{n}) \\ (12) & o_{t} = m_{t} \times o_{t - 1} + l_{t} \times tan k (a_{C} \times [k_{t - 1}, e_{t}]) \\ (13) & k_{t} = n_{t} \times tan k (o_{t}) \end{array}$

Where,

$e_{t} =$ input LSTM

$k_{t - 1} =$ hidden layer output

$l_{t}, m_{t}, n_{t} =$ input gate input

$a_{l}, a_{m}, a_{n} =$ weight matrix

$u_{l}, u_{m}, u_{n} =$ offset vector

$a_{C} =$ the cell unit’s weight in relation to the input

$k_{t} =$ hidden layer output

$p =$ sigmoid function

In the LSTM classifier, weight is optimally tuned by an improved cat and mouse optimization algorithm, which is explained clearly in Section 6.

5.2. Convolutional neural network (CNN)

The feature set F will be the input to CNN [1]. a CNN, a more developed variation of the ANN substitutes the mathematical calculation known as convolution for generic matrix multiplication in at minimum one of its layers. They are employed in image analysis and detection since they were created primarily to handle pixel data.

These layers carry out actions on the data in order to discover characteristics unique to the data. Convolution, activation or ReLU, and pooling are 3 of the most used layers.

Convolution layer: Inserts a series of convolutional filters across the input images, each one engaging different aspects of the images.

ReLU layer: Enables quicker and more efficient training by retaining positive attributes and mapping positive attributes to zero. Due to the fact that only the engaged features are carried over to the following layer, this is regularly alluded to as activation.

Pooling layer: Reduces the parameters count the system wants to understand by conducting nonlinear down sampling on the outcome.

Each layer learns to recognize various traits as these procedures are repeated across tens or even hundreds of levels. Figure 2 depicts the architecture of the proposed hybrid classification algorithm.

Fig. 2.

Architecture of proposed hybrid classification algorithm.

6. Improvised cat and mouse optimization (ICMO) algorithm for optimal weight tuning in LSTM

CMBO [2] theory is introduced in this part, followed by a presentation of its mathematical formula for use in optimizing a variety of situations. CMBO is a population-oriented technique that draws its design cues from the instinctive actions of a mouse being attacked by a cat and running away to safety. Where the optimization procedure or problem is designed and described, algorithms for optimization are used in all fields of study and in practical problems. Various goal functions can be minimized or maximized using the suggested CMBO. When choosing decision factors to maximize device performance, CMBO can be employed in engineering studies, and the best concepts. A matrix is used to calculate the system’s population which is defined in Eq. (14) and Eq. (15) $\begin{array}{l} (14) & X^{R} = [\begin{matrix} X_{1}^{R} \\ ⋮ \\ X_{i}^{R} \\ ⋮ \\ X_{N}^{R} \end{matrix}] = [\begin{matrix} g_{1, 1}^{r} \dots g_{1, d}^{r} \dots g_{1, v}^{r} \\ ⋮ \\ g_{i, 1}^{r} \dots g_{i, d}^{r} \dots g_{i, v}^{r} \\ ⋮ \\ g_{N, 1}^{r} \dots g_{N, d}^{r} \dots g_{N, v}^{r} \end{matrix}] \\ (15) & Q^{R} = [\begin{matrix} Q_{1}^{R} min (Q) \\ ⋮ \\ Q_{N}^{R} max (R) \end{matrix}] \end{array}$

Where,

$X^{R} =$ sorted population matrix

$X_{i}^{R} =$ sorted population’s ith member

$g_{i, d}^{r} = d^{t h}$ problem value

$Q^{R} =$ sorted vector

Two sets of mice and cats make up the population matrix in the suggested CMBO. In the CMBO, it is presumed that the cat population is made up of all the halves of the population participants who produced lower limits for the goal function and the other halves of the members of the population who offered excellent value for the optimization problem. Equations (16) and (17), indicates the population of mice and cats. $\begin{array}{l} (16) & T = [\begin{matrix} T_{1} = X_{1}^{R} \\ T_{i} = X_{i}^{R} \\ T_{N_{v}} = X_{N_{v}}^{R} \end{matrix}] = {[\begin{matrix} g_{1, 1}^{r} \dots g_{1, d}^{r} \dots g_{1, v}^{r} \\ ⋮ \\ g_{i, 1}^{r} \dots g_{i, d}^{r} \dots g_{i, v}^{r} \\ ⋮ \\ g_{N_{v}, 1}^{r} \dots g_{N_{v}, d}^{r} \dots g_{N_{v}, v}^{r} \end{matrix}]}_{N_{v} \times v} \\ (17) & Z = [\begin{matrix} Z_{1} = X_{N_{v} + 1}^{R} \\ ⋮ \\ Z_{j} = X_{N_{v} + j}^{R} \\ ⋮ \\ Z_{N_{c}} = X_{N_{v} + N_{c}}^{R} \end{matrix}] = {[\begin{matrix} g_{N_{v} + 1, 1}^{r} \dots g_{N_{v} + 1, d}^{r} \dots g_{N_{v} + 1, v}^{r} \\ g_{N_{v} + j, 1}^{r} \dots g_{N_{v} + j, d}^{r} \dots g_{N_{v} + j, v}^{r} \\ g_{N_{v} + N_{c}, 1}^{r} \dots g_{N_{v} +, N_{c}, d}^{r} \dots g_{N_{v} + N_{c}, v}^{r} \end{matrix}]}_{N_{c}, v} \end{array}$

Where,

$T =$ mice population matrix

$N_{v} =$ mice count

$T_{i} = j^{t h}$ mouse

$Z =$ cat population matrix

$Z_{j} = i^{t h}$ cat

Intially stage, the position shift of cats is formulated based on their natural habit and motion toward mice in updating the search criteria. The conventional update function is depicted in Eq. (18), Eq. (19), Eq. (20). $\begin{array}{l} (18) & Z_{j}^{new} : z_{j, d}^{new} = z_{j, d} + y \times (v_{h, d} - D \times z_{j, d}) j = 1, N_{z}, d = 1 : v, h \in 1 : N_{v} \\ (19) & D = rou (1 + ran) \\ (20) & Z_{j} = \{\begin{matrix} Z_{j}^{new} & Q_{j}^{z, new} ≺ Q_{j}^{z} \\ Z_{j} & else \end{matrix} \end{array}$

As per the suggested logic, the improved position update is illustrated in Eq. (21) $\begin{array}{l} (21) & Z_{j}^{new} : z_{j, d}^{new} = z_{j, d} + y \times (\frac{v_{h, d} - D \times z_{j, d}}{B} \times a_{i}) \end{array}$

Where,

$Z_{j}^{new} = j^{t h}$ cat new status

$z_{j, d}^{new} = d^{t h}$ problem new value

$v_{h, d} = d^{t h}$ dimension of $h^{t h}$ mouse

$Q_{j}^{z, new} =$ objective function value

$y =$ random number

$a_{i} =$ weight matrix

$B =$ radius which is calculated as $B = \frac{G}{2 π}$

$G =$ circumference of cat while catching mouse

By patterning the placements of some algorithmic elements, the location of the hideaways in the search area is generated at random. Equations (22), (23), (24) were used to model this step of updating the positions of mice analytically. $\begin{array}{l} (22) & U_{i} : u_{i, d} = g_{e, d} i = 1 : N_{v}, d = 1 : v, e \in 1 : N \\ (23) & T_{i}^{new} : v_{i, d}^{new} = v_{i, d} + y \times (u_{i, d} - D \times v_{i, d}) \times sig (Q_{i}^{v} - Q_{i}^{U}) i = 1 : N_{v}, d = 1 : v \\ (24) & T_{i} = \{\begin{matrix} T_{i}^{new}, & Q_{i}^{v, new} ≺ Q_{i}^{v} \\ T_{i}, & else \end{matrix} \end{array}$

As per the improved method, the random number y is referred by a circle map which is given in Eq. (25). $\begin{array}{l} (25) & Z_{h + 1} = Z_{h} + 0.5 - \frac{1.1}{π} sin (2 π Z_{k}) mod (1) \end{array}$

Algorithm 1

Pseudocode of ICMO

The pseudocode of the proposed model is shown in Algorithm 1.

7. Results and discussions

7.1. Simulation setup

The proposed Event Recommendation model was implemented in python. The data is collected from [3] [Access date: 2021-04-26]. The proposed method is more effective in event recommendation, analysis results includes positive measures (Accuracy, Sensitivity, Specificity, and Precision), negative measures (FNR, FPR), and other measures (F-measure, NPV and MCC). The system configuration of device specifications is displayed in Table 2. The adopted model Improvised Cat and Mouse Optimization (ICMO) was compared with Salp Swarm Algorithm (SSA), [4] Butterfly Optimization Algorithm (BOA), [27] DINGO Optimization Algorithm (DOA) [14] and Cat and Mouse Based Optimization (CMBO) [8] by varying the training percentage 60, 70, 80 and 90 respectively.

Table 2
System Configuration of Device specifications

Device Specifications

Device name DESKTOP-9FPAHOE

Processor AMD Ryzen 5 3450U with Radeen Vega Mobile Gfx 2.10 GHz

Installed RAM 16.0 GB(13.9 GB usable)

Device ID OFBA11F4-A5BO-44E7-B3F8-EB851E779A20

Product ID 00327-36322-31369-AAOEM

System type 64-bit operating system,X64-based processor

Pen and touch No pen or touch input is available for this display

Device	Specifications
Device name	DESKTOP-9FPAHOE
Processor	AMD Ryzen 5 3450U with Radeen Vega Mobile Gfx 2.10 GHz
Installed RAM	16.0 GB(13.9 GB usable)
Device ID	OFBA11F4-A5BO-44E7-B3F8-EB851E779A20
Product ID	00327-36322-31369-AAOEM
System type	64-bit operating system,X64-based processor
Pen and touch	No pen or touch input is available for this display

7.2. Performance analysis with respect to a positive measure

The positive measures of the suggested model ICMO has attained maximum when compared to the other approaches. The outcomes with respect to Accuracy, Sensitivity, Specificity, and precision are depicted in Fig. 3. The results visualized evidence that the suggested method is efficient than the other techniques whenever the learning percentages changes. The accuracy of the proposed ICMO model at 90% training percentage is 96.92%, which, in comparison to the greater value as SSA = 95.25%, DINGO = 95.13%, BOA = 94.07%, and CMBO = 93.70% (approximate). On observing the outcomes of the ICMO approach at 80% training percentage has obtained the highest sensitivity value as 95.19% while the others have attained the value as 93.89%, 93.35%, 92.36%, 92.24% for the extant approaches SSA, DINGO, BOA, CMBO. It has been observed that precision has experienced 2 to 4 percent increases for most of the classifiers. More improvement is seen with the adopted ICMO model. When compared to the existing methods, the value of precision rate of the proposed method was better than all the other classifiers. Thus from the evaluation of the positive measure, it is clear that the adopted ICMO-based recommended model is best for the event recommendations. This improvement is due to the influence of optimization strategy on solving the error minimization problem to determine appropriate recommendations.

Fig. 3.

Performance analysis on proposed approach ICMO in terms of a) accuracy b) sensitivity c) specificity and d) precision.

7.3. Performance analysis with respect to the negative measure

The negative measure of FNR and FPR are analysed for the adopted ICMO as well as the existing model in the different training percentages. Figure 4 shows that error value is less in the suggested method than it is highly efficient for event recommendation. The observed negative values, an impressive result has been obtained i.e. lower errors are noted for overall variations in the learning percentage. The FNR of the proposed is 2.927 which is the least value of the existing approaches like SSA = 3.45, DINGO = 3.54, BOA = 4.058, and CMBO = 4.32. Additionally, the FPR of the ICMO at 80% learning is 5.12, which is the minimum value when comparing to the others SSA, DINGO, BOA, and CMBO has obtained the values 10.53, 11.33, 14.20, 14.38 respectively. It is clear from the general evaluation that the suggested ICMO approach is less error-prone and that the likelihood of misclassification is extremely low.

Fig. 4.

Performance analysis on proposed approach ICMO in terms of a) FNR and b) FPR.

7.4. Performance analysis with respect to other measures

Figure 5 shows that F-measure, MCC, and NPV that are computed to add extra value to the proposed model. When the other measures for the adopted model ICMO are higher, it becomes highly efficient for event recommendation. Additionally, the proposed had achieved the highest value for all these measures. The f-measure of the ICMO at the 90% training percentage is 97.23% which is superior to the extant approaches like SSA, DINGO, BOA, and CMBO. Subsequently, the MCC measure of the 80% training percentage is 90.02% which is the highest value when compared to the existing approaches like SSA = 83.59%, DINGO = 82.25%, BOA = 78.49%, and CMBO = 78.19%. Thus from the evaluation, it is clear that the proposed ICMO model is much more efficient for event recommendations.

Fig. 5.

Performance analysis on proposed approach ICMO in terms of a) F-measure b) MCC and c) NPV.

7.5. Ablation study

Machine learning systems components are typically deleted or replaced as part of an experiment called an ablation study to determine how these changes affect the systems

Performance. The analysis of the proposed ICMO methods with different metrics as shown in Table 3. This section has proved the variation of performance of the presented study with the conventional feature set and optimization. The tabulation has been addressed with respect to different positive and negative measures.

Table 3
Ablation study

Metrics Conventional Node Interest Conventional Edge Weight Without optimization ICMO

Accuracy 0.874596 0.841981 0.907691 0.950522

Sensitivity 0.901546 0.881645 0.90805 0.951973

Specificity 0.845581 0.800842 0.907283 0.948782

Precision 0.862031 0.820527 0.915637 0.95708

F1-score 0.881345 0.849989 0.911827 0.95452

MCC 0.756776 0.692906 0.833671 0.900291

NPV 0.889481 0.867891 0.898817 0.942747

FPR 0.154419 0.199158 0.092717 0.051218

FNR 0.098454 0.118355 0.09195 0.048027

Metrics	Conventional Node Interest	Conventional Edge Weight	Without optimization	ICMO
Accuracy	0.874596	0.841981	0.907691	0.950522
Sensitivity	0.901546	0.881645	0.90805	0.951973
Specificity	0.845581	0.800842	0.907283	0.948782
Precision	0.862031	0.820527	0.915637	0.95708
F1-score	0.881345	0.849989	0.911827	0.95452
MCC	0.756776	0.692906	0.833671	0.900291
NPV	0.889481	0.867891	0.898817	0.942747
FPR	0.154419	0.199158	0.092717	0.051218
FNR	0.098454	0.118355	0.09195	0.048027

7.6. Analysis of classifiers

Table 4 shows the detailed study on classifier analysis for SSA, BOA, DINGO, CMBO, and ICMO. The developed method ICMO attained greater accuracy value of 0.95 when compared to the existing techniques like LSTM = 0.90, CNN = 0.87, RNN = 0.80, and NN = 0.98. The accepted model’s f-measure is 0.95, which is better than existing methods like LSTM, CNN, RNN, and NN. The FNR has the highest value of 0.21 for the existing method RNN, moreover, the suggested ICMO approach has acquired the lesser error’ of 0.07. Furthermore, the analysis also taken in the classifiers, 0.94 and 0.90 is the obtained NPV and MCC measure of the proposed model. The experimental results are evaluated with FPR, interestingly, it is noticeable the suggested method obtained lesser error values. Similarly, the positive measure of Sensitivity, Specificity, and precision of the proposed ICMO model maintain the highest value of 0.951, 0.94, and 0.957. The analysis results of the proposed model for the positive and the other measure is higher than the existing approaches and the negative measure is low.

Table 4
Analysis on Classifiers

ICMO LSTM CNN RNN NN

Accuracy 0.950522 0.907592 0.878838 0.807331 0.908433

Sensitivity 0.951973 0.900495 0.875184 0.78301 0.9293

Specificity 0.948782 0.915982 0.882839 0.828904 0.886002

Precision 0.95708 0.926845 0.891072 0.802355 0.89757

F-measure 0.95452 0.91348 0.883056 0.792565 0.913159

MCC 0.900291 0.814759 0.757513 0.612908 0.816934

NPV 0.942747 0.886199 0.865931 0.811548 0.920999

FPR 0.051218 0.084018 0.117161 0.171096 0.113998

FNR 0.048027 0.099505 0.124816 0.21699 0.0707

	ICMO	LSTM	CNN	RNN	NN
Accuracy	0.950522	0.907592	0.878838	0.807331	0.908433
Sensitivity	0.951973	0.900495	0.875184	0.78301	0.9293
Specificity	0.948782	0.915982	0.882839	0.828904	0.886002
Precision	0.95708	0.926845	0.891072	0.802355	0.89757
F-measure	0.95452	0.91348	0.883056	0.792565	0.913159
MCC	0.900291	0.814759	0.757513	0.612908	0.816934
NPV	0.942747	0.886199	0.865931	0.811548	0.920999
FPR	0.051218	0.084018	0.117161	0.171096	0.113998
FNR	0.048027	0.099505	0.124816	0.21699	0.0707

7.7. Statistical analysis

Table 5 shows that the evaluation of the Statistical analysis for the proposed ICMO model over the conventional schemes (SSA, DINGO, BOA, and CMBO). On having a glance at compared to other existing methods, the statistical analysis for the suggested model is less error-prone. The standard deviation of the ICMO method is 0.00662 and it converged in the mean as 1.085. Next, the median and maximum obtain 1.089 and 1.097. Finally, it converged in the minimum value of 1.07981 compared to the extant approaches. From the CMBO algorithm, standard deviation = 0.00857, Mean = 1.086, Median = 1.082, Max = 1.102 and Min = 1.081 get the highest value in the minimum analysis, which shows that the model is not efficient for event recommendations. Additionally, the minimum value for the SSA, DINGO, and BOA is 1.0844, 1.081, and 1.0814. The overall analysis has proved that the proposed model is must better at event prediction with less error rate.

Table 5
Statistical Analysis

Method Standard Deviation Mean Median Max Min

SSA 0.004191 1.085514 1.084476 1.101936 1.084411

DINGO 0.009143 1.085381 1.081091 1.105755 1.081091

BOA 0.007803 1.08427 1.081421 1.118115 1.081421

CMBO 0.008579 1.086937 1.082579 1.102331 1.081879

ICMO 0.006662 1.085076 1.081938 1.097233 1.07981

Method	Standard Deviation	Mean	Median	Max	Min
SSA	0.004191	1.085514	1.084476	1.101936	1.084411
DINGO	0.009143	1.085381	1.081091	1.105755	1.081091
BOA	0.007803	1.08427	1.081421	1.118115	1.081421
CMBO	0.008579	1.086937	1.082579	1.102331	1.081879
ICMO	0.006662	1.085076	1.081938	1.097233	1.07981

7.8. Performance evaluation

Table 6 illustrates the analysis of results between the adopted ICMO approach and the other methods such as W-GWO and BSH-EHA. The maximum accuracy of the ICMO is 0.9505 which is superior to the other two approaches, W-GWO = 0.9196 and BSH-EHA = 0.9470. Moreover, the sensitivity and specificity evaluation metrics of the ICMO are 0.958 and 0.947, respectively, which are higher than those of the other models. The W-GWO and BSH-EHA attained the lowest value in precision whereas the suggested model has a value of 0.9570. Furthermore, analysis for the negative measure (FPR, FNR) was also taken. The FPR measure of the developed model is 0.0512, W-GWO is 0.11 and BSH-EHA is 0.018; the adopted model has the least error values. The F-score of the suggested ICMO method is improved by approximately 0.933 and 0.94 over that of W-GWO and BSH-EHA. It is observed that ICMO has the highest value for the positive and other measures as well as obtained the least error.

Table 6
Performance evaluation

ICMO W-GWO [26] BSH-EHA [25]

Accuracy 0.950522 0.919632 0.94709

Sensitivity 0.951973 0.933844 0.916256

Specificity 0.948782 0.887264 0.982857

Precision 0.95708 0.933844 0.984127

F-measure 0.95452 0.933844 0.94898

MCC 0.900291 0.887264 0.902913

NPV 0.942747 0.902679 0.910053

FPR 0.051218 0.112736 0.087143

FNR 0.048027 0.066156 0.083744

	ICMO	W-GWO [26]	BSH-EHA [25]
Accuracy	0.950522	0.919632	0.94709
Sensitivity	0.951973	0.933844	0.916256
Specificity	0.948782	0.887264	0.982857
Precision	0.95708	0.933844	0.984127
F-measure	0.95452	0.933844	0.94898
MCC	0.900291	0.887264	0.902913
NPV	0.942747	0.902679	0.910053
FPR	0.051218	0.112736	0.087143
FNR	0.048027	0.066156	0.083744

7.9. Convergence graph analysis

The convergence graph analysis of SSA, BOA, DINGO, CMBO, and the ICMO is shown in Fig. 6. Once more the ICMO approach exhibits the largest improvement with respect to convergence rate. In Fig. 4, the adopted approach gets the value 1.095 in the iteration 0 to 10 then it converged to 1.085. Finally, it converged in iterations 30 to 50 and maintains the same least value 1.082. Next, the algorithm BOA attained the highest error rate up to iteration 5, and then it starts to converge, Finally, the model obtained 1.08456 in iterations 30 to 50. Further convergence analysis for the algorithm SSA, in iterations 10 to 50 maintains the same as 1.085 in comparison to the suggested model. The CMBO approach acquires the high error value of 1.1003 in the iteration 0 to 10, which finally converged to reach the value of 1.08432. This analysis has proved with better convergence of the least error rate when compared to other models.

Fig. 6.

Convergence graph for the proposed ICMO schemes over the existing methods SSA, BOA, DINGO, and CMBO.

7.10. ROC and AUC analysis

The examination of the ROC and AUC curve is shown in Fig. 7. AUC results varied from 0.5 to 1, with 0.5 designating an ineffective classifier and values closer to 1 designating an effective classifier. AUC achieved is 0.86, which is nearly 1. We evaluate the value of AUC in comparison to some of the earlier works. It is clear that the test data span a larger area than the training dataset between the upper-left corner and the curve diagonal. AUC for the test dataset was reached at 0.99, which is closer than AUC for the training dataset.

Fig. 7.

Analysis of ROC and AUC curve.

8. Conclusion

In this paper, a brand-new Improvised Cat and Mouse Optimization (ICMO) algorithm was presented. A set of features like personal willingness, community willingness, informative content, edge weight, and node interest degree were extracted. A hybrid classification was performed for the event recommendation system. LSTM and CNN models were combined to optimize weight in order to improve classification performance. In our upcoming work, we’ll aim to apply GPT-2 to the domain of recommendations in the hopes that we may investigate a compromise strategy to improve performance in recommendations tasks. Additionally, one of our key interests is in adapting our suggested approach to various contemporary application fields, such as Data protection, huge data comparison searching, and grouping, etc. Finally, the best result was then determined by evaluating the whole performance.

References

https://in.mathworks.com/discovery/convolutional-neural-network-matlab.html.

https://www-mdpi-com-s.web.bisu.edu.cn/1424-8220/21/15/5214/htm.

https://www.kaggle.com/c/event-recommendation-engine-challenge/data.

Bairathi and

Gopalani, Salp swarm algorithm (SSA) for training feed-forward neural networks, in: Soft Computing for Problem Solving, Springer, Singapore, 2019, pp. 521–534. doi:10.1007/978-981-13-1592-3_41.

Chen, APSO-LSTM: An Improved LSTM Neural Network Model Based on APSO Algorithm, 2020.

T.K.

Dang,

Q.P.

Nguyen and

V.S.

Nguyen, A study of deep learning-based approaches for session-based recommendation systems, SN COMPUT. SCI. 1 (2020), 216. doi:10.1007/s42979-020-00222-y.

E.G.L.

de Murillas,

H.A.

Reijers and

W.M.P.

van der Aalst, Case notion discovery and recommendation: Automated event log building on databases, Knowl Inf Syst 62 (2020), 2539–2575. doi:10.1007/s10115-019-01430-6.

Dehghani,

Š.

Hubálovský and

Trojovský, Cat and mouse based optimizer: A new nature-inspired optimization algorithm, Sensors 21(15) (2021), 5214. doi:10.3390/s21155214.

Du,

Meng and

Zhang, CVTM: A content-venue-aware topic model for group event recommendation, IEEE Transactions on Knowledge and Data Engineering 32(7) (2020), 1290–1303. doi:10.1109/TKDE.2019.2904066.

10.

Du,

Meng,

Zhang and

Lv, GERF: A group event recommendation framework based on learning-to-rank, IEEE Transactions on Knowledge and Data Engineering 32(4) (2020), 674–687. doi:10.1109/TKDE.2019.2893361.

11.

Gu,

Wang and

Yin, Social community detection and message propagation scheme based on personal willingness in social network, 2018.

12.

Gupta and

Srinath, A recommender system based on collaborative filtering, graph theory using HMM based similarity measures, International Journal of System Assurance Engineering and Management 13(Suppl 1) (2022), 533–545.

13.

Jelodar,

Wang,

Xiao et al., Recommendation system based on semantic scholar mining and topic modeling on conference publications, Soft Comput 25 (2021), 3675–3696. doi:10.1007/s00500-020-05397-3.

14.

Jiankui,

Zishu and

Benyong, Maximum likelihood DOA estimation using particle swarm optimization algorithm, in: 2006 CIE International Conference on Radar, 2006.

15.

D.S

Kudori, Event recommendation system using hybrid method based on mobile device, Journal of Information Technology and Computer Science 6(1) (2021), 107–116. doi:10.25126/jitecs.202161221.

16.

Lesage,

Deaconu,

Lejay et al., A recommendation system for car insurance, Eur. Actuar. J. 10 (2020), 377–398. doi:10.1007/s13385-020-00236-z.

17.

Li,

Wang,

Cheng et al., An event recommendation model using ELM in event-based social network, Neural Comput & Applic 32 (2020), 14375–14384. doi:10.1007/s00521-019-04344-0.

18.

Li,

Lv,

Zhu and

Song, Spatial-temporal topic model for cold-start event recommendation, IEEE Access 8 (2020), 214050–214060. doi:10.1109/ACCESS.2020.3040778.

19.

Li,

Zhu,

Fan and

Song, Hybrid deep framework for group event recommendation, IEEE Access 8 (2020), 4775–4784. doi:10.1109/ACCESS.2019.2962780.

20.

Liang,

Huang,

Bao and

Xuc, Sequential dynamic event recommendation in event-based social networks: An upper confidence bound approach, Information Sciences (2020).

21.

Liao,

Deng,

Wan and

Liu, Group event recommendation based on graph multi-head attention network combining explicit and implicit information, Information Processing and Management (2021).

22.

Liao,

Huang,

Xiong,

Wan and

Mao, Softwarized attention-based context-aware group recommendation technology in event-based industrial cyber-physical systems, IEEE Transactions on Industrial Informatics 17(10) (2021), 6894–6905. doi:10.1109/TII.2021.3054364.

23.

Mahajan and

P.D.

Kaur, Harnessing user’s social influence and IoT data for personalized event recommendation in event-based social networks, Soc. Netw. Anal. Min. 11 (2021), 14. doi:10.1007/s13278-021-00722-6.

24.

Mahajan and

P.D.

Kaur, Three-tier IoT-edge-cloud (3T-IEC) architectural paradigm for real-time event recommendation in event-based social networks, J Ambient Intell Human Comput 12 (2021), 1363–1386. doi:10.1007/s12652-020-02202-9.

25.

B.N.

Nithya, Meta-Heuristic assisted Contextual Post-Filtering Method for Event Recommendation System, 2023.

26.

B.N.

Nithya,

D.E.

Geetha and

Kumar, Optimization-assisted personalized event recommendation for event-based social networks, Advances in Engineering Software 176 (2023), 103368. doi:10.1016/j.advengsoft.2022.103368.

27.

Pelikan,

D.E.

Goldberg and

Cantú-Paz, BOA: The Bayesian optimization algorithm, in: Proceedings of the Genetic and Evolutionary Computation Conference GECCO-99 (Vol. 1), 1999, pp. 525–532.

28.

R.G.

Sakthivelan,

Rjendran and

Thangavel, A video analysis on user feedback based recommendation using A-FP hybrid algorithm, Multimed Tools Appl 79 (2020), 3847–3859. doi:10.1007/s11042-019-7293-1.

29.

Trinh,

Wu,

Wang et al., An effective content-based event recommendation model, Multimed Tools Appl 80 (2021), 16599–16618. doi:10.1007/s11042-020-08884-9.

30.

Y.J.

Tseng and

Wilson, Changes in recommendation rating systems, analyst optimism, and investor response, J Bus Ethics 166 (2020), 369–401. doi:10.1007/s10551-019-04139-2.

31.

Wu,

Lyu and

Liu, A personalized recommendation system for multi-modal transportation systems, Multimodal Transportation (2022).

32.

Wu,

Li,

Li et al., Graph embedding based real-time social event matching for EBSNs recommendation, World Wide Web 25 (2022), 335–356. doi:10.1007/s11280-021-00934-y.

33.

Zhang,

Li,

Li et al., Context-aware recommendation system using graph-based behaviours analysis, J. Syst. Sci. Syst. Eng. 30 (2021), 482–494. doi:10.1007/s11518-021-5499-z.