User-centric recommendations on energy-efficient appliances in smart grids A Multi-task learning approach

Deploying energy-efficient appliances is one of the most effective ways to save energy bills for residents. However, the existing recommender systems for energy-efficient appliances passively rely on energy consumption patterns without the knowledge of users ’ true needs. This paper proposes a user-centric energy-efficient appliance personalized recommender system (EEA-PRS) based on information collected from load monitoring platforms and e-commerce websites. The proposed system is built in a novel multi-task learning approach to collaboratively infer user ’ s preference on: (1) common types of appliances that appear in historical data; (2) energy-efficient models of common appliances; and (3) types of appliances that are novel to the users. The proposed system provides supervisory recommendation services with user feedback preferences on appliances as data labeling, which enables closed-loop evaluation to adhere to users ’ needs and interests. Simulation studies with comparative analysis have been conducted to validate its leading recommendation performance in terms of conforming to user preferences.


Background and motivations
With the rapid development of two-way communication facilities, advanced meter Infrastructure (AMI) with non-intrusive load monitoring (NILM), and smart appliances, countless energy consumption data with small time intervals from the residential side are collected and utilized in the smart grid context.The large dataset allows the utility to understand residences' energy usage patterns and develop more efficient applications which not only leads to more interaction between both sides but also saves more energy.For example, the Home Energy Management System (HEMS) [1] provides decision support for residential users to operate appliances under different pricing schemes efficiently.Another example is the electricity plan recommender system [2] which recommends appropriate electricity retailing plans to residential users by analyzing their energy usage data.
Since the residents' power consumption pattern is tightly connective to their habits and requirements, this paper focuses on another application that could potentially be achieved by using residents' energy consumption data, that is, how to detect residents' potential buying intentions on household appliances and further guide them to buy the energyefficient one either for replacing their outdated appliances or buying new types of appliances?In the literature, the necessity of deploying energyefficient appliances has been well studied.The study [3] showed a 20-year-old refrigerator could use 1700 kWh of electricity every year.By contrast, a similar-sized new refrigerator only consumes 450 kWh per year [4].From another perspective, the running costs of some electrical appliances can far exceed the sale price [5].In this case, upgrading outdated appliances into new appliance models with higher energy efficiency shows great potential to benefit residents in terms of reducing their energy bills.
Although energy efficiency can significantly affect overall energy consumption and running costs, these factors are hidden from residential users in most cases.Firstly, residents seldom change appliances, so they have few opportunities to experience the benefits of improving appliance energy efficiency, which makes them even less motivated to make this change.This forms a dead-end loop for residents.Secondly, manually selecting an appropriate energy-efficient appliance from thousands of candidate products is a challenging and inconvenient job for most residential users due to a lack of domain knowledge and professional advice.[6] shows the scarcity of knowledge about eco-labeling and energy consumption greatly impedes the appropriate selection of energy-efficient appliances.Therefore, rapidly and accurately detecting residents' potential needs and interests in appliances, as well as providing personalized recommendations on energy-efficient appliances, would be imperative to improve energy efficiency awareness and facilitate the appliance upgrade/procurement plan of residential energy users.

Review of energy-related personalized recommender systems (PRS)
Since the author in [7] provided the conceptual framework of the energy-related PRS in smart grids to help end-users select appropriate energy-related products, several PRS-related techniques have been applied to the energy system domain [2,[8][9][10][11][12][13][14][15].[2,[8][9][10][11] developed electricity plan recommender systems to recommend the cheapest retailing plan for the user.Since the price is the only criterion in this kind of application, every training user has "ratings" on various electricity plans, which means there is no data sparsity issue available.However, the utilization of basic Collaborative filtering (CF)-based and XGBoost techniques cannot explore the high-order relationships between users' features and provide great recommendation accuracy.[12][13] designed an appliance usage plan recommender system.[12] uses CF to recommend the energy-efficient energy usage pattern from the target user's neighbors who have similar living habits.[13] make recommendations for device operation based on the deep Reinforcement Learning (RL) method.The RL agent learns from the residents' overrides of the device operations as their feedback and finally optimizes energy consumption while minimizing discomfort to the residents.[21] designed a home battery recommender system based on users' preferences for battery products.
To the best of the authors' knowledge, few researchers investigated the recommendation systems in the field of household appliances.[14] aims to classify home appliances based on their features and recommend suitable ones for users in different regions, considering weather conditions.However, multiple essential factors are neglected, such as energy consumption and the price of each device.Reference [15] proposed a personalized energy-saving appliance recommender system based on NILM techniques, but there is no way to validate the user's preference for appliances of a certain type.
Despite the existing related works [14][15] showing a certain level of effectiveness in PRS on appliances, they still struggle to truly meet household needs and interests with the following practical concerns: 1) Recommendations passively relied on users' historical energy consumption data, which did not consider other supportive information, such as users' subjective needs and interests in appliance selection and their personal information (e.g., house type and region).2) Only the common appliance types appearing in users' historical data and extractable by NILM techniques were considered for recommendation, which narrows down the vision to the users and impedes users from exploring new appliances in the market.3) Due to the lack of user feedback, it is difficult to evaluate or enhance the recommendation performance.There is usually no evidence to know how well the recommended appliances fit the needs and interests of the target user.

Contributions of this work
A competent PRS on energy-efficient appliances would essentially aim to make reasonable recommendations on specific appliance models adhering to the needs and interests of the users.In practice, the critical challenge is the difficulty of knowing the users' exact interests because users can only give highly sparse feedback on numerous models of energy-efficient appliances.With this consideration, this paper leverages the appliance type recommendation as another highly related task with more concentrated user feedback and develops a multi-task learning approach to exert its correlation to the original task of appliance model recommendation.Meanwhile, CF-based techniques are incorporated to further help users explore novel types of appliances to fit their futureproof needs in the market.
The main contributions of the proposed energy-efficient appliance PRS (EEA-PRS) can be summarized as follows: 1) The proposed EEA-PRS enables supervisory recommendation services with user feedback on appliances as data labeling for recommendation performance evaluation.Such feedback is collected from user-case surveys delivered on load monitoring platforms and Ecommerce websites, which provides a more comprehensive understanding of users' needs and interests in energy-efficient appliances.2) Based on multi-task learning under a Multi-gate Mixture-of-Experts (MMoE) structure, the proposed EEA-PRS improves the classification performance about the user's preference on the specific models of the energy-efficient appliance by importing another highly related task of predicting users' preference on the appliance types.3) Based on the incorporation of multi-task learning and recommendation techniques, the proposed EEA-PRS is not only designed for common appliance recommendation but also adaptive to novel appliance recommendation.The common appliances aim to replace users' existing appliances, while the recommendations on novel appliances aim to guide users to explore new appliance types and help promote new products in the market.4) The proposed system is designed in a closed-loop manner.Based on the recommendations, users' further feedback and responses will help enrich the database and update the intelligent models to meet the ongoing changes in the market.
The rest of the paper is organized as follows.The framework of the proposed EEA-PRS is presented in Chapter 2. Chapter 3 introduces the data collection module.Chapter 4 presents the principle of the user interest inference module, followed by the mechanism of recommendation module elaborated in Chapter 5. Chapter 6 demonstrates and discusses the experiment results.Finally, Chapter 7 concludes the paper.

Framework of user-centric energy-efficient appliance personalized recommender system
In this paper, a new type of PRS is developed to recommend suitable energy-efficient appliances to residential users by leveraging multi-task learning and item-based CF in a closed-loop structure.This section first gives a brief technical introduction to PRS and then presents the overall framework of the proposed PRS with an overview of its key components.

Personalized recommender system
The principal function of a PRS for energy-efficient home appliances is to provide personalized information filtering on the numerous products available in the market [16].It helps users identify the items that best fit their personalized tastes and needs from a large repository of items.PRS can be employed in many fields, such as healthcare [17], e-commerce [18], tourism [19], etc.The unknown preferences of customers on the target item are usually predicted from the known data through the PRS algorithms such as CF [20] and the model-based method [21].Sometimes, the recommendation performance can be very inaccurate due to the sparsity of data (i.e., users only give a few feedback on each single item).In this case, two methods can be applied: (1) Getting aid from side information of users and items hence the model can learn interactions from features that are more widely available [21]; (2) Integrating other related tasks with less sparse data into one machine learning model, which is called multi-task learning [22][23][24][25][26].In the social recommendation field, [23] apply multi-task learning to learn social relations and user-item interactions simultaneously.The framework can provide better recommendations for users with few historical X.Guo et al. interactions by sharing user latent representations.[24] combine explicit and implicit feedback of users and mine cross-features between them based on a multi-task learning framework to recommend movies and music.To better achieve infant product recommendation, [25] take into account multiple types of user behavior data such as purchases, views and clicks.Each behavior type is treated as a task, and the model optimizes its performance on all tasks simultaneously.Basically, the multi-task framework can learn critical complement information from other tasks to support the recommendation's robustness.

The proposed framework and key components
The framework of the proposed user-centric EEA-PRS is shown in Fig. 1, where three modules work collaboratively to implement the PRS as below.
1) The data collection module is established to collect data collaboratively from the load monitoring platform and e-commerce website.
Four types of data are collected: (i) the user's side information, including their background information and energy-related data on appliances; (ii) an up-to-date list of energy-efficient appliances and their specifications; (iii) the user's interests in the type of common appliances which are normally recorded in the load monitoring platform.This part of the data is used as the "training labels" for appliance type preference prediction to enable supervised learning; (iv) the user's interests in the representative models of common appliances are to be collected from the e-commerce websites.This part of the data is used as a "training label" of appliance model preference prediction correspondingly.
2) The user interest inference module is to train a multi-task learning model through a supervised learning process.This machine learning model is to concurrently predict target users' preference ratings on (i) the types of common appliances and (ii) the models of energyefficient appliances in each type.
3) The recommendation module aims to make the final recommendations on energy-efficient appliances to the target user based on the predicted user's preferences on appliance types and models (i.e., (i) and (ii) in step ( 2)).This includes the recommendations on common and novel appliances as two separate outputs.Users' further responses based on our recommendations and their energy-related data will be fed back to enrich the database and update the PRS, which forms a closed-loop structure of PRS to be scalable to the changes in energy-efficient appliances and users' preferences.

Ownership, privacy, and benefits
The ownership of the proposed EEA-PRS can be flexible, but we recommend the utility companies as the main service provider or facilitators since they have large customer bases and can easily access customers' side information through the load monitoring platform and/ or survey study.They also have multiple ways to display the recommendation results, including smart meter displayers, electronic electricity bills and the e-commerce website.Meanwhile, appliance manufacturers can be attracted to display their products on the platform.During the whole process, the users' sensitive data can still be under the control of the utility company.
For the three parties involved in the transaction, they can all get benefits.The utility company can not only charge management fees from the appliance manufacturers but also foster the use of energyefficient appliances to alleviate the ever-increasing energy demand.The appliance manufacturers are facilitated to sell their products to residential users.The residents can save their electricity bills in the long run without compromising their well-being.

Data collection module
The data to build machine learning models is collaboratively collected from two sources: the load monitoring platform and the Ecommerce website.
Since residents' usage behaviors on home appliances are tightly connected to their interests and needs, this part of the data is considered significant side information for users.Load monitoring platforms will reliably provide this information by capturing the energy-related data of individual common appliances used by the residents.The specific features can be divided into three categories: (1) the types of common appliances in the target user's home; (2) the utilization time and frequency of these common appliances; and (3) the existing time of current appliances.Meanwhile, the platform can also collect target users' preferences in those types of common appliances in the form of short usercase surveys.
The E-commerce website first filters the qualified models of energyefficient appliances according to their energy rating labels and then provides the critical side information of the qualified appliance models, including ID, price, complexity of function and, which may influence the target user's preference.Besides, E-commerce provides the region and family income of users as part of their side information.Such background information of users can help enrich the potential feature interactions and support user preference learning.
Meanwhile, the target user's interests in the representative models of Fig. 1.The schematic of the proposed user-centric EEA-PRS.
X. Guo et al. common appliances can also be extracted implicitly from users' historical browsing on these appliance models or explicitly from pop-up surveys.The interest-related data can be considered as labels to facilitate the supervised multi-task learning on user's interests and needs.

User interest inference module
The multi-task learning structure of the user interest inference module is shown in Fig. 2. Based on the information from the data collection module as inputs, this module delivers two outputs simultaneously from two paths.Output A: Is the user interested in this type of appliance?Output B: Does the user have a potential preference for this specific appliance model?

The input representation
Considering that N instances make up the training dataset.For each instance, the data can be represented as D= (X, Y A , Y B ), where X includes the collected information of a pair of a user and an appliance, and Y A , Y B ∈{0,1} are the corresponding labels, indicating the user's preference on the type of the appliance and the specific appliance model, respectively.(Y i =1 means "like," and Y i =0 otherwise).X includes both categorical fields (e.g., place of residence and gender) and quantitative fields (e.g., energy consumption data).Each categorical field is transferred to a onehot-encoding vector.The X can be represented as a D-dimensional vector with m fields and inputted into the system with corresponding labels (i.e., Y A and Y B ) for training purposes.

Multi-task learning structure
In the proposed system, two highly related tasks mentioned earlier will be predicted simultaneously and incorporated by a single model to provide different recommendation services and improve the prediction accuracy complementarily.A common practice for multi-task learning is sharing the bottom structure of the neural network to learn the general information and only separating the two tasks in the last few layers of the neural network to study the specific features of each task.This is called "shared bottom" [22].However, the huge amount of shared information at the general bottom layers is mostly irrelevant, seen from each task, which undoubtedly decreases the prediction accuracy in individual tasks.In this case, a flexible framework that can distinguish the significant information from shared parts for the specific task is vital.This can be achieved by Multi-gate Mixture-of-Experts (MMoE) [26] which is the multi-task learning structure used in the proposed EEA-PRS.As can be seen from Fig. 2, for each task k, there is a corresponding gate network g k that receives the input features and performs a linear transformation with a softmax layer [27]: where ω g k ∈ R n×d is the weight matrix.Then, the outputs assemble the channels which can be different machine learning models with different weights (the sum of weights is 1) where channel i denotes the i th machine learning model.f k (X) is the weighted sum of the output of multiple models.In this way, different tasks can choose a diverse mixture of channel combinations and learn the most significant information and relationship between tasks for their own goals.After that, the results of assembling channels are passed into independent neural networks called "Tower" in order to learn the corresponding information for each task.
where h k denotes the independent network with the sigmoid function [28] for the k-th task, and the final output of the k-th task is Ŷk whose values are in the range of [0,1], representing the preference of the user on this type or model of appliance.Considering that both tasks are binary classification (i.e., customers are either interested or not interested), we define the loss function Loss as the addition of two binary cross entropy functions to pursue the balance in classification performance on the two tasks: Fig. 2. The multi-task learning structure of the user interest inference module.
X. Guo et al.

Channel algorithm
The choice of channel algorithm decides how much effective information can be chosen by the gate network.The simple machine learning model cannot cover all-round feature interactions.For example, logistic regression can well handle linear relationships between different fields but is weak in generalizing information and excavating non-linear relationships between fields, while the deep neural network (DNN) shows the opposite.Although a more complicated model tends to perform greatly during training, it may perform very poorly with new data cause the model takes care of every feature interaction, including noise features in the training dataset [29].As for our proposed application, the input features contain different types of data which have high dimensions after data preprocessing.Hence, there will be an abundant linear and non-linear relationship between features.On the other hand, the data volume is limited, the extremely complicated channel algorithm may perform overfitting.Under comprehensive consideration, we select deepFM [30] as our channel algorithm which uses a broadened structure to efficiently handle feature interactions in both linear and non-linear relationships and the numbers of weights needed to learn in the model is moderate.
As depicted in Fig. 3, the deepFM is comprised of two components: (1) the Factorization Machine (FM) [31] part and (2) the DNN part, which shares identical input features and the embedding layer.After one-hot encoding, each field of the input vector is sparse and of different lengths.This suggests an embedding layer to compress the input vector to a low-dimensional, dense vector before further feeding into two components.Otherwise, the network is difficult to converge.The output of the embedding layer is: where e i is the embedding of i th field with same length.Meanwhile, the e i participate in the training of FM by replacing initialX fieldi thus providing references for the training of DNN part.The goal of the FM part is to learn the order-1 and order-2 feature interaction.
where ω i is the weighting matrix to weigh the importance of the order-1 feature.v i is the latent vector for e i .〈...〉is the scalar product, indicating the impact of the order-2 feature combination.The bulk of the DNN component is a feed-forward neural network which is expected to improve the generalization of model and dig implicit feature combinations.After the embedding process, the α (0) is fed into the deep neural network, and the forward propagation process is: where l and L denote the current number and the total number of the hidden layer.σ is the activation function.α (l) , ω (l) , and b (l) are the output, weighting matrix, and bias of the L-th layer.Finally, the output of the DNN part is: Then, the output of DNN and FM parts are concatenated and ready to be weighed by the gate network in Fig 2.

Recommendation module
For a target user, the recommendation module serves as the final module to consolidate the multi-task prediction outputs of user interests and make the final recommendations on both common and novel appliances.

Models of common appliances recommendation
The preference ratings of the target user on common appliance models have been predicted by the user interest inference module (i.e., output B).Through the conversion of the sigmoid function, all predicted ratings are confined between 0 and 1.For common PRS systems, the value greater than 0.5 is considered as "like" [32].Similarly, candidate appliance models with at least 0.5 predicted preference value are considered as preferred items for the target user.These appliance models are then sorted in descending order, and the top n 1 products will be recommended to the target user.

Models of common appliances recommendation
In the process of preparation, each candidate novel appliance will compare their similarity with available common appliances.The similarity in this paper refers to how similar two types of appliances are in their characters and functions (e.g., amusement and cooling).We use a cosine-based approach to measure this similarity level.
where p and q denote the target novel type of appliance and the common Fig. 3.The structure of deepFM.
X. Guo et al. type of appliance on the e-commerce website.sim(p,q) denote the similarity degree between p and q; p → and q → denote the feature vectors of the type of appliance p and q.Based on the similarity level, the top-K most similar common appliances compared to the novel appliance will make up a set U q and are involved in the weighted aggregation in the process of prediction: wherer up is the predicted preference rating for the target user on this novel type of appliance.r uq is the predicted preference rating from the last stage (i.e., output A).The formula (12) repeats for every novel type of appliance; hence, each novel type of appliance gets its predicted rating and is sorted in descending order.The top-n 2 novel type appliance will update the new survey in the data collection module for the target user to get further responses.

Simulation study
In this section, we conduct experiments to validate the effectiveness of the proposed EEA-PRS.For convenience purposes, we name the type preference prediction on household appliances as task A, and the model preference prediction on the energy-efficient appliances as task B. The experimental process was developed in Python3.85.

Data preparation 6.1.1. Side information of energy-efficient appliances
In this paper, we consider 13 types of common appliances that are always recorded in the energy disaggregation dataset.We assume these types of appliances are common appliances.For each type of common appliance, we select 4 representative models with at least 3 energy rating stars which are considered energy-efficient appliances from the website of JD [33].These 4 representative models are selected to represent 4 distinctive appliance categories: economical & low power, economical & medium power, deluxe & medium power, and deluxe & high power.The parameters of these models are also collected.They are the price, complexity of function, and strength.The criterion for strength is different depending on the type of appliance.For example, the light is the irradiated area, while the air conditioner is the cooling area.Besides, we prepare 16 other types of novel appliances with different functions and characters.

Side information of users
We establish the profile of 2238 users based on the dataset [34] issued by the authors before.The side information of users can be classified into two parts: (i) Users' background information including region, usual residents, family income, house type and electricity bills.(ii) Users' energy-related data on common home appliances, including the current home appliances and their existing time, and users' average usage time/frequency on these common appliances.In the real world, energy-related data can be extracted from the load monitoring platform.Since this kind of data is scarce in current practice, we collect them from the survey to represent similar information.

User's interest data on the type of common appliances
We collect these 2338 respondents' subjective willingness on 13 types of common appliances based on the dataset [34].The rating value is either 1 or 0 (like or dislike).The number of label 1 accounts for 42.6% of total rating counts (i.e., 2338×13).This part of the data is considered the ground truth of task A.

User's interest data on the model of common appliances
The user's preference for the models of the common appliance should be collected by the e-commerce website either in implicit ways or explicit ways in practical situations.Due to the scarcity of this kind of dataset, we set predefined rules and apply it to individual users to generate personalized recommendations.The method is called rulebased recommender system [35], which is commonly used in previous energy-related recommender system [9][10][11]15,21].Several low-order and high-order rules are simulated based on basic popular trends in the real world: (i) The user with a high-income level would prefer the model with complex functions and high prices.(ii) The user with large energy consumption would like the model with high power output.(iii) The user with more family members would like the model with high strength.The rating value is either 1 or 0 (like or dislike).The number of label 1 accounts for 10.7% of the whole rating space (i.e.,2238×52).This part of the data is regarded as the ground truth of task B.

Other simulation setup
80% of the data is used for training, while the remaining 20% is for testing.For the deepFM, the embedding length is set as 4.Although a larger embedding length allows the model to learn more complex feature interactions, our data volume is not big enough to fit training data well with a large embedding length.Since the user's preference data is rule-based, the complexity of the proposed problem is not as high as in practical situations.After a grid search, the DNN which has one hidden layer with 128 neurons can satisfy our needs.For the MMoE, the increase in the number of channels require huge computation cost, and it may lead to overfitting problems.After experiments, we use 4 channels, which is the optimal number of channels with the best classification performance selected from 1 to 5. Adam [36], with a learning rate decay, is used as an optimizer.It is shown that the dataset has an imbalanced distribution of positive and negative samples, that is, for each type of appliance and each representative model of appliance, only a small proportion of respondents have positive feedback.Considering the data imbalance, the Area Under Receiver Operating Characteristic (AUC) [37] is used apart from F-Measure [38] to evaluate the models' classification performances.

Comparative study
The features collected from residents and appliances are complicated.It is important to select a comprehensive model to learn useful information from feature interactions well.We compare our proposed channel algorithm (i.e., deepFM) with four other benchmarks: (1) Logistic Regression (LR), which is effective for learning linear feature interactions, and (2) Factorization Machine (FM).The algorithm can effectively capture linear and pairwise feature interactions and is well suited for high-dimensional datasets; (3) Multiple layer perception (MLP), which is effective for learning non-linear feature interactions.It is an essential component of a deep learning-based algorithm; (4) Neural collaborative filtering (NCF), which leverages the strength of both neural networks and general matrix factorization to capture complex, non-linear patterns.
As can be seen from Fig. 4, in either task A or task B, the AUC and Fmeasure of deepFM are in the lead.For simple low-order feature interaction learning methods (the LR and FM), they cannot well learn the complex relationships between features that affects users' preferences on appliances.The poor performance, especially for task B, on F-measure indicates that these models are unable to choose the right side of the minority class.In other words, they cannot accurately predict users' preferences for appliance models.In comparison, the high-order feature interaction learning method (the MLP) has a better performance than LR and FM since it captures complex non-linear relationships between features and the target variable.However, the linear transformation in the input layer is not enough to capture linear relationships for these two tasks.The combinations of lower and high-order feature interaction methods (the NCF and deepFM) perform great since they leverage strengths of linear and non-linear modeling.The lack of pairwise feature learning in NCF makes it not as good as deepFM in this rich feature dataset.In general, the deepFM is good at capturing both linear and nonlinear feature interactions in both tasks due to its FM and MLP structures.This makes it well-suited for our proposed tasks which include rich features for users and home appliances.
We also compare the proposed system (i.e., deepFM+MMoE) with 11 benchmark methods to validate the effectiveness of multi-task learning.
We divide those methods into four groups: (1) the FM-based group; (2) the MLP-based group; (3) the NCF-based group; (4) the deepFM-based group.In each group, we compare the performance of these bulk models based on different structures: (1) sole model, which means the multi-task learning framework is not deployed; (2) shared bottom framework; (3) our proposed MMoE framework.The performance of different methods is shown in Fig. 5.As can be seen, the results clearly show the significant advantage of the deepFM +MMoE framework.For both task A and task B, it greatly improves the performance in both criteria.It is worth noting that the performance of methods with MMoE (right points in each figure) is better than others in their respective groups.This is attributed to the gating networks in MMoE which efficiently extract effective task-specific information from the common channels.This allows the model to capture feature connections in each task and improve the performance.In comparison, for methods with the shared bottom framework, the performance is different according to the types of bulk models.For models with FM structure, they learn excessive general non-linear information from the shared bottom during the process of pairwise feature combination.The branch of each task fails to capture key information for its corresponding task, which derates the classification performance.In comparison to the sole models for each individual task, the sharing bulk models in the proposed system receive both features of the two tasks and are trained from end to end.These traits help the model find complementary and connective information for each task, which makes it perform better than the sole model for every single task.

Case study
In this section, we randomly select two typical users #1209 and #1079 from the testing dataset to demonstrate the recommendation performance of the proposed system.The former has a single resident with the lowest degree of energy consumption and family income, living in a dorm, while the latter has above 5 residents with the highest degree of energy consumption and family income, living in a house.

Recommendation results on common appliances
The predicted preference rating of both the type and model of common appliances is shown in Fig. 6.Due to the limited space, we use the model index to substitute the names of appliance models.For each type of appliance, the sequence of models is economical & low power, economical & medium power, deluxe & medium power, and deluxe & high power.Basically, the model gets a high predicted rating only when their belonged type also gets a high predicted rating, which proves that our proposed model well handles the relationship between these two related tasks.For user #1209, the models with economical & low power get the highest predicted ratings in each appliance type, while for user #1079, the models with deluxe & high power get the highest predicted ratings.These recommendation results are expected to well meet the interests of the two users based on the conformance to their background profiles.

Recommendation results on novel appliances
The similarity between common appliances and each novel appliance is evaluated, and top-3 (k = 3) similarities are picked out and made weighted sum by the Eq. ( 11).The final predicted rating of each novel appliance and the proportion of its top-3 similar common appliances are shown in Fig. 7.
It can be seen that the average predicted ratings of novel appliances are higher for user #1079 than for user #1209.This is because the majority of the selected novel appliances are more similar to AC and light which are the high-rated common appliances of user #1079 as observed in Fig. 6.For user #1209, one example is that the garbage crusher gets the highest predicted rating because the garbage crusher's most similar common appliances (i.e., dishwasher and cleaner) have high predicted ratings to this user.
Overall, the proposed system can conform to the user's habits and potential interests to recommend novel appliances.

Conclusion and future works
In this paper, we extract to develop a user-centric EEA-PRS under a supervised multi-task learning approach.The data input is the user and appliance information extracted from load-monitoring platforms and ecommerce websites.The multi-task learning framework can collaboratively predict users' preferences on (1) the types of appliances and (2) the models of energy-efficient appliances.Item-based CF technique is further applied to make personalized recommendations on not only the common appliances appearing in users' historical energy consumption data but also the novel appliances that are new to the users.The simulation tests and comparative studies validate that the proposed system and methodologies can better handle two tasks collaboratively and improve user-centric recommendation performances.

Further discussion
The works in this paper aim to present a new PRS framework that helps understand users' preferences for energy-efficient appliances, provides supervisory recommendation services to meet the ongoing users' needs and interests, and encourages users to explore new types of appliances in the market.Based on the proposed framework, the following two factors could be further considered in the PRS to meet the practical challenges.Firstly, the temporal dimension of data can reflect the trend in users' preferences and behaviors recently.For instance, a user tends to buy cooling-related appliances during the summer or prefers to buy warming-related appliances during the winter.The lack of   time series data may lead to irrelevant suggestions in specific time periods in reality.Secondly, the user's economic benefits, such as savings and payback period brought by specific energy-efficient appliances, could be further considered in the items' side information.By doing so, even more personalized recommendations can be provided to meet users' financial needs and interests in energy-efficient appliances, aiming to reduce or eliminate the discouragement from adapting to energyefficient appliances.

Future works
Accordingly, in the future, a time-series-based PRS will be introduced to capture sequential dependencies in users' usage behaviors and preferences on home appliances.In this way, we can model the changes in users' preferences over time.Moreover, the economic benefits of energy-efficient appliances will be calculated based on users' smart meter data and current electricity plans.

Fig. 6 .
Fig. 6.The predicted rating of models and types of common appliances for users #1209 and #1079.

X
.Guo et al.